Generative AI and Firm Values

NBER WORKING PAPER SERIES

GENERATIVE AI AND FIRM VALUES

Andrea L. Eisfeldt

Gregor Schubert

Miao Ben Zhang

Working Paper 31222

http://www.nber.org/papers/w31222

NATIONAL BUREAU OF ECONOMIC RESEARCH

1050 Massachusetts Avenue

Cambridge, MA 02138

May 2023

Gregor Schubert gratefully acknowledges funding from the UCLA Fink Center for Finance and

the UCLA Easton Technology Management Center. The views expressed herein are those of the

authors and do not necessarily reflect the views of the National Bureau of Economic Research.

NBER working papers are circulated for discussion and comment purposes. They have not been

peer-reviewed or been subject to the review by the NBER Board of Directors that accompanies

official NBER publications.

sections of text, not to exceed two paragraphs, may be quoted without explicit permission

provided that full credit, including © notice, is given to the source.

Generative AI and Firm Values

Andrea L. Eisfeldt, Gregor Schubert, and Miao Ben Zhang

NBER Working Paper No. 31222

May 2023

JEL No. E0,G0

ABSTRACT

What are the effects of recent advances in Generative AI on the value of firms? Our study offers

a quantitative answer to this question for U.S. publicly traded companies based on the exposures

of their workforce to Generative AI. Our novel firm-level measure of workforce exposure to

Generative AI is validated by data from earnings calls, and has intuitive relationships with firm

and industry-level characteristics. Using Artificial Minus Human portfolios that are long firms

with higher exposures and short firms with lower exposures, we show that higher-exposure firms

earned excess returns that are 0.4% higher on a daily basis than returns of firms with lower

exposures following the release of ChatGPT. Although this release was generally received by

investors as good news for more exposed firms, there is wide variation across and within

industries, consistent with the substantive disruptive potential of Generative AI technologies.

Andrea L. Eisfeldt

Anderson School of Management

University of California at Los Angeles

110 Westwood Plaza

Suite C4.10

Los Angeles, CA 90095

and NBER

[email protected]

Gregor Schubert

UCLA Anderson

School of Management

[email protected]

Miao Ben Zhang

University of Southern California

701 Exposition Blvd

HOH-722

Los Angeles, CA 90089

[email protected]

Recent advances in Generative Artiﬁcial Intelligence are widely seen as a major technology

shock with important implications for ﬁrm values. Relative to earlier artiﬁcial intelligence

models, Generative AI models can digest more complex inputs, and can produce human-

like output, making Generative AI more versatile and scalable than prior innovations in AI

and machine learning. As a result, Generative AI has the potential for widespread corporate

adoption, with implications for ﬁrm values both across and within a wide array of industries.

One of the biggest questions surrounding advances in Generative AI is what eﬀect these

technologies will have on corporate valuations as a result of the impact of Generative AI on

ﬁrms’ labor inputs. We construct a novel dataset containing ﬁrm-level workforce exposures

to Generative AI. We provide a quantitative measure of the impact of Generative AI based

on our ﬁrm-level exposure data combined with ﬁnancial market data. Using this measure

we compute the ﬁrst estimates of the eﬀect of Generative AI on ﬁrm values by studying the

impact of the release of ChatGPT on ﬁrms with varying exposures to the technology shock.

We measure the impact of a major event in the advancement and dissemination of Gen-

erative AI technology, namely, the public release of ChatGPT, on equity returns at the ﬁrm

level. This event had a substantial impact on ﬁrm returns, consistent with Generative AI

advancement representing a major technological shock, one for which we can measure the

arrival and impact in real time. While ﬁrms may progressively adopt the technology, the

unmatched media attention and user base that ChatGPT has garnered within just months

indicates that ﬁrms and investors are actively assessing the potential fast diﬀusion of this

technology. We show that Twitter mentions and earnings call mentions of Generative AI

increased substantially following the release of ChatGPT. Moreover, the massive information

gathering and processing ability of ChatGPT itself allows us to assess each ﬁrm’s exposure

to ChatGPT’s disruption in real-time.

Our key ﬁnding is that the arrival of ChatGPT had a sizable positive eﬀect on the value

of ﬁrms whose labor forces are more exposed to Generative AI and related Large Language

Models (LLMs). Firms with higher exposure to the release of ChatGPT, as measured by

the exposure of their labor force to being made more productive by tools like ChatGPT,

outperform ﬁrms with lower exposures by over 40 basis points in daily excess returns during

the two weeks following its release. Notably, these return diﬀerences are not only due to

diﬀerences in labor force exposures across industries. Returns of ﬁrms with high labor

force exposures also outperform ﬁrms with low exposures by about 40 basis points daily in

Recent studies of Generative AI include Eloundou, Manning, Mishkin, and Rock (2023) who study the

impact of Generative AI on industries’ labor forces, Noy and Zhang (2023) who study the displacement

eﬀects of Generative AI on professional writing tasks, and Brynjolfsson, Li, and Raymond (2023) who study

the eﬀects of Generative AI on customer support agents, and Felten, Raj, and Seamans (2023) who consider

heterogeneity in occupational exposure.

industry-neutral portfolios.

Our methodology builds on the idea that ChatGPT and related technologies will in-

crease ﬁrm-level free cash ﬂows through a labor eﬀect that can work through two potential

channels. First, ﬁrms whose labor force can be substituted for with cheaper Generative

AI-based capital will experience higher free cash ﬂows by lowering input costs.

Second,

ﬁrms whose labor inputs are more complementary to Generative AI will experience higher

cash ﬂows due to the technological improvement in an input that is complementary to their

workforce.

While we do not take a stand on whether (and for which workers) Generative AI

is a substitute for, or a complementary to, labor, we are able to show that ﬁrms that have

a higher share of occupations exposed to Generative AI experience gains in value across a

wide array of industries. At the same time, the eﬀect of the release of ChatGPT on ﬁrm

values varies widely across industries, as well as within industries across ﬁrms. Indeed, we

ﬁnd a signiﬁcantly negative impact from the release of ChatGPT for some industries. Value

losses for incumbents are consistent with the idea that for some industries Generative AI

will lead to new entrants and displacement of existing ﬁrms. While advances in Generative

AI can have eﬀects through the product market as well as through the labor market (for

example, increasing demand for cloud computing services), our results support the idea that

AI advances will have a broad impact on the economy through its eﬀects on labor inputs.

The fact that the overall impact of the arrival of ChatGPT on ﬁrms with more exposure

to Generative AI is signiﬁcantly positive is consistent with recent studies showing that it is

increasingly more diﬃcult for new entrants to displace incumbent ﬁrms.

We measure ﬁrm-level exposure to Generative AI in two steps. First, we build on Eloun-

dou et al. (2023) and use ChatGPT itself to assess whether each of the 19,265 tasks cur-

rently performed by various occupations can be done by the current ChatGPT or by future

ChatGPT after investment in additional capabilities. Following Eloundou et al. (2023), we

aggregate the task-level exposure measures to the occupations in the O*NET database. Sec-

ond, and novel to our analysis, we map occupations to publicly-traded ﬁrms using data from

Revelio Labs. This dataset is constructed from millions of public employee proﬁles such as

LinkedIn. Our ﬁrm-level exposure measure thus captures the ability of the tasks currently

performed by labor at those ﬁrms to be performed (or made more eﬃcient) by Generative

AI. To the best of our knowledge, our study is the ﬁrst to create a ﬁrm-level measure of

exposure to Generative AI.

We next validate our labor-based measure of ﬁrms’ exposure to Generative AI by examin-

See Autor, Levy, and Murnane (2003) and Zhang (2019) for measures of ﬁrm exposure to automation

and Webb (2019) and Lane and Saint-Martin (2021) for the impact of AI on ﬁrms.

See Krusell, Ohanian, R´ıos-Rull, and Violante (2000) and Eisfeldt, Falato, and Xiaolan (2022).

See, for example, Guti´errez and Philippon (2019) and Akcigit and Ates (2020).

ing ﬁrms’ earnings call transcripts in 2023. We document a strong relationship between our

measure of exposure to Generative AI and ﬁrms’ discussions of Generative AI and related

technologies in ﬁrms’ earnings calls following the release of ChatGPT. In contrast, ﬁrms with

higher exposure to Generative AI do not increase discussions common technological topics

such as Engineering following the release of ChatGPT. Moreover, these ﬁndings remain even

after we exclude all ﬁrms from the most IT-related sectors,

suggesting that ﬁrms’ recent

discussions about Generative AI go beyond its impact on related products, and extend to

the impact on operations including labor inputs.

We start by showing the types of occupations that will be aﬀected by advances in Gen-

erative AI. We ﬁnd that the most aﬀected occupations are those that involve non-routine

cognitive tasks. This is in stark contrast with prior ﬁndings that automation mainly dis-

places occupations involving routine tasks (Autor et al. (2003)). Indeed, the most aﬀected

occupations are those with a high share of non-routine cognitive analytical tasks or routine

cognitive tasks, while manual physical tasks are relatively unaﬀected. Interpersonal tasks lie

in between cognitive and manual tasks in terms of their exposure to Generative AI. Occupa-

tions with higher wages also have higher exposure to Generative AI. Our result is consistent

with recent ﬁndings by Kogan, Papanikolaou, Schmidt, and Seegmiller (2019), who ﬁnd that

technological advances impact workers at the higher end of the wage distribution.

Exposure to Generative AI through ﬁrms’ labor inputs has an intuitive relationship to

average ﬁrm characteristics across and within industries. At the industry level, more exposed

sectors have higher wages, consistent with those sectors employing more workers in higher-

paid occupations that also tend to be more exposed to Generative AI. Regarding labor

inputs, ﬁrms in more exposed industries tend to have higher labor intensity in terms of the

number of employees per unit of capital, and lower asset tangibility. More exposed ﬁrms also

have higher ratios of organizational to total capital.

For the characteristics related to ﬁrm

valuation, more exposed sectors have lower average ﬁrm size as measured by total assets and

higher Tobin’s Q. Importantly, we also observe similar relationships between ﬁrms’ exposure

to Generative AI and ﬁrm characteristics within industry sectors. The robust patterns of

variation in industry and ﬁrm-level exposures with ﬁrm characteristics support our study of

stock returns both across and within industries.

Firms with higher exposure to Generative AI experience higher volatility around the

release of ChatGPT. However, it appears that it takes some time for the information in

ChatGPT’s release to be impounded into stock prices. The cumulative excess returns for

To be precise, we exclude the NAICS 51 “Information” and NAICS 54 “Professional, Scientiﬁc, and

Technical Services” sectors.

See Eisfeldt and Papanikolaou (2014) and Eisfeldt and Papanikolaou (2013).

Figure 1: Generative AI exposure quintile portfolio returns over time: market factor-

adjusted. The graph shows the cumulative excess realized returns on portfolios based on value-weighted

sorts. All portfolio returns shown are net of the risk-free rate. The data set consists of daily stock returns

from Yahoo Finance for Nov. 15, 2022 - March 31, 2023. The ﬁgure shows returns adjusted for market factor

exposure.

the highest-exposure quintile of ﬁrms vs. the lowest-exposure quintile diverge for several

weeks following the release of ChatGPT. Figure 1 plots the returns of the highest-exposure

quintile, the lowest-exposure quintile, and a long-short portfolio, which we denote AMH for

“Artiﬁcial Minus Human”. Cumulative returns to holding the AMH portfolio that is long

the highest-exposure quintile, and short the lowest-exposure quintile from the released date

through March 31, 2023, are over 9%.

We study the eﬀect of Generative AI on ﬁrm values by comparing the returns of ﬁrms with

higher and lower occupational exposure to Generative AI during and outside the two-week

window following the release of ChatGPT on November 30, 2022. The eﬀects are substantial,

and monotonic, within industries across Generative AI-exposure quintiles. Adjusting for the

market factor, the excess returns to quintile portfolios formed based on ﬁrm-level occupa-

tional exposure to Generative AI are monotonically increasing, with the highest-exposure

quintile of ﬁrms within industries earning positive excess daily returns of over 40 basis

points while the lowest exposure quintile experiences negative excess returns of around 25

basis points. The fact that these strong eﬀects exist within industries for many industries

provides evidence that Generative AI can have a broad impact on ﬁrm values through the

eﬀects on their labor inputs.

Across industries, the eﬀects of Generative AI on ﬁrm value also vary widely. Publishing,

information and computing-related industries have positive returns following the release of

ChatGPT, while ﬁnance and transportation-related industries experience negative returns

overall. Dispersion in industry returns is much higher during the two-week period following

the release of ChatGPT than over the full sample from November 30, 2022 to March 31,

2023 overall.

Our within-industry results also display striking diﬀerences across industry sectors. Within

ﬁnance, the return of more exposed ﬁrms relative to less exposed ﬁrms is substantially and

signiﬁcantly positive. Combined with the overall negative industry eﬀect, this is consistent

with some ﬁrms beneﬁtting greatly from Generative AI advances while overall the impact

of the release of ChatGPT was negative for value in the ﬁnance industry. Firms with higher

exposures to Generative AI within manufacturing as well as the administrative support,

waste management, and remediation services industry also signiﬁcantly outperform ﬁrms

with lower exposures. On the other hand, ﬁrms with higher exposures in the real estate

and rental and leasing industry signiﬁcantly underperform ﬁrms with lower exposures. This

could mean that existing ﬁrms with large exposures to Generative AI may be displaced by

new entrants in those industries. Finally, several industries do not display signiﬁcant re-

turn spreads following the release of ChatGPT, including construction of buildings, mining,

and heavy and civil engineering construction. The negligible impact in these industries is

consistent with manual tasks’ lower exposure to Generative AI.

Our study contributes to the literature studying the impact of disruptive technologies on

ﬁrm valuations.

Papanikolaou (2011) and Kogan and Papanikolaou (2014) study the eﬀects

of investment-speciﬁc technological changes on asset prices. Eisfeldt and Papanikolaou (2013)

and Eisfeldt and Papanikolaou (2014) study ﬁrms’ exposure to the organization capital

technology frontier. Zhang (2019) studies ﬁrms’ exposure to routine-biased automation.

In a series of papers, Babina, Fedyk, He, and Hodson (2020), Babina, Fedyk, He, and

Hodson (2021), and Babina, Fedyk, He, and Hodson (2022) study the eﬀects of AI on ﬁrm

growth, compensation, and workforce composition. See also Webb (2019) for the impact

of AI on ﬁrms. Kelly, Papanikolaou, Seru, and Taddy (2021) study ﬁrms’ exposure to

disruptive technological shocks using patent textual data, and Kogan et al. (2019) assesses

worker displacement from technological change over a very long sample. These studies oﬀer

important insights into investors’ and ﬁrms’ responses to technological shocks in historical

samples.

Our study departs from these works by focusing on measuring ﬁrms’ exposure to Gen-

See Greenwood, Hercowitz, and Krusell (1997) for an early contribution on the long-run impacts of

investment-speciﬁc technological change.

erative AI and assessing investors’ reaction to the technology shock upon its arrival. We

argue that the release of ChatGPT in November of 2022 is an observable, large technology

shock. We also highlight our contribution of measuring investors’ reactions to this shock in

real-time. Indeed, the information in market prices can potentially inform employees’ and

ﬁrms’ ultimate responses to technological disruption. Timely assessment of the market’s

expectations of Generative AI’s impact on ﬁrms can also help policy makers to eﬀectively

evaluate regulatory policies in response to the arrival of the new technology.

While other contemporaneous or recent studies such as Eloundou et al. (2023) also address

the exposure of occupations to Generative AI advances, our paper is novel in its contributions

to the eﬀect on ﬁrms. Our use of the Revelio Labs data to link occupations to ﬁrms yields

a unique opportunity to study corporate outcomes.

The paper proceeds as follows: Section I describes our data and measure of ﬁrms’ expo-

sure to Generative AI. Section II provides descriptive facts about Generative AI exposures

across occupations, industries, and ﬁrms. Section III documents corporate communications

to investors regarding Generative AI, and the relationship between those communications

and our measure of Generative AI exposures. Section IV presents our results documenting

the substantial changes in ﬁrm valuations following the introduction of ChatGPT. Finally,

Section V concludes.

I. Data and Measurement

We measure a ﬁrm’s labor exposure to Generative AI in two steps. First, we measure

each occupation’s exposure to Generative AI based on the occupation’s task statements from

the O*NET database. Second, we aggregate the occupation-level Generative AI exposure

measure to the ﬁrm level using the ﬁrm-occupational employment data from the Revelio

Labs Workforce Dynamics database.

A. Measuring occupational exposure to Generative AI

Occupational task data To assess whether an occupation will likely experience a change

in absolute or relative productivity as a result of Generative AI models becoming widely

available and used, we use a task-based approach. That is, similar to Eloundou et al. (2023),

Indeed, as we draft this study, IBM, the company ranked #1 in our exposure to Gen-

erative AI measure among the largest U.S. ﬁrms (see Table II) announced to halt hir-

ing of 7,800 jobs that could be replaced by AI. See https://www.businessinsider.

com/ibm-halts-hiring-for-7800-jobs-that-could-be-replaced-by-ai-report-2023-5?

utm_source=superhuman.beehiiv.com&utm_medium=newsletter&utm_campaign=

ibm-starts-replacing-jobs-with-ai&r=US&IR=T

we consider an occupation to be a set of tasks-to-be-done and evaluate for each task whether

it can be done more productively using ChatGPT and similar large language models (LLMs)

or future applications that will be built based on their capabilities.

We obtain information on the tasks involved in each occupation from the O*NET database,

which provides a list of task statements created by practitioners or experts.

A task state-

ment is usually one sentence, and an occupation has on average 22 tasks. The 19,265 pairs

of task statements and the occupations that they belong to then need to be coded as being

exposed to Generative AI technologies or not.

Task scoring We build on the approach for scoring tasks that was suggested and validated

by Eloundou et al. (2023). In particular, we use GPT itself to score exposure of tasks based

on whether the task can already be done directly using the ChatGPT interface, or can be

done with additional tools built on top of it. Two advantages of using an LLM to assess task

statements are that it allows for better replicability of the research in terms of cost and speed

of execution, and rapid scaling of the method to the full set of 19,265 task statements.

Speciﬁcally, we use Open AI’s GPT 3.5 Turbo model to classify the full set of task

statements and validate its reliability on a smaller subsample of tasks.

The model is given

an overall rubric for scoring LLM exposure, as well as two example interactions between a

user and an assistant that showcase the kind of output it is expected to produce. Then, a task

statement is submitted together with its O*NET title, and the model returns a score. The

scores capture whether the time taken to complete task is reduced by at least half, at constant

quality, if the worker can access ChatGPT-like tools. The scores fall into the following

categories: E0 indicates no exposure as the tool is either insuﬃciently useful for this task,

or cannot be brought to bear as a result of the intrinsic nature of the task, e.g. if it involves

physical activities; E1 is applied if a 50% reduction in completion time is already feasible

with the existing large language model interfaces; E2 requires that such a productivity gain

is feasible, but only once the current capabilities of the model can be deployed through

applications with further inputs (e.g. access to internet or proprietary databases), or if it is

trained on domain-speciﬁc issues or data; E3 is applied when the productivity increase would

require image processing capabilities in addition to current text processing. Importantly, the

This data can be accessed via the O*Net database at https://www.onetonline.org/

While similar large-volume classiﬁcation tasks in the past often relied on crowd-workers on online plat-

forms such as Amazon Mechanical Turk (MTurk), ChatGPT has recently been shown to outperform human

crowd-workers in accuracy in text classiﬁcation tasks, while also exhibiting lower variability in scores across

multiple runs of the program (Gilardi, Alizadeh, and Kubli, 2023). Economists have also recently used other

large language models to classify unstructured text from job postings and found that they outperformed

other machine learning methods Hansen, Lambert, Bloom, Davis, Sadun, and Taska (2023).

The structure of the prompt submitted to the Open AI GPT API is shown in Appendix A.

model is asked not only to respond with the score but also to explain its reasoning, which

allows the researcher to audit whether GPT is in fact understanding the prompt as intended

and interpreting the task correctly. Note that this exceeds the auditing capabilities that are

available in many instances of human text classiﬁcation - and that outputting this additional

information is enabled by the feature of LLMs that text generation is relatively cheap in terms

of time cost.

A random sample of scored tasks together with the model’s explanations can

be found in Table C1.

Consistency of Generative AI scoring To validate the consistency and replicability

of our procedure, we compare the scores assigned across 3 diﬀerent GPT runs (which may

vary in results due to the randomized order of example cases provided, or non-deterministic

features of the underlying LLM) for a randomly selected subsample of 100 task statements.

We compare the diﬀerent sets of scores as follows: First, we construct 3 diﬀerent classi-

ﬁcations for each task based on the assigned score: (1) “Current exposure”: score 1 has

been assigned. (2) “Expected exposure:” Either score 1 or 2 has been assigned. (3) “Broad

exposure:” Any score other than 0 has been assigned (this includes exposure conditional on

image capabilities becoming further developed). Then, we compute the agreement between

diﬀerent scoring runs with regard to which tasks belong in these categories. The compar-

ison between diﬀerent runs is shown in Appendix Table C2. We ﬁnd that the agreement

between diﬀerent GPT runs is very high - they arrive at the same score for at least 88%

of all cases independent of the exposure classiﬁcation considered. This validates that GPT

reliably provides classiﬁcations that are highly consistent across diﬀerent runs.

Scoring occupations’ exposure to Generative AI We next aggregate tasks’ exposure

to Generative AI to the occupation level. For each 8-digit Standard Occupational Classiﬁ-

caltion (SOC) occupation from the O*NET, we calculate the share of the total number of

tasks for each occupation that are aﬀected by Generative AI. We follow Eloundou et al.

(2023) and focus on an aggregation that takes into account that scores of 1 represent the

current direct feasibility of productivity improvements, while exposure scores of 2 rely on

investment in additional capabilities, such as interaction through custom-built applications

or the ability to search local or online databases, that complement the current LLM chat

interface. Therefore, our main measure of the share of an occupation’s exposed tasks counts

both the number of tasks with exposure rubric 1, (N

) and those with exposure rubric 2

) but applies half the weight to the latter. That is, our exposure score at the occupation

The model was also asked to return a conﬁdence score (low/medium/high) for its prediction, but while

this may have led the LLM to focus on reﬁning its answer in this regard, we do not use this dimension of

the response. In the large majority of cases, the model expresses “high conﬁdence” in its assessment.

level for each occupation o is:

tasks in o

+ 0.5 ∗ N

+ N

Finally, we aggregate across 8-digit O*Net occupation codes to the 6-digit SOC level to match

the occupation-level exposure measure to ﬁrms’ occupational data. Note that the Gener-

ative AI exposure measure is bounded by 100% on the upper end, which would represent

that all tasks in that occupation can be done at least 50% faster with the already-existing

functionality of ChatGPT and similar tools. On the lower end, 0% indicates that none of

the tasks involved in the occupation are likely to be more productive now, or even after

additional applications have been built on top of current Generative AI technology. The

full set of 6-digit occupations for which we compute Generative AI exposures consists of 778

occupations, of which 678 are also contained in the ﬁrm-level employment structure data

described below. The mean and median exposure in the latter set of occupations are 23%

and 18%, respectively, with a standard deviation of 21 ppt. The inter-quartile range extends

from 6% to 38% exposure.

B. Measuring ﬁrms’ exposure to Generative AI

To estimate a ﬁrm’s exposure to Generative AI, we use data on ﬁrms’ occupational struc-

ture to aggregate our occupational exposure measure. We obtain data on ﬁrms’ occupational

employment from Revelio Labs, which collects information on job titles and employers from

LinkedIn and other resume proﬁles and constructs occupation-by-ﬁrm employment counts.

Our customized data deﬁne a ﬁrm at the unique Compustat identiﬁer gvkey and deﬁne an

occupation using the 6-digit SOC. We use the employment counts for each gvkey-SOC6 as

of March 2022, which is the latest month in our data.

We construct a ﬁrm’s Generative AI exposure as the weighted average of its occupations’

Generative AI exposure, using the ﬁrm’s occupational employment as weights. That is,

occupations in f

EmpShare

f,o

∗ E

where EmpShare

f,o

emp

f,o

emp

is the employment share of occupation o in ﬁrm f. The result

of this procedure is a cross-section of 2,518 publicly-traded ﬁrms with predicted exposure to

Generative AI and basic company characteristics from Compustat. Summary statistics for

the distribution of Generative AI exposure across this set of ﬁrms are shown in Table I.

C. Other data

Firm earnings call transcripts data We manually collect ﬁrm earnings call transcripts

from the Seeking Alpha website for the tickers of all S&P 500 ﬁrms over the 2018-2023

period. To account for variations in the word forms, we applied standard natural language

pre-processing techniques to the text data, such as tokenizing and lowercasing. We then

counted the frequency of each keyword in each transcript and created a panel data set that

contained each ticker, quarter, year, and word counts for particular topics as variables.

Social media attention data To measure social media attention to GPT and related

technologies, we obtain data on Twitter mentions of “GPT” and “ChatGPT” by day for

2022 and 2023 from the media search platform Media Cloud.

II. Descriptive Facts about Generative AI Exposures

A. Generative AI Exposure and occupation characteristics

Table C3 shows an overview of the 20 occupations with the highest and lowest Genera-

tive AI exposure scores.

Note that among the highest exposure occupations, many, such

as “Telemarketers”, “ Computer programmers”, and “Interpreters and translators”, map

closely onto some of the key recent technological advances in Generative AI regarding its

ability to hold natural text-based conversations, generate functioning code based on high-

level descriptions of a programming task, and translate texts accurately between languages

and styles.

To better understand which occupation characteristics are associated with higher expo-

sure, Panel A of Figure 2 shows the relationship between the average wage level of each 2-digit

major occupation group in 2021 and our estimated Generative AI exposure. As the graph

shows, higher-wage occupations are generally more likely to be exposed to ChatGPT-like

technological advances making their constituent tasks more productive. One notable excep-

tion are the relatively low-wage “Oﬃce and Administrative Support” occupations, which are

URL: https://mediacloud.org/

Note that the lowest score category only shows a subset of a larger set of occupations with zero Generative

AI exposure.

While “Mathematicians” might seem out-of-place among occupations with high exposure to Generative

AI, note that Fields Medal winner Terence Tao of UCLA remarked in April, 2023, that “GPT-4 has saved

[him] a signiﬁcant amount of tedious work”, noting that “while these AI tools do not directly assist [him]

in core tasks such as trying to attack an unsolved mathematical problem, they are quite useful for a wide

variety of peripheral (but still work-related) tasks (though often with some manual tweaking afterwards).”

(Source: https://pandaily.com/mathematician-terence-tao-comments-on-chatgpt/)

also predicted to be highly exposed. The positive relation between Generative AI exposure

and wage is also observed at the industry level. In Panel B of Figure 2, we aggregate oc-

cupational Generative AI exposure and occupational wages to the NAICS 2-digit industry

level using the 2021 BLS National Employment Matrix. We observe a similar pattern that

high-wage industries also tend to have high exposure to Generative AI.

Occupational skills and Generative AI exposure Our measure of exposure to Gen-

erative AI technologies is based on the ability of tools like ChatGPT to make certain tasks

more productive. Thus, we would like to understand how our exposure measure relates to

other classiﬁcations of occupations which have been previously been deﬁned based on the

tasks that are involved in them. In particular, Acemoglu and Autor (2011) suggest that

technology-based productivity changes of past decades can be understood by scoring oc-

cupations based on the skills they involve. They suggest a characterization based on the

degree to which an occupation involves particular combinations of routine vs. non-routine,

cognitive vs. manual, and analytical vs. interpersonal aspects. To understand how their

categories map onto occupations that can be made more productive by ChatGPT-like tech-

nologies, we regress our 6-digit occupation Generative AI exposure measure jointly on the

set of occupational skill scores deﬁned by Acemoglu and Autor (2011). That is, we run the

following regression:

= α +

∗ Skill

+ ε

The results are shown in Figure 3. We ﬁnd that occupations with higher Generative AI

exposure are more likely to involve non-routine cognitive analytical skills or routine cognitive

skills, and less likely to involve diﬀerent kinds of manual skills, or interpersonal skills.

The literature on previous waves of computer-based automation argued that routine work

was most likely to be substituted by computers and to complement non-routine communica-

tion and problem-solving tasks (Autor et al., 2003). This “routinization” hypothesis assumed

that “computers and computer-controlled equipment are highly productive and reliable at

performing the tasks that programmers can script - and relatively inept at everything else”

(Acemoglu and Autor, 2011). While routine tasks and jobs were taken over by comput-

ers, workers skilled in “abstract” tasks were in high demand, leading to wage polarization.

However, while in the past abstract jobs requiring “problem-solving, intuition, persuasion,

and creativity” (Acemoglu and Autor, 2011) appeared safe from substitution by comput-

ers, Figure 3 suggests that the labor market impact due to recent Generative AI advances

may be diﬀerent. Tools like ChatGPT can interpret and respond to relatively unstructured

inputs, display a surprising amount of common sense in ﬁlling in gaps in instructions, and

can respond with relatively complex outputs, such as texts in diﬀerent styles, or – in the

case of image generators like Stable Diﬀusion or Midjourney – even with new and original

images (Bubeck, Chandrasekaran, Eldan, Gehrke, Horvitz, Kamar, Lee, Lee, Li, Lundberg,

et al., 2023). As a result, this wave of technological change may diﬀer from previous waves

in that many tasks in non-routine cognitive analytical jobs that were safe from automation

by previous technologies are now suddenly more likely to be substituted for by software and

computers. Combined with the unprecedented speed with which this wave of computer-based

automation tools is being adopted, this portends that the eﬀects on wages and inequality

across diﬀerent demographic groups may look very diﬀerent this time around than for the

automation waves of the past.

B. Summary of ﬁrms’ Generative AI exposure

Table I shows the summary statistics of our sample which includes 2,518 publicly traded

ﬁrms in the cross-section of 2022. On average, the ﬁrms in our sample have a mean and

median task exposure score of 35%, with a standard deviation of 8 ppt. The ﬁrm-level

measure of exposure to Generative AI spans from 27% at the 10th percentile to 44% at the

90th percentile. Figure 4 shows that the variation of ﬁrm-level exposure to Generative AI

has both across-industry and within-industry components. While industry sectors such as

“Information” and “Professional, Scientiﬁc, and Technical Services” have an average ﬁrm

exposure to Generative AI of about 13 ppt greater than industries such as “Accommodation

and Food Services,” there are substantial variations of ﬁrms’ exposure to GPT within each

industry. A variance decomposition shows that industry diﬀerences explain about 18% of

the ﬁrm-level variation in exposure to Generative AI.

Table II lists the 15 ﬁrms with the highest and lowest exposure to Generative AI, re-

spectively, among the top 100 largest U.S. ﬁrms by market capitalization in 2022. While

many IT ﬁrms, such as IBM and Intuit, not surprisingly have a large fraction of employees

exposed to Generative AI, we also observe manufacturing ﬁrms, such as 3M, and adminis-

trative conglomerates, such as S&P Global, in this high-exposure category. The large U.S.

ﬁrms ranked at the bottom of the exposure distribution include restaurants, such as Star-

bucks and McDonald’s, retail ﬁrms, such as Target and Walmart, transportation ﬁrms, such

as UPS, and manufacturing ﬁrms, such as Tesla, suggesting that they have a smaller fraction

of employees exposed to Generative AI.

The rich cross-industry and within-industry variations in ﬁrms’ exposure to Generative

AI motivate us to explore our ﬁrm-level empirical analyses within and across industries.

Importantly, our within-industry analyses also help to diﬀerentiate our labor-based mech-

anism from the product-based mechanism when studying the eﬀects of ﬁrms’ exposure to

Generative AI on ﬁrm values.

How do ﬁrms’ and industries’ exposure to Generative AI relate to their other character-

istics? Panel A of Table III shows that ﬁrms with higher exposure to Generative AI tend

to be smaller, have greater Tobin’s Q, and are less proﬁtable. These ﬁndings are consistent

with the notion that such small and high-growth ﬁrms tend to focus their workforce on

cognitive tasks such as R&D. Indeed, we also observe that ﬁrms with high R&D intensity

are more exposed to Generative AI. Moreover, we also observe that ﬁrms with higher la-

bor intensity, higher organizational capital ratio (Eisfeldt and Papanikolaou, 2013) and less

tangible capital are more exposed to Generative AI. These cross-ﬁrm ﬁndings hold consis-

tently within-industry and cross-industry as well. Panel B of Table III shows that the above

ﬁndings are qualitatively similar after we include NAICS 2-digit industry ﬁxed eﬀects in

the regressions. Figure 5 plots the relationship between industry sectors’ mean exposure to

Generative AI (averaged across Compustat ﬁrms) and industries’ other characteristics. In

line with the ﬁrm-level patterns, we observe a consistent picture that industries with higher

exposure to Generative AI have ﬁrms that are smaller, have higher Tobin’s Q and lower

current proﬁtability, and feature higher organizational capital and lower tangible capital.

III. Investor Communication and Exposure to

Generative AI Technology

To understand whether ﬁrms communicate that they are aﬀected by the technological

change as a result of the evolution of large language models and other Generative AI tools,

and to validate our bottom-up measure of language model exposure based on ﬁrms’ employ-

ment structure, we ﬁrst analyze the public communication between ﬁrms and investors. We

focus on earnings conference calls, as ﬁrms use earnings conference calls to communicate

their views regarding risks and opportunities, (see, e.g. Hassan, Hollander, Van Lent, and

Tahoun (2019)) as well as past and expected future performance. Moreover, they respond

to questions by analysts who may reﬂect investors’ perspectives of which issues particular

ﬁrms should focus on. If the technological change resulting from the recent rapid evolution

of large language models is indeed aﬀecting ﬁrms as we postulate, we should see both a time

pattern of increasing communication regarding these issues that coincides with the launch of

ChatGPT, as well as a cross-sectional pattern of larger increases in this kind of ﬁrm-investor

communication among ﬁrms for which our bottom-up measure predicts higher Generative

AI exposure.

A. Measuring earnings conference call mentions of technologies

We use the Seeking Alpha website to manually collect a panel of the earnings conference

call transcripts for S&P 500 ﬁrms, for calls that were held from July 2018 to March 2023. For

each of these earnings calls, we assign a calendar month, quarter, and year (as distinct from

the ﬁscal year and quarter referenced in the call), based on the time stamp of the earnings

call transcript. We process each transcript by converting it into a list of lower-case tokens,

creating separate lists of unigrams (one-word tokens) and bigrams (two-word tokens).

Topic deﬁnitions We deﬁne four categories of words (incl. their plural form where ap-

plicable) that we compare to the list of unigrams and bigrams in each earnings call: (1)

Machine learning-related words that are not speciﬁc to Generative AI technologies.

(2)

Generative AI -speciﬁc words: “LLM”, “ChatGPT”, “GPT”, “GPT3”, “GPT4”, “genera-

tive”, “language model”. (3) Generic engineering-related words: “engineer”, “engineering”.

Our sample size and composition vary somewhat across years, and we may not fully cap-

ture all earnings calls for each ﬁrm, but we capture the vast majority of relevant earnings

calls, with the number of unique S&P 500 companies with calls in our sample in each quar-

ter varying from 377 to 474. Moreover, our sampling procedure of hand-collecting publicly

available transcripts from Seeking Alpha is unlikely to be biased with regard to Generative

AI exposure - which would be the selection bias of relevance.

B. Aggregate trends in earnings call mentions

To see how mentions of these topics evolve over time, we compute two aggregate variables

for each month and topic: First, we capture the share of calls in our sample that mention

a word from that category, which captures the extensive margin of whether there is any

suﬃciently large perceived ﬁrm exposure for investors or ﬁrm management to mention it.

Second, we compute the mean number of category words per call in our sample, which

additionally captures the intensive margin of longer discussions of the topics. The aggregate

trend in these variables for the Generative AI topic, which is our proxy for ﬁrms discussing

ChatGPT-like technologies, is shown in Figure 6. For comparison, each graph also shows the

equivalent trends for other machine learning topics, and for generic mentions of engineering.

To allay concerns that only investors in software companies are likely to be knowledgeable

about, or have an interest in, the Generative AI topic, we also show the same trends excluding

the “Information” (NAICS 51) and “Professional, Scientiﬁc, and Technical Services” (NAICS

These are: “deep learning”, “ML”, “machine learning”, “deep learning”, “natural language”, “neural

net”, “neural network”, “NLP”.

54) sectors in Panels B and D, as determined by Compustat industry codes.

While the sample only reﬂects companies that had earnings calls before March 2023, there

is an unmistakeable break in the trend shortly after the release of ChatGPT in November

2022: both measures show that an increasing number of ﬁrms seem to consider language

models and generative AI as important enough that they are discussed in earnings conference

calls. Moreover, these trends are separate from discussions of machine learning or engineering

in general, which do not show a large increase in 2023 – so this is not simply a reﬂection of

companies in general - or in our sample - becoming more technical in their language over

time. Generative AI-related topics were discussed in 27% of all earnings conference calls

by March 2023, and in 13% of calls outside the software-related sectors. The number of

mentions per call in Panel C rises even more steeply than the share of ﬁrms discussing the

topic, suggesting that discussions of the topic increase at the intensive margin in addition to

the extensive margin. Overall, earnings call mentions show that the rapid rise in importance

of language model technologies is reﬂected in the communication between ﬁrms and analysts

- suggesting that they are also likely to inﬂuence valuations by investors.

C. Firm-level Generative AI exposure and earnings call topics

While Figure 6 shows that general interest in LLM-related topics is on the rise following

the release of ChatGPT in November 2022, we also want to understand which ﬁrms are more

likely to discuss related topics. To see how our ﬁrm-level predicted Generative AI exposure

relates to these changes in earnings call mentions, we run repeated cross-sectional regressions

of the form,

[Topic X]

i,t

= α

+ β

+ γ [Topic X]

i,2019

+ ε

i,t

(1)

where the dependent variable is a binary indicator for whether company i’s earnings calls

in quarter t have any mentions of topic X, and E

is the ﬁrm’s Generative AI exposure.

We also control for whether the ﬁrm already mentioned the topic in any 2019 earnings calls.

This means that the coeﬃcient β

here estimates the degree to which ﬁrms with higher

Generative AI exposure are more likely to start (or stop) discussing topic X in earnings calls

in comparison to 2019.

The quarterly β

coeﬃcients over 2020-2023 for each topic are shown in Figure 7. The

ﬁndings validate our Generative AI exposure measure’s ability to pick up on ﬁrms’ exposure

to a technology shock driven by Generative AI. Panels A and B conﬁrm that exposed ﬁrms

– according to our bottom-up measure – were more likely to discuss Generative AI topics

than other ﬁrms, both in the full sample and when we exclude sectors with many software

and technology companies. Moreover, they only became more likely to discuss Generative

AI topics after the recent wave of innovation in those ﬁelds in Q4 of 2022 - so our exposure

measure does not identify companies that were involved in discussions about these technolo-

gies before ChatGPT was released. Put diﬀerently, the fact that our Generative AI exposure

measure does not predict a higher likelihood of mentioning these technologies before the re-

cent productivity shock suggests that it does a reasonable job of identifying companies that

are newly exposed to this particular way of using AI to improve productivity. Moreover, the

similar patterns in Panels B and D, where we exclude ﬁrms in the “Information” (NAICS 51)

and “Professional, Scientiﬁc, and Technical Services” (NAICS 54) sectors, conﬁrms that this

pattern is not limited to investor communication for IT companies. The eﬀect on mentions

of generic engineering topics in Panels C and D shows that the break in trend with regard to

the eﬀect of our Generative AI exposure measure on whether a ﬁrm mentions Generative AI

does not extend to higher mentions of engineering topics more generally. This validates our

ﬁrm-level exposure measure capturing the potential impact of technologies like ChatGPT in

particular.

The magnitude of these eﬀects is large: The β

coeﬃcients indicate the percentage point

change in the probability that the ﬁrm mentions the topic in an earnings call in that year,

relative to 2019, for each percentage point change in the share of tasks in that ﬁrm that is

exposed to Generative AI . That is, the March 2023 coeﬃcient in Panel A suggests that a 1

ppt increase in ﬁrm exposure is associated with a more than 1 ppt higher likelihood in 2023

of the ﬁrm mentioning Generative AI relative to 2019.

Firm-level panel regressions of earnings call topics on Generative AI exposure

In order to be able to more ﬂexibly control for diﬀerent ﬁrm characteristics in determining

the association between our Generative AI exposure measure and topics mentioned in ﬁrm

earnings conference calls, we also estimate ﬁrm-level panel regressions. We aggregate our

monthly data into a ﬁrm-quarter panel in order to ensure that most ﬁrms have a continu-

ous time series of earnings calls over the period of Q3 2018–Q1 2023. Then, we estimate

regressions of the form

[Gen. AI Topic Mentioned]

i,t

=α

+ α

+ β

× [Post-ChatGPT] (2)

+ γ [Post-ChatGPT] + ε

i,t

(3)

where we regress whether a ﬁrm mentions Generative AI related words in an earnings con-

ference call in that quarter on the measure of ﬁrm exposure to to the technology and its

interaction with an indicator of whether the quarter is “post-ChatGPT”, which corresponds

to Q4 2022 and Q1 2023 in our sample. This speciﬁcation allows us to quantify the precise

degree to which higher exposure under our measure is associated with a higher likelihood of

mentioning Generative AI technologies after the launch of ChatGPT compared to before –

which was not possible in the repeated cross-sectional regressions in Figure 7.

The results of estimating Equation 3 in our sample are shown in Table IV. In the ﬁrst four

columns, we add increasingly stringent combinations of ﬁxed eﬀects to the baseline regression.

In column 1 we control for calendar quarter ﬁxed eﬀects, which will capture the degree to

which people are in general more likely to speak about the topic in particular periods. We

also add ﬁrst industry sector ﬁxed eﬀects, and then ﬁrm ﬁxed eﬀects, to control for the degree

to which particular sectors and ﬁrms are more likely to mention particular technology topics

at all times. The most stringent speciﬁcation in column 4 suggests that a 10 ppt increase in

Generative AI exposure is associated with a 7.6 ppt increase in the likelihood of mentioning

the Generative AI topic – or that the interquartile range of exposure to Generative AI is

associated with about a 7 ppt diﬀerence in the likelihood of talking about the topic. To

ensure that this pattern is not limited to software companies, we repeat the most stringent

estimation in column 5, and exclude NAICS sectors 51 and 54. The result shows that,

even outside of these sectors, more exposed ﬁrms are signiﬁcantly more likely to mention

Generative AI after the ChatGPT launch, but the coeﬃcient is smaller, suggesting a 3.1 ppt

increase in the share of ﬁrms mentioning it for a 10 ppt increase in exposure. In general,

our results on investor communication reveal that the technological advances in question are

salient to investors, rapidly rising in importance based on their prevalence in communication,

and more likely to be mentioned by ﬁrms that our bottom-up measure predicts to be more

exposed.

IV. Stock Market Impact of Generative AI Exposure

A. Stock return volatility, social media attention, and ﬁrms’ Generative AI

exposure

Our measure of ﬁrm-level Generative AI exposure is intended to capture the relative

degree to which tasks in a ﬁrm can be made more productive by advances in language models.

However, it is ex-ante ambiguous whether greater exposure should be associated with higher

returns: while higher productivity should lead to lower costs of producing the output of ﬁrms

in a particular industry, the degree to which this cost improvement is captured by incumbent

ﬁrms can vary. For instance, if new entrants are more likely to be ﬂexible enough to take

advantage of the beneﬁts of large language models, incumbent proﬁts could actually fall in

response to the technology shock.

Even if the sign of the impact of Generative AI exposure on stock returns is unclear, the

release of new information about the technology can be expected to increase the volatility

of returns for aﬀected companies. Conversely, if our exposure measure is a valid proxy for

ﬁrm-level characteristics that are speciﬁc to GPT-like technologies, higher return volatility

associated with this exposure should be associated with new information about these tech-

nologies. First, we explore whether Generative AI exposure is in fact associated with higher

stock return volatility on the days after the release of ChatGPT on November 30th, 2022,

which was the beginning of the period when the technological advances around large lan-

guage models started to become widely known. For each trading day from November 15th

to March 31st, we run separate regressions of the form

|r|

= α

+ α

ind

+ β

GenAIExposure

+ ε

where β

captures the degree to which our ﬁrm-level measure of Generative AI exposure

predicts higher return volatility on that day, and we control for 2-digit industry sector ﬁxed

eﬀects α

ind

to capture the degree to which Generative AI exposure might simply correlate

with industry news coming out around the same time. The time series of t-statistics testing

whether this coeﬃcient is zero on each day are shown in Panel A and B of Figure 8 for

the speciﬁcations without and with industry ﬁxed eﬀects. There are only a small number

of periods when the stock return volatility eﬀect of Generative AI exposure clearly exceeds

conventional signiﬁcance thresholds with a t-statistic well above 2, but in both graphs they

include November 30th, 2022, which is the day that ChatGPT was released to the public, and

at least one other day in the two weeks after the release, which corresponds to the period

when many major news outlets were ﬁrst covering the surge in interest in large language

models, that resulted from encounters with ChatGPT’s capabilities. Note that, even though

our exposure measure uses no information other than a ﬁrm’s employment structure, its

association with stock returns actually identiﬁes the relevant period when the capabilities of

ChatGPT were ﬁrst becoming known, suggesting that it captures exposure to this particular

technology.

Panels C and D of Figure 8 conﬁrm that the days on which our ﬁrm employment-structure

based measure of Generative AI exposure predicts higher stock return volatility around the

release of ChatGPT closely coincide with surging social media attention to the technology:

Twitter mentions of “ChatGPT” and “GPT” spike relative to their trend right after the

release of ChatGPT, supporting the argument that the higher association between stock

volatility and Generative AI exposure in this period indeed reﬂects investor incorporating

new information into stock prices. The fact that the volatility impact of our Generative AI

exposure measure reliably identiﬁes the key period when major related news was released

validates that it captures a dimension of exposure to the associated technologies that is

relevant to market participants.

Deﬁning the ChatGPT “event” period In order to identify the eﬀect of Generative

AI exposure on ﬁrm returns in this relatively short time series, we want to focus on days

when market participants were incorporating substantial news about related technologies

into their valuation of ﬁrms. The method above suggests that Generative AI exposure was

associated with high stock return volatility particularly during the initial ChatGPT launch

and in the weeks thereafter as the market discerned the likely impact of this technological

advance and aggregated relevant information. This period also coincides with a time period

when the growth in Twitter mentions of the topic suggests a high level of interest and new

information. We, therefore, focus on the “ChatGPT release period” consisting of November

30, 2022, and the two weeks following it in our main analysis of the stock return impact of

the new technology.

B. Realized returns and Generative AI exposure

Forming Generative AI exposure portfolios To estimate whether Generative AI ex-

posure aﬀects the realized returns of stocks during the event window, we ﬁrst form value-

weighted high and low exposure quintile portfolios, and also a high-minus-low portfolio (H-L)

– which we will also refer to as the “Artiﬁcial Minus Human” (AMH ) portfolio – that repre-

sents the zero net investment portfolio long high exposure (H) stocks and short low exposure

(L) stocks.

All portfolios are formed based on market capitalization weights as of October

31, 2022, and exposure measures based on March 2022 ﬁrm-level employment structures,

and weights are adjusted daily to mimic passive buy-and-hold exposure. The returns data

set consists of daily stock returns from Yahoo Finance for Nov. 15, 2022 - March 31st,

2023, as well as Fama-French 5-Factor data and risk-free returns from Ken French’s web-

site. We also form industry-neutral portfolios by ﬁrst forming within-2-digit NAICS industry

value-weighted quintile portfolios, and then averaging portfolio returns for the same quintiles

across industries while applying industry market capitalization as weights.

Generative AI exposure returns after ChatGPT release We compute realized excess

daily stock returns across Generative AI exposure portfolios during diﬀerent time periods,

To be precise, we use returns on all trading days from Nov. 30, 2022 to Dec. 14, 2022.

See the Appendix for details on the portfolio construction.

estimated from regressions of the form

− r

=α

release

[ChatGPT release period]

+ α

not release

[Not ChatGPT release period]

+ β

(Mkt

− r

) + ε

where the intercept represents either the mean return for the full period, or is allowed to vary

with whether the day is part of the ChatGPT release period or not, as deﬁned above. Table

V shows the mean excess realized returns by portfolio for diﬀerent speciﬁcations, comparing

returns for the full time period of Nov. 15, 2022 - March 31, 2023, and returns on the days

when we would expect the Generative AI exposure to matter for returns, i.e. the two weeks

after the ChatGPT release.

Panel A of Table V shows raw excess returns: The AMH high-minus-low exposure port-

folio has positive daily excess returns of 0.4% (t-statistic > 3) on average during the post-

ChatGPT release dates, but not outside of that period. This key ﬁnding – that highly

exposed companies have higher average returns only on days when advances in language

model technology become known – also validates our bottom-up measure of technology ex-

posure, as it predicts higher returns if, and only if, the dates are likely to be associated with

updating about the potential of the recently released technology.

In Panel B of Table V, we additionally control for each portfolio’s exposure to the market

return factor. The magnitude of the positive return eﬀect is largely unaﬀected, and continues

to be statistically highly signiﬁcant and large. One additional concern might be that the

Generative AI exposure quintiles load on particular industries, such as technology companies,

and that the exposure to the relative performance of these industries drives the realized

return variation across exposure quintiles. While the fact that the portfolio outperformance

occurs only on GPT news release dates makes it unlikely that this is an important issue

in our analysis, in Panel C we show the performance of sector-neutral factor portfolios,

also adjusted for market factor exposure.

In this speciﬁcation, the excess returns to the

AMH portfolio are almost identical to those in Panel B, suggesting that the Generative

AI exposure factor is not simply driven by an association between exposure and particular

industry sectors.

Note that, across the diﬀerent models in Table V, mean daily excess returns across the

period as a whole and returns outside of the ChatGPT release period are small, statistically

indistinguishable from zero, and almost ﬂat across Generative AI exposure quintiles for the

These portfolios are formed by sorting ﬁrms into quintiles within each industry and then forming quintile

portfolios by taking a weighted average of the corresponding quintile in each industry sector, thereby pre-

venting the quintiles from loading excessively on particular industries. See the Appendix for further details

on the portfolio construction.

period as a whole. This reduces concerns that Generative AI exposure spuriously correlates

with other ﬁrm characteristics that are driving diﬀerential returns over the period in question.

In contrast, on the days when exposure to large language model productivity changes should

actually matter, returns vary signiﬁcantly with whether a ﬁrm is more or less exposed based

on our bottom-up task-based measure. In the two week period after the ChatGPT release,

excess returns monotonically increase across Generative AI exposure quintiles, as shown in

Figure 9 for the market-factor adjusted, industry-neutral exposure quintiles.

As a robustness check for our analysis, we also consider whether the excess returns on

the Generative AI-exposed stocks after the release of ChatGPT can be explained by an

association between this exposure and other risk factors. We replicate the previous portfolio

returns analysis but additionally control for the returns on the factors in the Fama French

5-factor model (Fama and French, 2015). The results are shown in Appendix Table C7.

We ﬁnd that the 5-factor-adjusted excess returns (and the industry-neutral version) on the

AML portfolio during the ChatGPT release period are smaller, albeit still sizable, in this

speciﬁcation, suggesting daily excess returns of 0.3%, and statistically highly signiﬁcant.

Moreover, Appendix Figure B2 shows that excess return variation across Generative AI

exposure portfolios is highly monotonic even in this speciﬁcation.

Time series of Generative AI exposure portfolio returns To further validate that

the Generative AI exposure factor returns are associated with the advances in the related

technology, we also consider the time series of daily excess returns on industry-neutral port-

folios in Figures 1 and 10 – which show the results for global sorts and industry-neutral sorts

– comparing the cumulative excess returns of the high, the low, and the high-minus-low

AMH portfolios. The graphs of the high and low exposure portfolios show that, if anything,

that the high exposure portfolio had lower returns in the two weeks preceding the release

of ChatGPT. On the other hand, the returns of the “A” and “H” portfolios diverge rapidly

thereafter, with most of the gains to the high exposure portfolio concentrated in the two-week

period after the release. The cumulative returns to the high-minus-low zero net investment

portfolio over the full time period through March 31st are large. Cumulative returns to

holding a portfolio that is long the highest-exposure quintile, and short the lowest-exposure

quintile from the release date (Nov. 30, 2022) through March 31, 2023 are 9.4% for the

global sort across all stocks, and 5.1% for the sector-neutral sorts (see Appendix Figure B4

for the time series of the cumulative excess returns without adjusting for market factors).

The corresponding values when adjusting for the market factor are 8.5% and 5.0%.

See Figure B1 for the variation across portfolios sorted jointly across all stocks (not industry-neutral) in

market-factor adjusted returns.

C. Heterogeneity in returns to Generative AI exposure

The results in the previous section showed our key result that, on average, a higher

exposure to Generative AI technology advances due to a prevalence of aﬀected tasks in a ﬁrm

was associated with higher realized stock returns after the release of ChatGPT. However,

there is reason to believe that these eﬀects might vary in size, and perhaps also in sign,

across diﬀerent industries. On the one hand, our exposure measure focuses on the labor

cost and productivity dimensions of the impact of these technology shocks. However, there

will likely also be simultaneous eﬀects on product market competitiveness for diﬀerent ﬁrms,

depending on whether their product complements, or substitutes for, the capabilities of

products like ChatGPT for end-consumers. Similarly, some companies may have valuable

intangible capital that is made obsolete – e.g. because ChatGPT obviates the productivity

advantage of proprietary internal software and processes – or increases in value, for instance

if the company controls access to data sources that can be more eﬀectively analyzed in

conjunction with large language model capabilities.

In addition, industry structure and contestability might aﬀect the impact of higher Gen-

erative AI exposure on publicly traded incumbents. If entry into the product market becomes

easier as Generative AI technology lowers the cost of putting together minimum viable prod-

ucts, or if start-ups in an industry are more likely to have the organizational ﬂexibility to

quickly incorporate these new technologies into their workﬂows, then incumbents with high

exposure to Generative AI technologies may lose out in relative terms, even if the overall

impact on the industry’s productivity is positive. While it is too early to make deﬁnitive

statements about which ﬁrms will be winners or losers as a result of recent technology ad-

vances, this section will provide early suggestive evidence of what industry characteristics

have been associated with higher return eﬀects.

Industry portfolio heterogeneity in ChatGPT release period returns One way

of looking at industry heterogeneity is to consider which industry portfolios exhibit returns

after the release of ChatGPT that exceed or fall short of that predicted by their Generative

AI exposure. Figure 11 shows the variation in market factor-adjusted excess realized returns

of 3-digit industry portfolios either during the two weeks after the ChatGPT release (upper

panel), or on all other days of the Nov 15, 2022 - March 31, 2023, period (lower panel).

There is a positive association across industries between average industry level Generative

AI exposure and returns during the weeks after the ChatGPT release. Moreover, as the

lower panel conﬁrms, this is not due to the fact that high Generative AI exposure industries

incidentally experienced systematically higher or lower returns over the nearly 5 months in

our study: the average excess daily returns of almost all 3-digit industry portfolios are small

and do not vary with Generative AI exposure over this period. That is, the Generative AI

-related returns eﬀect is evident at the industry level only on the event days when ChatGPT

news is most likely being incorporated into valuations.

However, Figure 11 also shows that some industries with high Generative AI exposure

perform even better on news days than we would expect based on a simple linear relationship

between exposure and returns, i.e. their news day returns lie above the red line. For example,

the highest returns on news days among large subsectors accrue to ﬁrms in the industry

subsector “Publishing industry (except internet)”, in which the largest ﬁrms by market cap

are Microsoft, Salesforce, and Intuit. While Microsoft is directly associated with the success

of ChatGPT and related products through its investment in Open AI, the company releasing

this tool, Salesforce and Intuit operate software businesses focused on customer relationships,

tax, personal ﬁnance, and accounting data. It is intuitive that better analytical tools based

on Generative AI can, for instance, make the proprietary data resulting from these business

lines more valuable.

Within-industry Generative AI exposure eﬀects While there is substantial variation

across industries in their overall exposure and performance after the release of ChatGPT,

there may also be diﬀerences in the degree to which high-exposure ﬁrms outperform low-

exposure ﬁrms relative to other ﬁrms within the same industry. That is, in some industries

the ability to convert relatively higher exposure to the technology shock into higher ﬁrm-

level returns may be better than in others, independent of whether the industry as a whole is

beneﬁting from higher Generative AI exposure or not. Moreover, exploring how the within-

industry outperformance varies also provides evidence for how widespread the Generative AI

technology eﬀects are outside of the “Information” sector that is most likely to see impacts

distorted by simultaneous product market eﬀects.

Figures 12 and 13 show the average daily return alpha during the ChatGPT release period

for H-L portfolios within each sector or subsector, respectively, for quintiles at the NAICS

2-digit sector and terciles at the 3-digit subsector level.

Although the standard errors for these within-industry portfolio regressions are necessar-

ily larger, as some industries have a limited number of ﬁrms in the sample, we do indeed ﬁnd

that the Generative AI exposure eﬀect is signiﬁcantly diﬀerent from zero in several sectors

and subsectors. Focusing on the more detailed categories in Figure 13, we ﬁnd signiﬁcant

Each graph only includes sector or subsectors that have (1) a large enough sample of ﬁrms in the data,

consisting of at least 10 ﬁrms across the high and low exposure quantile, and (2) a large enough Generative

AI exposure spread between the high and low exposure portfolio, with cutoﬀs set at a 10 ppt spread between

the high and low quintiles for 2-digit sectors, and at 5 ppt between the high and low tercile for 3-digit

subsectors.

positive returns within ‘credit intermediation & related activities”, which contains several

large banks and brokerages, “publishing industries (except internet)”, which contains Mi-

crosoft, Salesforce, and Intuit, as noted above, and also the “administrative and support

services” subsectors. In contrast, the “real estate”, “other information services”, and “food

manufacturing” subsectors” all show negative within-industry returns to higher Generative

AI exposure. Contrasting this with non-release period within-industry returns shown in

Appendix Figures B5 and B6, we can see that the latter are more precisely estimated and

small, as well as not signiﬁcant with few exceptions. This suggests that the variation in

within-industry returns to Generative AI exposure are likely to be driven by information in-

corporated during the ChatGPT release period, rather than generally during our 4.5 month

sample period.

Industry characteristics and within-industry Generative AI exposure eﬀects The

previous analysis raises the question of what industry characteristics can explain the observed

heterogeneity in whether industries are more likely to exhibit a high return on the within-

industry AMH exposure portfolio. We explore this question by running regressions of the

form

news,AMH

= η + γCharacteristic

+ ε

which relate the subsector AMH Generative AI exposure portfolio alpha to various industry

characteristics that are computed as either the average of the characteristics for the industry

(left panel of Figure 14), or the diﬀerence between the high and low Generative AI exposure

portfolios in the characteristic within that industry (right panel of Figure 14). In either case,

both the dependent and independent variables are transformed into standardized Z-scores,

such that coeﬃcients capture the relationship between a change in the characteristic by one

standard deviation and a standardized change in within-industry returns to Generative AI

exposure.

As shown in Figure 14, industry subsectors with larger ﬁrms are more likely to have high

within-industry returns to diﬀerences in Generative AI exposure. Moreover, high Generative

AI exposure ﬁrms are more likely to outperform the low exposure ﬁrms in the same industry,

if the former have higher ROE, ROA, market capitalization, gross proﬁtability, or organi-

zational capital. However, these associations are only suggestive: the precise characteristics

that allow some ﬁrms to take better advantage of language model-driven technological ad-

vances will hopefully be considered in future studies.

V. Conclusion

Market prices indicate that the arrival and diﬀusion of large language models and Gen-

erative AI represent a major technology shock with important eﬀects on the overall value

of ﬁrms, as well leading to winners and losers. This paper uses occupational exposures to

Generative AI, along with ﬁrm-level measures of occupational composition, to assess the ex-

posure to Generative AI innovations at the ﬁrm level for publicly traded U.S. corporations.

We ﬁnd that the eﬀect of the release of ChatGPT on ﬁrm values was large, driving a diﬀer-

ence in ﬁrm returns of approximately .4% daily, translating to over 100% on an annualized

basis. These diﬀerences were realized both within and across industries, and display wide

variation which is correlated with ﬁrm characteristics such as organizational capital or gross

proﬁtability. According to investors, ChatGPT represents an important shock to corporate

valuations.

REFERENCES

Acemoglu, Daron, and David Autor, 2011, Skills, tasks and technologies: Implications for

employment and earnings, in Handbook of labor economics, volume 4, 1043–1171 (Elsevier).

Akcigit, Ufuk, and Sina T Ates, 2020, Slowing business dynamism and productivity growth

in the united states, in Kansas City Federal Reserve. Paper prepared for the 2020 Jackson

Hole conference.

Autor, David H, Frank Levy, and Richard J Murnane, 2003, The skill content of recent

technological change: An empirical exploration, The Quarterly journal of economics 118,

1279–1333.

Babina, Tania, Anastassia Fedyk, Alex He, and James Hodson, 2021, Artiﬁcial intelligence,

ﬁrm growth, and product innovation, Firm Growth, and Product Innovation (November

9, 2021) .

Babina, Tania, Anastassia Fedyk, Alex Xi He, and James Hodson, 2020, Artiﬁcial intelli-

gence, ﬁrm growth, and industry concentration, Firm Growth, and Industry Concentration

(November 22, 2020.

Babina, Tania, Anastassia Fedyk, Alex Xi He, and James Hodson, 2022, Firm investments

in artiﬁcial intelligence technologies and changes in workforce composition, Available at

SSRN .

Brynjolfsson, Erik, Danielle Li, and Lindsey R Raymond, 2023, Generative ai at work,

Technical report, National Bureau of Economic Research.

Bubeck, S´ebastien, Varun Chandrasekaran, Ronen Eldan, Johannes Gehrke, Eric Horvitz,

Ece Kamar, Peter Lee, Yin Tat Lee, Yuanzhi Li, Scott Lundberg, et al., 2023, Sparks of ar-

tiﬁcial general intelligence: Early experiments with gpt-4, arXiv preprint arXiv:2303.12712

Donangelo, Andres, 2014, Labor mobility: Implications for asset pricing, The Journal of

Finance 69, 1321–1346.

Eisfeldt, Andrea L., Antonio Falato, and Mindy Xiaolan, 2022, Human capitalists, in Mar-

tin Eichenbaum, Erik Hurst, and Valery Ramey, eds., NBER Macroeconomics Annual,

chapter 1, 2 (University of Chicago Press, Chicago).

Eisfeldt, Andrea L., and Dimitris Papanikolaou, 2013, Organization capital and the cross-

section of expected returns, The Journal of Finance 68, 1365 – 1406.

Eisfeldt, Andrea L., and Dimitris Papanikolaou, 2014, The value and ownership of intangible

capital, American Economic Review 104, 189–94.

Eloundou, Tyna, Sam Manning, Pamela Mishkin, and Daniel Rock, 2023, Gpts are gpts: An

early look at the labor market impact potential of large language models, arXiv preprint

arXiv:2303.10130 .

Fama, Eugene F, and Kenneth R French, 2015, A ﬁve-factor asset pricing model, Journal of

ﬁnancial economics 116, 1–22.

Felten, Edward W, Manav Raj, and Robert Seamans, 2023, Occupational heterogeneity in

exposure to generative ai, Available at SSRN 4414065 .

Gilardi, Fabrizio, Meysam Alizadeh, and Ma¨el Kubli, 2023, Chatgpt outperforms crowd-

workers for text-annotation tasks, arXiv preprint arXiv:2303.15056 .

Greenwood, Jeremy, Zvi Hercowitz, and Per Krusell, 1997, Long-run implications of

investment-speciﬁc technological change, The American economic review 342–362.

Guti´errez, Germ´an, and Thomas Philippon, 2019, The failure of free entry, Technical report.

Hansen, Stephen, Peter John Lambert, Nicholas Bloom, Steven J Davis, Raﬀaella Sadun,

and Bledi Taska, 2023, Remote work across jobs, companies, and space, Technical report,

National Bureau of Economic Research.

Hassan, Tarek A, Stephan Hollander, Laurence Van Lent, and Ahmed Tahoun, 2019, Firm-

level political risk: Measurement and eﬀects, The Quarterly Journal of Economics 134,

2135–2202.

Kelly, Bryan, Dimitris Papanikolaou, Amit Seru, and Matt Taddy, 2021, Measuring techno-

logical innovation over the long run, American Economic Review: Insights 3, 303–320.

Kogan, Leonid, and Dimitris Papanikolaou, 2014, Growth opportunities, technology shocks,

and asset prices, The journal of ﬁnance 69, 675–718.

Kogan, Leonid, Dimitris Papanikolaou, Lawrence Schmidt, and Bryan Seegmiller, 2019,

Technology-skill complementarity and labor displacement: Evidence from linking two cen-

turies of patents with occupations, Available at SSRN 3585676 .

Krusell, Per, Lee E Ohanian, Jos´e-V´ıctor R´ıos-Rull, and Giovanni L Violante, 2000, Capital-

skill complementarity and inequality: A macroeconomic analysis, Econometrica 68, 1029–

1053.

Lane, Marguerita, and Anne Saint-Martin, 2021, The impact of artiﬁcial intelligence on the

labour market: What do we know so far? .

Noy, Shakked, and Whitney Zhang, 2023, Experimental evidence on the productivity eﬀects

of generative artiﬁcial intelligence, Available at SSRN 4375283 .

Papanikolaou, Dimitris, 2011, Investment shocks and asset prices, Journal of Political Econ-

omy 119, 639–685.

Underwood, Ted, 2023, Using gpt-4 to measure the pas-

sage of time in ﬁction, https://tedunderwood.com/2023/03/19/

using-gpt-4-to-measure-the-passage-of-time-in-fiction/.

Webb, Michael, 2019, The impact of artiﬁcial intelligence on the labor market, Available at

SSRN 3482150 .

Willison, Simon, 2023, A simple python wrapper for the chatgpt api, https://til.

simonwillison.net/gpt3/chatgpt-api.

Zhang, Miao Ben, 2019, Labor-technology substitution: Implications for asset pricing, The

Journal of Finance 74, 1793–1839.

Figure 2: Generative AI exposure and wages by major occupation group and

industry sectors.

(A) Major occupation groups

(B) NAICS 2-digit industry sectors

Figure 3: Occupational skills and Generative AI exposure. The graph below shows

the results of regressing our 6-digit occupation’s Generative AI exposure measure jointly

on a set of occupational skill scores deﬁned in Acemoglu and Autor (2011), converted into

standard z-scores. That is, we run the following regression:

Exposure

GP T

= α +

Skill

+ ε

The regression sample contains 690 occupations. The bars around each coeﬃcient show 95%

conﬁdence intervals based on heteroskedasticity-robust standard errors.

Figure 4: Generative AI exposure Across and Within Industries This ﬁgure plots

the average and the standard deviation of Compustat ﬁrms’ Generative AI exposure within each

NAICS 2-digit sector.

0.15

0.20

0.25

0.30

0.35

0.40

0.45

0.50

Accommodation and Food Services

Retail Trade

Construction

Mining, Quarrying, and Oil and Gas Extraction

Transportation and Warehousing

Other Services (except Public Administration)

Arts, Entertainment, and Recreation

Health Care and Social Assistance

Utilities

Manufacturing

Real Estate and Rental and Leasing

Finance and Insurance

Wholesale Trade

Educational Services

Agriculture, Forestry, Fishing and Hunting

Administrative and Support and Waste Management and

Remediation Services

Information

Professional, Scientific, and Technical Services

Figure 5: Generative AI exposure and average ﬁrm characteristics by industry

sector. The graphs show the market capitalization weighted average of the characteristic

among ﬁrms in our Generative AI exposure sample in each NAICS sector, and plots against

it the weighted average of our Generative AI exposure measure in the same sector. The

ﬁrm characteristics are based on Compustat ﬁscal year 2021 data., and the wages are from

the Bureau of Labor Statistics. The red line is the market-capitalization-weighted linear ﬁt

across the sector level aggregates.

Figure 6: Topic mentions in company earnings conference calls. The graphs below

show data on the share of calls mentioning Generative AI related words (Panels A and B) or

the average number of such mentions per call across the sample (Panels C and D). The data

set is a manually collected panel of earnings conference call transcripts for S&P 500 ﬁrms’

tickers from the Seeking Alpha website. The graphs show monthly statistics for calls that

were held from Jan 2019 to March 2023. Calls are assigned a calendar month, quarter and

year based on the time stamp at the beginning of the transcript. Each transcript is converted

into a list of lower-case unigrams and bigrams. Word tokens in these lists are counted as

Generative AI-speciﬁc words if they are in the following set: “llm”, “chatgpt”, “gpt”, “gpt3”,

“gpt4”, “generative”, “language model”. Trends in other machine learning-related words and

generic engineering terms are shown for comparison - see text for details. in Panels B and

D, the sample of ﬁrms excludes the “Information” (NAICS 51) and Professional, Scientiﬁc,

and Technical Services (NAICS 54) sectors, as determined by Compustat industry codes.

(A) Share of calls: all ﬁrms

(B) Share of calls: excl. Information &

Profess., Scient., and Tech. Svcs. sectors

(D) Mentions per call: excl. Information &

Profess., Scient., and Tech. Svcs. sectors

Figure 7: Firm-level Generative AI exposure and topics in company earnings

conference calls: Each graph show the result of estimating regression speciﬁcations of the

form

[Topic X]

i,t

= α

+ β

+ γ [Topic X]

i,2019

+ ε

i,t

for each topic for the ﬁscal quarters 2019 Q1-2023 Q1 for a manually collected panel of

earnings conference call transcripts for S&P 500 ﬁrms’ tickers from the Seeking Alpha website.

Calls are assigned a calendar month, quarter and year based on the time stamp at the

beginning of the transcript. Each transcript is converted into a list of lower-case unigrams

and bigrams. Word tokens in these lists are counted as LLM/Generative AI-speciﬁc words

if they are in the following set: “llm”, “chatgpt”, “gpt”, “gpt3”, “gpt4”, “generative”,

“language model”. The association of Generative AI exposure with generic engineering

terms is shown for comparison - see text for details of other topic deﬁnitions. in Panels

B and D, the sample of ﬁrms excludes the “Information” (NAICS 51) and Professional,

Scientiﬁc, and Technical Services (NAICS 54) sectors, as determined by Compustat industry

codes. Quarterly sample sizes for the regressions are 341-416 for the full sample, and 299-368

ﬁrms for the sample exluding NAICS 51 and 54. Dotted lines show 95% conﬁdence intervals

based on heteroskedasticity-robust standard errors.

(A) LLM/Gen. AI: all ﬁrms

(B) LLM/Gen. AI: excl. Information &

Profess., Scient., and Tech. Svcs. sectors

(D) Engineering: excl. Information & Profess.,

Scient., and Tech. Svcs. sectors

Figure 8: Stock return volatility, Generative AI exposure, and social media

attention: Panel A shows the t-statistic for the hypothesis β

= 0 in the regression |r|

+α

ind

+β

GenAIExposure

+ε

, estimated at the ﬁrm level, for each trading day from Nov.

15, 2022, to March 31, 2023. Panel shows the estimates without, and Panel B including, 2-

digit NAICS sector ﬁxed eﬀects. The dependent variable is the absolute value of daily stock

returns, and the independent variable is the ﬁrm-level measure of task exposure to Generative

AI technology productivity changes. Standard errors are heteroskedasticity-robust. The

dashed horizontal line indicates where t-statistics exceed 2. Panel C shows the total count

of Twitter mentions of “ChatGPT” or “GPT” reported by Media Cloud, and Panel D shows

the growth rate (using the average base method) of these Twitter mentions.

(A) Exposure & stock return volatility: no

industry FEs

(B) Exposure & stock return volatility: incl.

industry FEs

Figure 9: Industry-neutral Generative AI exposure quintile portfolio returns: market-factor

adjusted. Each bar shows the average daily return alpha during the ChatGPT release period and also

for all days, for portfolios based on industry-neutral Generative AI exposure sorts, and also for the net

zero investment high-minus-low exposure portfolio that represents the “Artiﬁcial Minus Human” (AMH )

factor. Industry-neutral portfolios are computed by ﬁrst forming within-industry value-weighted quintile

portfolios, and then averaging portfolio returns for the same quintiles across industries. Returns within-

industry quintiles are value-weighted, while across-industry averages are industry market-cap. weighted.

The data set consists of daily stock returns from Yahoo Finance for Nov. 15, 2022 - March 31, 2023, and

factors from Ken French’s website. The ﬁgure shows alphas estimated from portfolio-level regressions of the

form

−r

= α

release

[ChatGPT release period]

+α

not release

[Not ChatGPT release period]

+β

(Mkt

−r

)+ε

where the bars show either the intercept for the “ChatGPT release period”, or the intercept from a regression

where all days have the same intercept. Error bars indicate 95% conﬁdence intervals computed using Newey

West standard errors with ﬁve lags.

Figure 10: Generative AI exposure industry-neutral quintile portfolio returns over time:

market factor-adjusted. The graph shows the cumulative excess realized returns on portfolios based on

industry-neutral Generative AI exposure sorts. Industry-neutral portfolios are computed by ﬁrst forming

within-industry value-weighted quintile portfolios, and then averaging portfolio returns for the same quintiles

across industries. Returns within-industry quintiles are value-weighted, while across-industry averages are

industry market-cap. weighted. All portfolio returns shown are net of the risk free rate. The data set consists

of daily stock returns from Yahoo Finance for Nov. 15, 2022 - March 31, 2023. The ﬁgure shows returns

adjusted for market factor exposure.

Figure 11: Industry Portfolio Stock Returns in ChatGPT Release Period. Each graph shows

the average daily return alpha for the 3-digit NAICS industry portfolios corresponding to the 50 largest

subsectors. The data set consists of daily stock returns from Yahoo Finance for Nov. 15, 2022 -March 31,

2023. The table shows market-factor adjusted alphas estimated from regressions of the form

−r

= α

release

[ChatGPT release period]

+α

not release

[Not ChatGPT release period]

+β

(Mkt

−r

)+ε

where the intercept is allowed to vary with whether the day is part of the “ChatGPT release period” (Nov.

30, 2022, and the following two weeks), as deﬁned in the text, or not. The upper panel shows the excess

daily returns estimates for the GPT release period, while the lower panel shows excess returns on all other

days for the Nov. 15, 2022 to March 31 2023, period. The ten largest 3-digit industry subsectors by market

capitalization are indicated in red and labeled. Grey lines indicate 95% conﬁdence intervals computed using

Newey West standard errors with ﬁve lags.

Figure 12: Within-Sector AMH Generative AI exposure Portfolio Realized Returns: ChatGPT Release

Period. The graph shows the average daily return alpha for the ChatGPT release period for AMH Generative AI exposure

portfolios within each industry, at the 2-digit NAICS sector level Each AMH portfolio is formed by taking the value-weighted

highest and lowest quintiles of Generative AI exposure within each industry (based on NYSE stock cutoﬀs) and forming zero

net investment AMH portfolio returns as the equal-weighted diﬀerence in the daily realized returns between these portfolios,

and then subtracting the daily risk-free return. The industries shown omit any sectors with fewer than 10 ﬁrms combined in

the highest and lowest quantiles in the sample, as well as sectors with less than a 10 ppt Generative AI exposure spread at the

sector level between the high and low quintile. The data set consists of daily stock returns from Yahoo Finance for Nov. 15,

2022 - March 31, 2023. The graphs show market-factor adjusted alphas estimated from regressions of the form

− r

= α

release

[ChatGPT release period]

+ α

not release

[Not ChatGPT release period]

+ β

(Mkt

− r

) + ε

where the intercept is allowed to vary with whether the day is in the release period deﬁned by Nov. 30, 2022, and the following

two weeks, or is one of the other trading days in the sample. The returns are shown in units of average daily excess realized

returns (controlling for the market factor). Red error bars indicate 95% conﬁdence intervals computed using Newey-West

standard errors with ﬁve lags.

Figure 13: Within-Subsector AMH Generative AI exposure Portfolio Realized Returns: ChatGPT Release

Period. Each graph shows the average daily return alpha for the ChatGPT release period for AMH Generative AI exposure

portfolios within each industry, at the 3-digit subsector level. Each AMH portfolio is formed by taking the value-weighted

highest and lowest terciles of Generative AI exposure within each industry (based on NYSE stock cutoﬀs) and forming zero

net investment H-L portfolio returns as the equal-weighted diﬀerence in the daily realized returns between these portfolios, and

then subtracting the daily risk-free return. The industries shown omit any subsectors with fewer than 10 ﬁrms combined in the

highest and lowest quantiles in the sample, as well as subsectors with less than a 5 ppt Generative AI exposure spread between

the high and low tercile. The data set consists of daily stock returns from Yahoo Finance for Nov. 15, 2022 - March 31, 2023.

The graphs show market-factor adjusted alphas estimated from regressions of the form

− r

= α

release

[ChatGPT release period]

+ α

not release

[Not ChatGPT release period]

+ β

(Mkt

− r

) + ε

where the intercept is allowed to vary with whether the day is in the release period deﬁned by Nov. 30, 2022, and the following

two weeks, or is one of the other trading days in the sample. The returns are shown in units of average daily excess realized

returns (controlling for the market factor). Red error bars indicate 95% conﬁdence intervals computed using Newey-West

standard errors with ﬁve lags.

Figure 14: Industry Characteristics and Heterogeneity in Within-Industry Returns to Gen-

erative AI exposure. Each graph shows the result of a regression where the dependent variable is the

average daily return alpha during the ChatGPT release period on the AMH tercile portfolio within each

industry subsector, corresponding to the estimates shown in Figure 13. The independent variable is either

the average of the characteristics for the industry (left panel), or the diﬀerence between the high and low

Generative AI exposure portfolios in the characteristic within that industry, in either case transformed into

standardized Z-scores. Each coeﬃcient in the graph represents an estimate of ˆγ from a regression of the form

release

= η + γCharacteristic

+ ε

where z

release

is the within-industry H-L Generative AI exposure portfolio return for each 3-digit in-

dustry subsector i as shown in Figure 13, transformed into a standardized Z-score. Each characteristic

Characteristic

has also been transformed into a standard Z-score. Only subsectors with a total of at least

10 ﬁrms in the high and low tercile are included, as well as with a Generative AI exposure spread of at

least 5 ppt between the highest and lowest tercile. Depending on the characteristic, the sample size varies

from 19 and 22 subsectors for regressions involving R& D or organizational capital to 26 subsectors for

other characteristics. Error bars indicate 95% conﬁdence intervals computed using heteroskedasticity-robust

standard errors.

Table I: Summary statistics for selected ﬁrm characteristics. Generative AI exposure

is our bottom-up task-based measure of occupational exposure, aggregated to the ﬁrm-level

based on the ﬁrm’s occupational employment structure. Log Size is the natural logarithm of

total assets. Labor Intensity is the logarithm of the ratio of Property, Plant & Equipment

(PP&E) to Total Assets, based on Donangelo (2014). Tangibility is the ratio of PP& E to

Total Assets. Org. Capital Ratio is the ratio of the organizational capital stock from Eisfeldt

and Papanikolaou (2013) divided by Total Assets.

Measure Mean Std. Dev. p10 p50 p90 Obs.

Generative AI Exposure 0.354 0.078 0.268 0.353 0.442 2,518

Log Size 1.876 2.381 -1.091 2.063 4.636 2,517

Tobin’s Q 3.667 10.081 1.354 2.176 5.832 2,380

ROA -0.011 0.891 -0.181 0.093 0.216 2,513

Labor Intensity 0.761 1.761 -2.157 1.126 2.512 2,387

Org. Capital Ratio 1.190 4.416 0.137 0.712 2.179 1,571

Tangibility 0.301 0.266 0.036 0.198 0.758 2,515

Table II: Major U.S. Firms with the Highest and Lowest Exposure to GPT. This

table shows the 15 large U.S. publicly traded ﬁrms with the highest employee exposure to

ChatGPT in Panel A and the 15 ﬁrms with the lowest exposure in Panel B. We select the

large U.S. publicly traded ﬁrms as the top 100 ﬁrms with the largest market capitalization

as of November 1, 2022, which also have headquarters in the U.S. Generative AI exposure is

the ﬁrm’s labor exposure to ChatGPT-like technologies deﬁned in Section I. MktCap is the

ﬁrm’s market capitalization as of November 1, 2022, in B. Sector is deﬁned at the NAICS

2-digit level.

Panel A: Top 15 Large U.S. Companies with Highest Exposure to ChatGPT

Company Name Generative AI exposure MktCap Sector

Int. Business Machines Corp 0.488 125 Information

Intuit Inc. 0.480 111 Information

QUALCOMM Inc. 0.479 132 Manufacturing

Fiserv Inc. 0.475 66 Information

NVIDIA Corporation 0.468 337 Manufacturing

S&P Global Inc 0.452 103 Admin. & Support Services

Broadcom Inc 0.449 195 Manufacturing

Verizon Communications Inc 0.444 157 Information

Microsoft Corp 0.442 1,700 Information

3M Co 0.442 69 Manufacturing

Advanced Micro Devices Inc 0.441 96 Manufacturing

ServiceNow Inc 0.434 85 Information

Adobe Inc 0.427 147 Information

PayPal Holdings Inc 0.418 96 Information

Thermo Fisher Scientiﬁc Inc 0.411 203 Manufacturing

Panel B: Bottom 15 Large U.S. Companies with Lowest Exposure to ChatGPT

Company Name Generative AI exposure MktCap Sector

Starbucks Corp 0.119 100 Accommodation & Food Svcs

McDonald’s Corp 0.194 201 Accommodation & Food Svcs

Dollar General Corporation 0.212 57 Retail Trade

Target Corp 0.235 76 Retail Trade

Walmart Inc 0.235 385 Retail Trade

Lowe’s Cos Inc 0.238 120 Retail Trade

TJX Companies Inc 0.243 83 Retail Trade

Costco Wholesale Corp 0.252 221 Retail Trade

Union Paciﬁc Corp 0.253 121 Transportation & Warehousing

CSX Corp 0.256 61 Transportation & Warehousing

United Parcel Service Inc 0.256 123 Transportation & Warehousing

Home Depot Inc 0.261 303 Retail Trade

Tesla Inc 0.283 719 Manufacturing

Northrop Grumman Corp 0.291 83 Manufacturing

Mondelez International Inc 0.292 85 Manufacturing

Table III: Firm Generative AI exposure and Firm Characteristics This table re-

gresses our ﬁrms’ Generative AI exposure measure on ﬁrm characteristics using the cross-

section of U.S. publicly traded ﬁrms in 2022. See Table I for variable deﬁnitions. Panel B

controls for ﬁxed eﬀects at the NAICS 2-digit level. Standard errors are clustered at the

industry level and reported in parentheses.

∗

∗∗

, and

∗∗∗

indicate signiﬁcance at the 10%,

5%, and 1% level, respectively.

Panel A: Across All Firms

(1) (2) (3) (4) (5) (6)

Log Size -2.653

∗∗

(1.230)

Tobin’s Q 3.076

∗∗

(1.217)

ROA -21.516

∗

(11.882)

Labor Intensity 7.892

∗∗

(2.785)

Org. Capital Ratio 9.139

∗∗∗

(3.081)

Tangibility -89.931

∗∗∗

(22.426)

Observations 2517 2380 2513 2387 1571 2515

Adjusted R

0.006 0.013 0.006 0.038 0.022 0.107

Panel B: Within-Industry

(1) (2) (3) (4) (5) (6)

Log Size -2.101

∗∗∗

(0.665)

Tobin’s Q 1.499

∗∗∗

(0.490)

ROA -13.850

(8.733)

Labor Intensity 5.476

∗∗

(2.586)

Org. Capital Ratio 3.742

∗∗

(1.668)

Tangibility -72.084

∗∗

(25.866)

Observations 2517 2380 2513 2387 1571 2515

Adjusted R

0.200 0.210 0.197 0.227 0.211 0.227

Table IV: Firm-level eﬀects of Generative AI exposure on earnings call topic

mentions. The table shows ﬁrm-quarter panel regressions over the period of Q3 2018 - Q1

2023 of the form

[Gen. AI Topic]

i,t

= α

+ α

+ β

× [Post-ChatGPT] + γ [Post-ChatGPT] + ε

i,t

where we regress whether a ﬁrm mentions Generative AI related words in an earnings con-

ference call in that quarter on the measure of ﬁrm exposure to Generative AI technologies

and its interactions with an indicator of whether the quarter is “post-ChatGPT”, which

corresponds to Q4 2022 and Q1 2023 in our sample. The ﬁrm sample consists of all S&P 500

ﬁrms for which we were able to collect earnings call transcripts in columns 1-4. In column 5,

we exclude the “Information” (NAICS 51) and “Professional, Scientiﬁc, and Technical Ser-

vices” (NAICS 54) sectors, as determined by Compustat industry codes. The panel includes

437 ﬁrms for the most restrictive speciﬁcation with all ﬁrms, and 385 ﬁrms when excluding

NAICS 51 and 54. T-statistics in parentheses based on standard errors clustered at the

NAICS sector level.

Dep. var.: [Generative AI Topic Mentioned]

i,t

(1) (2) (3) (4) (5)

[Post-ChatGPT]

× Gen. AI Exposure

0.660** 0.659** 0.644** 0.761** 0.310**

(2.307) (2.303) (2.319) (2.407) (2.474)

R-squared 0.01 0.01 0.04 0.27 0.19

Observations 6235 6235 6235 6228 5502

Fixed Eﬀects & Controls

Quarter FEs X X X X

Industry Sector FEs X

Firm FEs X X

Excl. NAICS 51 and 54 X

Table V: Realized returns of portfolios sorted on Generative AI exposure after ChatGPT

release. This table reports daily excess stock returns of value-weighted portfolios of ﬁrms sorted on Gener-

ative AI exposure. AMH is the ”Artiﬁcial Minus Human” zero net investment portfolio long high exposure

(H) stocks and short low exposure (L) stocks. Quintile thresholds that deﬁne value-weighted portfolios are

solely based on the sample of stocks listed on NYSE as of the sorting date. All quintile portfolios are formed

based on value weights on October 31, 2022, and weights are adjusted based on daily returns to mimic

passive buy-and-hold exposure. Industry-neutral portfolios are computed by ﬁrst forming within-industry

value-weighted quintile portfolios, and then averaging portfolio returns for the same quintiles across indus-

tries. Returns for within-industry quintiles are value-weighted, while across-industry averages are industry

market-cap. weighted. The data set consists of daily stock returns from Yahoo Finance for Nov. 15, 2022 -

March 31, 2023. The market factor and risk free returns are obtained from Ken French’s website. The table

shows alphas estimated from regressions of the form

−r

= α

release

[ChatGPT release period]

+α

not release

[Not ChatGPT release period]

+β

(Mkt

−r

)+ε

where the intercept is either for the full sample, or is allowed to vary with whether the day is in the ChatGPT

release period consisting of Nov 30, 2022 - Dec. 14, 2022. Panel A does not include the market factor. T-

statistics in parentheses are computed using Newey-West standard errors with ﬁve lags.

Portfolios

Sample Q1 Q2 Q3 Q4 Q5 AMH

A: Excess returns (%)

All days 0.042 0.027 0.050 0.037 0.134 0.076

(0.38) (0.25) (0.54) (0.31) (1.17) (1.33)

Not ChatGPT release period 0.066 0.030 0.038 0.037 0.110 0.028

(0.57) (0.25) (0.36) (0.28) (0.85) (0.44)

ChatGPT release period -0.137 0.000 0.137 0.034 0.316 0.437

(-0.53) (0.00) (0.43) (0.11) (0.82) (3.08)

B: Market factor-adjusted alpha (%)

All days 0.014 -0.005 0.020 0.003 0.098 0.068

(0.38) (-0.15) (0.56) (0.07) (2.50) (1.11)

Not ChatGPT release period 0.042 0.003 0.013 0.008 0.079 0.021

(1.07) (0.08) (0.32) (0.17) (1.83) (0.32)

ChatGPT release period -0.196 -0.066 0.076 -0.036 0.240 0.420

(-4.88) (-0.55) (0.82) (-0.63) (5.29) (5.42)

C: Ind.-neutral mkt. factor-adjusted alpha (%)

All days -0.003 0.000 0.011 0.005 0.065 0.052

(-0.07) (0.01) (0.54) (0.24) (2.60) (0.95)

Not ChatGPT release period 0.031 0.013 -0.001 -0.008 0.050 0.002

(0.69) (0.26) (-0.03) (-0.41) (1.86) (0.05)

ChatGPT release period -0.257 -0.091 0.100 0.106 0.182 0.423

(-5.32) (-2.82) (4.78) (4.37) (5.88) (6.20)

Appendix A. Appendix: Methodology Notes

Generative AI exposure portfolio construction.

Portfolios for the main realized return analysis are formed from quintiles of stocks that

have Yahoo Finance data for Nov. 15, 2022 - March 31, 2023. Quintile thresholds that

deﬁne value-weighted portfolios within industries or for all stocks are solely based on the

sample of stocks listed on NYSE as of the sorting date. All portfolios are formed based on

value weights on October 31, 2022, and weights are adjusted based on daily returns to mimic

passive buy-and-hold exposure. Industry-neutral portfolios are computed by ﬁrst forming

within-industry value-weighted quintile portfolios, and then averaging portfolio returns for

the same quintiles across industries. Returns for within-industry quintiles and for all global

(not industry-neutral) portfolio sorts are value-weighted, while across-industry averages are

industry market-cap. weighted. AMH is the ”Artiﬁcial Minus Human” (also referred to as

H-L) is the zero net investment portfolio long highest exposure quintile (H) stocks and short

lowest exposure quintile (L) stocks.

GPT prompt for exposure scoring

The following prompt structure was based on the rubric language by Eloundou et al.

(2023), as well as insights by Willison (2023) and Underwood (2023) about how to best

structure API calls for GPT classiﬁcation. Here are the instruction prompts submitted be-

fore asking GPT 3.5 Turbo to classify each task statement (using the version as of March

28th, 2023). Note that the order in which the two user-assistant interactions are provided

to the API is randomized for each task, and the GPT temperature is set to 0:

systemprompt = ”Consider the most powerful OpenAI large language model (LLM).

This model can complete many tasks that can be formulated as having text input and text

output where the context for the input can be captured in 2000 words. The model also

cannot draw up-to-date facts (those from <1 year ago) unless they are captured in the

input. Assume you are a worker with an average level of expertise in your role trying to

complete the given task. You have access to the LLM as well as any other existing software

or computer hardware tools mentioned in the task. You also have access to any commonly

available technical tools accessible via a laptop (e.g. a microphone, speakers, etc.). You do

not have access to any other physical tools or materials. You are a helpful research assistant

who wants to label the given tasks according to the rubric below. Equivalent quality means

someone reviewing the work would not be able to tell whether a human completed it on their

own or with assistance from the LLM. If you aren’t sure how to judge the amount of time a

task takes, consider whether the tools described exposed the majority of subtasks associated

with the task.

# Exposure rubric:

## E1 - Direct exposure: Label tasks E1 if direct access to the LLM through an interface

like ChatGPT or the OpenAI playground alone can reduce the time it takes to complete

the task with equivalent quality by at least half. This includes tasks that can be reduced

to: - Writing and transforming text and code according to complex instructions, - Providing

edits to existing text or code following speciﬁcations, - Writing code that can help perform

a task that used to be done by hand, - Translating text between languages, - Summarizing

medium-length documents, - Providing feedback on documents, - Answering questions about

a document, - Generating questions a user might want to ask about a document, - Writing

questions for an interview or assessment, - Writing and responding to emails, including ones

that involve refuting information or engaging in a negotiation (but only if the negotiation is

via written correspondence), - Maintain records of written data, - Prepare training materials

based on general knowledge, or - Inform anyone of any information via any written or spoken

medium.

## E2 - Exposure by LLM-powered applications: Label tasks E2 if having access to the

LLM alone may not reduce the time it takes to complete the task by at least half, but

it is easy to imagine additional software that could be developed on top of the LLM that

would reduce the time it takes to complete the task by half. This software may include

capabilities such as: - Summarizing documents longer than 2000 words and answering

questions about those documents, - Retrieving up-to-date facts from the Internet and using

those facts in combination with the LLM capabilities, - Searching over an organization’s

existing knowledge, data, or documents and retreiving information, - Retrieving highly

specialized domain knowledge, - Make recommendations given data or written input, -

Analyze written information to inform decisions, - Prepare training materials based on highly

specialized knowledge, - Provide counsel on issues, and - Maintain complex databases. ##

E3 - Exposure given image capabilities: Suppose you had access to both the LLM and a

system that could view, caption, and create images as well as any systems powered by the

LLM (those in E2 above). This system cannot take video as an input and it cannot produce

video as an output. This system cannot accurately retrieve very detailed information from

image inputs, such as measurements of dimensions within an image. Label tasks as E3 if

there is a signiﬁcant reduction in the time it takes to complete the task given access to a LLM

and these image capabilities: - Reading text from PDFs, - Scanning images, or - Creating

or editing digital images according to instructions. The images can be realistic but they

should not be detailed. The model can identify objects in the image but not relationships

between those options

## E0 - No exposure: Label tasks E0 if none of the above clearly decrease the time it takes

for an experienced worker to complete the task with high quality by at least half. Some

examples: - If a task requires a high degree of human interaction (for example, in-person

demonstrations) then it should be classiﬁed as E0. - If a task requires precise measurements

then it should be classiﬁed as E0. - If a task requires reviewing visuals in detail then it

should be classiﬁed as E0. - If a task requires any use of a hand or walking then it should

be classiﬁed as E0. - Tools built on top of the LLM cannot make any decisions that might

impact human livelihood (e.g.hiring, grading, etc.). If any part of the task involves collecting

inputs to make a ﬁnal decision (as opposed to analyzing data to inform a decision or make a

recommendation) then it should be classiﬁed as E0. The LLM can make recommendations.

- Even if tools built on top of the LLM can do a task, if using those tools would not save

an experienced worker signiﬁcant time completing the task, then it should be classiﬁed as

E0. - The LLM and systems built on top of it cannot do anything that legally requires a

human to perform the task. - If there is existing technology not powered by an LLM that

is commonly used and can complete the task then you should mark the task E0 if using an

LLM or LLM-powered tool will not further reduce the time to complete the task. When in

doubt, you should default to E0.”

user prompts = [””, ””, ””, ””] user prompts[0] = ”Read the following occupation title

and one of the tasks that belong to that occupation. Then do three things. 1: Reason step

by step to decide which of the labels (E0/E1/E2/E3) from the exposure rubric you were

given applies to the task’s exposure to LLM. Report. Give an explanation for which label

applies and report the label that you think ﬁts best. Do not say zero or N/A.

2: Report only the label that you determined for the task, which should match the label in

step 1. Do not reply N/A.

3: Given the amount of speculation required in step 1, describe your certainty about the

estimate–either high, moderate, or low.

The task to label follows:

Occupation: Inspectors, Testers, Sorters, Samplers, and Weighers. Task: Adjust, clean, or

repair products or processing equipment to correct defects found during inspections.”

assistant prompts = [””, ””, ””, ””] assistant prompts[0] = ”1: Label (E0/E1/E2/E3):

E0. Explanation: The model does not have access to any kind of physicality, and more than

half of the task (adjusting, cleaning and repairing equipment) described requires hands or

other embodiment.

2: E0.

3: High conﬁdence.”

user prompts[1] = ”Read the following occupation title and one of the tasks that belong

to that occupation. Then do three things. 1: Reason step by step to decide which of the

labels (E0/E1/E2/E3) from the exposure rubric you were given applies to the task’s expo-

sure to LLM. Report. Give an explanation for which label applies and report the label that

you think ﬁts best. Do not say zero or N/A.

2: Report only the label that you determined for the task, which should match the label in

step 1. Do not reply N/A.

3: Given the amount of speculation required in step 1, describe your certainty about the

estimate–either high, moderate, or low.

The task to label follows:

Occupation: Computer and Information Research Scientists. Task: Apply theoretical ex-

pertise and innovation to create or apply new technology, such as adapting principles for

applying computers to new uses.”

assistant prompts[1] = ”1: Label (E0/E1/E2/E3): E1. Explanation: The model can

learn theoretical expertise during training as part of its general knowledge base, and the

principles to adapt can be captured in the text input to the model.

2: E1.

3: Medium conﬁdence.”

Appendix B. Appendix Figures

Figure B1: Generative AI exposure quintile portfolio returns: market-factor adjusted. Each

bar shows the average daily return alpha during the ChatGPT release period and also for all days, for

portfolios based on Generative AI exposure sorts across all stocks in the sample, and also for the net zero

investment high-minus-low exposure portfolio that represents the “Artiﬁcial Minus Human” (AMH ) factor.

The data set consists of daily stock returns from Yahoo Finance for Nov. 15, 2022 - March 31, 2023, and

factors from Ken French’s website. The ﬁgure shows alphas estimated from portfolio-level regressions of the

form

−r

= α

release

[ChatGPT release period]

+α

not release

[Not ChatGPT release period]

+β

(Mkt

−r

)+ε

where the bars show either the intercept for the “ChatGPT release period”, or the intercept from a regression

where all days have the same intercept. Error bars indicate 95% conﬁdence intervals computed using Newey

West standard errors with ﬁve lags.

Figure B2: Industry-neutral Generative AI exposure quintile p ortfolio returns: Fama French

5-factor adjusted. Each bar shows the average daily return alpha during the ChatGPT release period and

also for all days, for portfolios based on industry-neutral Generative AI exposure sorts, and also for the net

zero investment high-minus-low exposure portfolio that represents the “Artiﬁcial Minus Human” (AMH )

factor. Industry-neutral portfolios are computed by ﬁrst forming within-industry value-weighted quintile

portfolios, and then averaging portfolio returns for the same quintiles across industries. Returns within-

industry quintiles are value-weighted, while across-industry averages are industry market-cap. weighted.

The data set consists of daily stock returns from Yahoo Finance for Nov. 15, 2022 - March 31, 2023, and

Fama French factors from Ken French’s website. The ﬁgure shows alphas estimated from portfolio-level

regressions of the form

−r

= α

release

[ChatGPT release period]

+α

not release

[Not ChatGPT release period]

fac∈F F 5

fac

+ε

where the bars show either the intercept for the “ChatGPT release period”, or the intercept from a regression

where all days have the same intercept. Error bars indicate 95% conﬁdence intervals computed using Newey

West standard errors with ﬁve lags.

Figure B3: Generative AI exposure quintile portfolio returns over time: Fama French 5-

factor adjusted. The top graph shows Generative AI exposure quintile sorts across all stocks in the

sample. The bottom graph shows the cumulative excess realized returns on portfolios based on industry-

neutral Generative AI exposure sorts. Industry-neutral portfolios are computed by ﬁrst forming within-

industry value-weighted quintile portfolios, and then averaging portfolio returns for the same quintiles across

industries. Returns within-industry quintiles are value-weighted, while across-industry averages are industry

market-cap. weighted. All portfolio returns shown are net of the risk free rate. The data set consists of

daily stock returns from Yahoo Finance for Nov. 15, 2022 - March 31, 2023, and Fama French factors from

Ken French’s website. The Fama French 5-factor model robust regressions include the market factor, HML,

SMB, RMW, and CMA factor returns in the regression.

Figure B4: Generative AI exposure quintile portfolio returns over time. The top graph shows

Generative AI exposure quintile sorts across all stocks in the sample. The bottom graph shows the cumulative

excess realized returns on portfolios based on industry-neutral Generative AI exposure sorts. Industry-

neutral portfolios are computed by ﬁrst forming within-industry value-weighted quintile portfolios, and then

averaging portfolio returns for the same quintiles across industries. Returns within-industry quintiles are

value-weighted, while across-industry averages are industry market-cap. weighted. All portfolio returns

shown are net of the risk free rate. The data set consists of daily stock returns from Yahoo Finance for Nov.

15, 2022 - March 31, 2023.

Figure B5: Within-Sector H-L Generative AI exposure Portfolio Realized Returns: Not ChatGPT Release

Period. The graph shows the average daily return alpha for days that are not in the ChatGPT release period for H-L Generative

AI exposure portfolios within each industry, at the 2-digit NAICS sector level Each H-L portfolio is formed by taking the value-

weighted highest and lowest quintiles of Generative AI exposure within each industry (based on NYSE stock cutoﬀs) and

forming zero net investment H-L portfolio returns as the equal-weighted diﬀerence in the daily realized returns between these

portfolios, and then subtracting the daily risk-free return. The industries shown omit any sectors with fewer than 10 ﬁrms

combined in the highest and lowest quantiles in the sample, as well as sectors with less than a 10 ppt Generative AI exposure

spread at the sector level between the high and low quintile. The data set consists of daily stock returns from Yahoo Finance

for Nov. 15, 2022 - March 31, 2023. The graphs show market-factor adjusted alphas estimated from regressions of the form

− r

= α

release

[ChatGPT release period]

+ α

not release

[Not ChatGPT release period]

+ β

(Mkt

− r

) + ε

where the intercept is allowed to vary with whether the day is in the release period deﬁned by Nov. 30, 2022, and the following

two weeks, or is one of the other trading days in the sample. The returns are shown in units of average daily excess realized

returns (controlling for the market factor). Red error bars indicate 95% conﬁdence intervals computed using Newey-West

standard errors with ﬁve lags.

Figure B6: Within-Subsector H-L Generative AI exposure Portfolio Realized Returns: Not ChatGPT

Release Period. Each graph shows the average daily return alpha for days that are not in the ChatGPT release period for

H-L Generative AI exposure portfolios within each industry, at the 3-digit subsector level. Each H-L portfolio is formed by

taking the value-weighted highest and lowest terciles of Generative AI exposure within each industry (based on NYSE stock

cutoﬀs) and forming zero net investment H-L portfolio returns as the equal-weighted diﬀerence in the daily realized returns

between these portfolios, and then subtracting the daily risk-free return. The industries shown omit any subsectors with fewer

than 10 ﬁrms combined in the highest and lowest quantiles in the sample, as well as subsectors with less than a 5 ppt Generative

AI exposure spread between the high and low tercile. The data set consists of daily stock returns from Yahoo Finance for Nov.

15, 2022 - March 31, 2023. The graphs show market-factor adjusted alphas estimated from regressions of the form

− r

= α

release

[ChatGPT release period]

+ α

not release

[Not ChatGPT release period]

+ β

(Mkt

− r

) + ε

where the intercept is allowed to vary with whether the day is in the release period deﬁned by Nov. 30, 2022, and the following

two weeks, or is one of the other trading days in the sample. The returns are shown in units of average daily excess realized

returns (controlling for the market factor). Red error bars indicate 95% conﬁdence intervals computed using Newey-West

standard errors with ﬁve lags.

Appendix C. Appendix Tables

Occupation Task

Task GPT

Score

GPT Explanation

Secretaries and

Administrative

Assistants, Except Legal,

Medical, and Executive

20283 Operate oﬃce equipment, such as fax

machines, copiers, or phone systems

and arrange for repairs when

equipment malfunctions.

0 The task requires physical interaction with oﬃce

equipment, which cannot be done by the LLM.

Nuclear Technicians 22315 Monitor instruments, gauges, or

recording devices under direction of

nuclear experimenters.

0 The task requires monitoring physical instruments

and gauges, which cannot be done by the LLM.

Veterinary Technologists

and Technicians

4180 Administer emergency ﬁrst aid, such

as performing emergency resuscitation

or other life saving procedures.

0 The task requires hands-on physical intervention

and cannot be completed through text

input/output alone.

Penetration Testers 21754 Document penetration test ﬁndings. 1 The model can be trained on penetration testing

and can generate text output based on the input.

Gambling Dealers 4453 Open and close cash ﬂoats and game

tables.

0 The task requires physical interaction with cash

and game tables, and the LLM cannot perform any

physical actions.

Agricultural Engineers 5337 Meet with clients, such as district or

regional councils, farmers, and

developers, to discuss their needs.

0 The task requires human interaction and

communication skills, which cannot be replaced by

the LLM.

Farm Equipment

Mechanics and Service

Technicians

13750 Maintain, repair, and overhaul farm

machinery and vehicles, such as

tractors, harvesters, and irrigation

systems.

0 The task requires physical maintenance and repair

of machinery and vehicles, which cannot be done by

the LLM.

Adult Basic Education,

Adult Secondary

Education, and English

as a Second Language

Instructors

6846 Meet with other professionals to

discuss individual students’ needs and

progress.

1 The model can assist in writing and responding to

emails, including those that involve discussing

student progress with other professionals.

Payroll and Timekeeping

Clerks

2526 Distribute and collect timecards each

pay period.

0 The task does not involve writing or transforming

text, nor does it require any complex

decision-making or analysis that the LLM could

assist with.

Environmental

Engineering Technologists

and Technicians

3647 Obtain product information, identify

vendors or suppliers, or order materials

or equipment to maintain inventory.

2 The model can help identify vendors or suppliers by

searching the internet and retrieving information.

It can also help order materials or equipment by

generating text that can be sent to suppliers.

Table C1: Examples of GPT scores assigned to task statements and GPT-provided explanations.

Table C2: Exposure score variation across GPT scoring runs

Agreement %

Score comparison Current Exposure Expected exposure Broad exposure

GPT #1 vs. GPT #2 95 90 90

GPT #1 vs. GPT #3 93 88 88

GPT #2 vs. GPT #3 96 88 88

SOC Code Occupation Title Exposure

Score

41-9041 Telemarketers .96

43-9081 Proofreaders and copy markers .95

43-3031 Bookkeeping, accounting, and auditing clerks .87

15-2021 Mathematicians .86

15-1251 Computer programmers .85

43-9022 Word processors and typists .85

43-3011 Bill and account collectors .83

27-3091 Interpreters and translators .82

43-9111 Statistical assistants .82

15-1254 Web developers .81

43-6011 Executive secretaries and executive administrative assistants .77

43-3051 Payroll and timekeeping clerks .77

43-6014 Secretaries and administrative assistants, except legal, medical, and executive .77

43-5061 Production, planning, and expediting clerks .76

15-1212 Information security analysts .75

43-6013 Medical secretaries and administrative assistants .75

27-3043 Writers and authors .75

43-4021 Correspondence clerks .74

43-9061 Oﬃce clerks, general .74

41-3091 Sales representatives of services, except advertising, insurance, ﬁnancial

services, and travel

.73

39-5093 Shampooers 0

51-6041 Shoe and leather workers and repairers 0

51-6042 Shoe machine operators and tenders 0

51-3023 Slaughterers and meat packers 0

47-2022 Stonemasons 0

47-2221 Structural iron and steel workers 0

51-2041 Structural metal fabricators and ﬁtters 0

29-9093 Surgical assistants 0

51-6052 Tailors, dressmakers, and custom sewers 0

47-2082 Tapers 0

49-9052 Telecommunications line installers and repairers 0

47-2053 Terrazzo workers and ﬁnishers 0

51-6064 Textile winding, twisting, and drawing out machine setters, operators, and

tenders

47-2044 Tile and stone setters 0

51-9197 Tire builders 0

49-3093 Tire repairers and changers 0

51-4194 Tool grinders, ﬁlers, and sharpeners 0

39-3031 Ushers, lobby attendants, and ticket takers 0

49-9064 Watch and clock repairers 0

53-7073 Wellhead pumpers 0

Table C3: Highest and lowest Generative AI exposure score occupations

NAICS

Code

Industry Title Exposure

Score

52 Finance and insurance .49

54 Professional, scientiﬁc, and technical services .49

55 Management of companies and enterprises .48

51 Information .47

42 Wholesale trade .35

91 Federal government .34

53 Real estate and rental and leasing .33

90 Government .3

22 Utilities .29

61 Educational services; state, local, and private .29

56 Administrative and support and waste management and remediation

services

.27

81 Other services (except public administration) .24

31-33 Manufacturing .24

44-45 Retail trade .22

62 Healthcare and social assistance .22

71 Arts, entertainment, and recreation .22

21 Mining, quarrying, and oil and gas extraction .21

48-49 Transportation and warehousing .2

23 Construction .17

72 Accommodation and food services .11

11 Agriculture, forestry, ﬁshing and hunting .086

Table C4: Generative AI exposure scores by industry

Table C5: Within-industry H-L Generative AI exposure returns: 2-digit sectors.

The table shows the realized excess return on GPT news days and no-GPT-news days on

the within-industry H-L Generative AI exposure quintile portfolio for the Nov. 15, 2022 -

March 31, 2023, period, in daily returns data from Yahoo Finance, controlling for average

market returns. Each H-L portfolio is formed by taking the value-weighted highest and

lowest quintiles of Generative AI exposure within each industry sector and forming zero net

investment H-L portfolio returns as the equal-weighted diﬀerence in the daily realized returns

between these portfolios, and then subtracting the daily risk-free return. The industries

shown are 2-digit NAICS sectors, omitting any sectors with fewer than 10 ﬁrms combined

in the high and low quintiles in the sample, as well as industries with less than a 10 ppt

Generative AI exposure spread between the high and low quintile. Standard errors shown

are Newey-West standard errors with a ﬁve period lag bandwidth.

Ind.

Code

Industry Title News

Day α

News

Day

t-Stat.

News

Day α

News

Day

t-Stat.

52 Finance and Insurance 0.81 2.54 0.11 0.82

33 Manufacturing: Other Products 0.81 2.22 -0.00 -0.02

56 Admin. and Support and Waste Mgmt and

Remediation Svcs

0.54 2.66 -0.02 -0.11

42 Wholesale Trade 0.24 1.47 0.05 0.57

51 Information 0.10 0.73 -0.03 -0.28

62 Health Care and Social Assistance 0.06 0.22 -0.07 -0.27

54 Professional, Scientiﬁc, and Technical Svcs -0.01 -0.05 -0.07 -0.89

44 Retail Trade: Motor Vehicle and Parts

Dealers

-0.08 -0.29 -0.29 -2.53

48 Transportation and Warehousing:

Transportation

-0.11 -0.64 0.02 0.16

72 Accommodation and Food Svcs -0.22 -1.56 0.00 0.01

53 Real Estate and Rental and Leasing -0.49 -3.05 0.07 0.40

Table C6: Within-industry H-L Generative AI exposure returns: 3-digit sub-

sectors. The table shows the realized excess return on GPT news days and no-GPT-news

days on the within-industry H-L Generative AI exposure tercile portfolio for the Nov. 15,

2022 - March 31 2023, period, in daily returns data from Yahoo Finance, controlling for

average market returns. Each H-L portfolio is formed by taking the value-weighted highest

and lowest terciles of Generative AI exposure within each industry subsector and forming

zero net investment H-L portfolio returns as the equal-weighted diﬀerence in the daily real-

ized returns between these portfolios, and then subtracting the daily risk-free return. The

industries shown are 3-digit NAICS subsectors, omitting any sectors with fewer than 10 ﬁrms

combined in the high and low quintiles in the sample, as well as industries with less than

a 5 ppt Generative AI exposure spread between the high and low tercile. Standard errors

shown are Newey-West standard errors with a ﬁve period lag bandwidth.

Ind.

Code

Industry Title News

Day α

News

Day

t-Stat.

News

Day α

News

Day

t-Stat.

511 Publishing Industries (exc. Internet) 1.19 2.14 -0.23 -1.10

522 Credit Intermed. and Rel. Activ. 1.09 3.52 0.22 1.17

336 Transp. Equipment Mfg. 0.68 1.21 -0.32 -1.04

561 Administrative and Support Svcs 0.55 2.53 0.07 0.65

441 Motor Vehicle and Parts Dealers 0.42 1.53 -0.10 -0.46

621 Ambulatory Health Care Svcs 0.36 1.75 0.07 0.26

523 Secur., Commod. Contracts, Other Fin. Inv. 0.25 0.59 -0.01 -0.14

334 Computer and Electronic Prod. Mfg. 0.22 1.52 0.18 1.37

423 Merchant Wholesalers, Durable Goods 0.19 0.74 -0.01 -0.15

213 Supp. Activ. for Mining/Oil/Gas Extraction 0.17 0.53 0.09 0.72

424 Merchant Wholesalers, Nondurable Goods 0.13 0.58 0.09 0.79

722 Food Svcs and Drinking Places 0.07 0.77 -0.01 -0.30

518 Data Processing, Hosting, and Related Svcs 0.03 0.20 -0.08 -0.37

236 Construction of Buildings 0.01 0.15 0.08 0.74

237 Heavy and Civil Engineering Construction -0.02 -0.05 0.03 0.18

212 Mining and Quarrying (except Oil and Gas) -0.04 -0.41 0.13 1.36

533 Lessors of Nonﬁn. Int. Assets (exc.

Copyrighted)

-0.06 -0.15 0.61 3.24

541 Professional, Scientiﬁc, and Technical Svcs -0.12 -0.58 -0.11 -1.62

481 Air Transp. -0.17 -1.11 0.08 0.68

335 Electr. Equip., Appliance, and Comp. Mfg. -0.20 -0.41 0.01 0.06

325 Chemical Mfg. -0.26 -1.87 0.11 2.24

332 Fabricated Metal Product Mfg. -0.26 -0.75 0.07 0.57

515 Broadcasting (except Internet) -0.29 -0.66 -0.02 -0.16

311 Food Mfg. -0.37 -2.40 -0.10 -1.85

531 Real Estate -0.39 -2.55 0.08 0.76

519 Other Information Svcs -0.65 -3.60 -0.01 -0.04

Table C7: Realized returns of portfolios sorted on Generative AI exposure after ChatGPT

release.: adjusted for Fama French 5 factors. This table reports daily excess stock returns of value-

weighted portfolios of ﬁrms sorted on Generative AI exposure. AMH is the ”Artiﬁcial Minus Human” is

the zero net investment portfolio long high exposure (H) stocks and short low exposure (L) stocks. Quintile

thresholds that deﬁne value-weighted portfolios are solely based on the sample of stocks listed on NYSE as of

the sorting date. All quintile portfolios are formed based on value weights on October 31, 2022, and weights

are adjusted based on daily returns to mimic passive buy-and-hold exposure. Industry-neutral portfolios are

computed by ﬁrst forming within-industry value-weighted quintile portfolios, and then averaging portfolio

returns for the same quintiles across industries. Returns for within-industry quintiles are value-weighted,

while across-industry averages are industry market-cap. weighted. The data set consists of daily stock

returns from Yahoo Finance for Nov. 15, 2022 - March 31, 2023. Daily market returns, risk free rates,

and additional factors are obtained from Ken French’s website. The table shows alphas estimated from

regressions of the form

−r

= α

release

[ChatGPT release period]

+α

not release

[Not ChatGPT release period]

fac∈F F 5

fac

+ε

where the intercept is either for the full sample, or is allowed to vary with whether the day is in the ChatGPT

release period consisting of Nov 30, 2022 - Dec. 14, 2022. The Fama French 5-factor model robust regressions

include the market factor, HML, SMB, RMW, and CMA factor returns r

fac

in the regression. T-statistics

in parentheses are computed using Newey-West standard errors with ﬁve lags.

Portfolios

Sample Q1 Q2 Q3 Q4 Q5 AMH

A: Excess returns (%)

All days 0.042 0.027 0.050 0.037 0.134 0.076

(0.38) (0.25) (0.54) (0.31) (1.17) (1.33)

Not ChatGPT release period 0.066 0.030 0.038 0.037 0.110 0.028

(0.57) (0.25) (0.36) (0.28) (0.85) (0.44)

ChatGPT release period -0.137 0.000 0.137 0.034 0.316 0.437

(-0.53) (0.00) (0.43) (0.11) (0.82) (3.08)

B: Fama French 5-factor-adjusted alpha (%)

All days 0.040 0.021 -0.024 0.036 0.068 0.011

(1.11) (0.71) (-1.05) (1.27) (2.40) (0.22)

Not ChatGPT release period 0.061 0.024 -0.025 0.034 0.057 -0.020

(1.57) (0.76) (-0.94) (1.09) (1.91) (-0.36)

ChatGPT release period -0.132 -0.000 -0.021 0.053 0.155 0.272

(-1.96) (-0.00) (-0.43) (0.86) (3.37) (3.09)

C: Ind.-neutral Fama French 5-factor-adjusted alpha (%)

All days 0.044 0.041 -0.006 -0.005 0.055 -0.005

(1.24) (1.84) (-0.33) (-0.28) (2.35) (-0.11)

Not ChatGPT release period 0.067 0.043 -0.016 -0.018 0.045 -0.039

(2.00) (1.69) (-0.75) (-0.99) (1.80) (-0.84)

ChatGPT release period -0.148 0.026 0.069 0.095 0.140 0.273

(-1.63) (0.86) (2.67) (3.51) (3.44) (2.64)

Table C8: Generative AI exposure for the Largest 100 U.S. Firms This table lists

the Generative AI exposure scores for the largest 100 publicly-traded ﬁrms with headquarters in the

U.S., where size is measured as the market capitalization as of November 1, 2022. Generative AI

exposure is the ﬁrm’s labor exposure deﬁned in Section I. MktCap is the ﬁrm’s market capitalization

as of November 1, 2022, in B. Sector is deﬁned at the NAICS 2-digit level.

Company Name Gen. AI exposure MktCap Sector

International Business Machines Corp 0.488 125 Information

Intuit Inc. 0.480 111 Information

QUALCOMM Inc. 0.479 132 Manufacturing

Fiserv Inc. 0.475 66 Information

NVIDIA Corporation 0.468 337 Manufacturing

S&P Global Inc 0.452 103 Administrative and Support and Waste Management and Remediation Services

Broadcom Inc 0.449 195 Manufacturing

Verizon Communications Inc 0.444 157 Information

Microsoft Corp 0.442 1,701 Information

3M Co 0.442 69 Manufacturing

Advanced Micro Devices Inc 0.441 96 Manufacturing

ServiceNow Inc 0.434 85 Information

Adobe Inc 0.427 147 Information

PayPal Holdings Inc 0.418 96 Information

Thermo Fisher Scientiﬁc Inc 0.411 203 Manufacturing

Intuitive Surgical Inc 0.404 87 Manufacturing

Automatic Data Processing Inc 0.398 101 Information

Comcast Corp 0.396 136 Information

Vertex Pharmaceuticals Inc 0.395 81 Manufacturing

Analog Devices Inc 0.392 74 Manufacturing

AbbVie Inc 0.391 260 Manufacturing

Regeneron Pharmaceuticals Inc 0.390 81 Manufacturing

Gilead Sciences Inc 0.388 99 Manufacturing

Micron Technology Inc. 0.388 60 Manufacturing

Intel Corp 0.386 117 Manufacturing

Bristol-Myers Squibb Co 0.385 165 Manufacturing

Illinois Tool Works Inc. 0.382 66 Manufacturing

Netﬂix Inc 0.381 128 Real Estate and Rental and Leasing

Meta Platforms Inc 0.381 217 Information

Lam Research Corp 0.380 56 Manufacturing

SALESFORCE INC 0.379 160 Information

General Dynamics Corp 0.378 69 Manufacturing

Abbott Laboratories 0.376 174 Manufacturing

AT&T Inc 0.375 131 Information

Applied Materials Inc 0.374 77 Manufacturing

Booking Holdings Inc 0.373 75 Information

General Electric Co 0.373 85 Wholesale Trade

Merck & Co Inc 0.372 253 Manufacturing

T-Mobile US Inc 0.371 189 Information

Johnson & Johnson 0.371 453 Manufacturing

Honeywell International Inc 0.368 137 Manufacturing

Alphabet Inc 0.366 546 Information

Amgen Inc 0.365 146 Manufacturing

Eli Lilly and Co 0.364 335 Manufacturing

Apple Inc 0.364 2,397 Manufacturing

Philip Morris International Inc 0.364 142 Manufacturing

DEERE & COMPANY 0.364 117 Manufacturing

Texas Instruments Inc 0.363 148 Manufacturing

Caterpillar Inc 0.358 115 Manufacturing

CVS Health Corp 0.356 124 Health Care and Social Assistance

Cisco Systems Inc 0.355 187 Manufacturing

Zoetis Inc 0.355 71 Manufacturing

Pﬁzer Inc 0.352 269 Manufacturing

Southern Co (The) 0.351 71 Utilities

Danaher Corp 0.350 186 Manufacturing

Procter & Gamble Co (The) 0.342 320 Manufacturing

Raytheon Technologies Corp 0.339 140 Manufacturing

Colgate-Palmolive Co 0.337 62 Manufacturing

Becton Dickinson and Co 0.331 67 Manufacturing

Dominion Energy Inc 0.330 58 Utilities

NextEra Energy Inc 0.329 154 Utilities

Walt Disney Co (The) 0.328 193 Information

Altria Group Inc 0.327 83 Manufacturing

Air Products and Chemicals Inc. 0.326 56 Manufacturing

Waste Management Inc. 0.325 64 Administrative and Support and Waste Management and Remediation Services

Duke Energy Corp 0.322 72 Utilities

EOG Resources Inc. 0.322 80 Mining, Quarrying, and Oil and Gas Extraction

Exxon Mobil Corp 0.320 466 Manufacturing

Amazon.com Inc 0.317 987 Retail Trade

Stryker Corp 0.317 83 Manufacturing

Schlumberger Ltd 0.316 73 Mining, Quarrying, and Oil and Gas Extraction

Conocophillips 0.316 163 Mining, Quarrying, and Oil and Gas Extraction

HCA Healthcare Inc 0.312 63 Health Care and Social Assistance

Marathon Petroleum Corp 0.308 59 Manufacturing

Occidental Petroleum Corp 0.307 69 Mining, Quarrying, and Oil and Gas Extraction

Coca-Cola Co (The) 0.306 258 Manufacturing

Boston Scientiﬁc Corp 0.305 61 Manufacturing

PepsiCo Inc 0.303 249 Manufacturing

Chevron Corp 0.301 357 Manufacturing

Berkshire Hathaway Inc 0.300 324 Finance and Insurance

Lockheed Martin Corp 0.299 127 Manufacturing

Boeing Co 0.298 85 Manufacturing

Sherwin-Williams Co (The) 0.296 58 Manufacturing

Activision Blizzard Inc 0.295 57 Information

Pioneer Natural Resources Co 0.294 60 Mining, Quarrying, and Oil and Gas Extraction

Mondelez International Inc 0.292 85 Manufacturing

Northrop Grumman Corp 0.291 82 Manufacturing

Tesla Inc 0.283 719 Manufacturing

Home Depot Inc. (The) 0.261 303 Retail Trade

United Parcel Service Inc 0.256 123 Transportation and Warehousing

CSX Corp 0.256 61 Transportation and Warehousing

Union Paciﬁc Corp 0.253 121 Transportation and Warehousing

Costco Wholesale Corp 0.252 221 Retail Trade

TJX Companies Inc (The) 0.243 83 Retail Trade

Lowe’s Cos Inc 0.238 120 Retail Trade

Walmart Inc 0.235 385 Retail Trade

Target Corp 0.235 76 Retail Trade

Dollar General Corporation 0.212 57 Retail Trade

McDonald’s Corp 0.194 201 Accommodation and Food Services

Starbucks Corp 0.119 100 Accommodation and Food Services