SPB Git
spboucher.ai

spb/wp17_uqo

Public

FIRC0002 Stage E (Agent 10): §1 introduction and §2 literature rewritten — contribution narrowed and 'first' claim withdrawn (LITE-01/02/08), positioning table vs Eisfeldt/Basu/Borri, AI-washing and cybersecurity strands (LITE-03/05/06), Hassan/Davis/Kim/Cao/Hong–Lim–Stein/Bloom (EDIT-14, LITE-07), §2.1 negation claim withdrawn (AI_C-07), hypotheses H3a/b, H4, H5a/b (THEO-11, ADVE-07), U8 sentences removed (ADVE-14)

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Simon-Pierre Boucher committed (Nov 5, 2026) parent 430ae85 Browse files

2 changed files +281 −179

2 changed files
Unified Split
modified paper/sections/introduction.tex +108 −82
@@ -1,92 +1,118 @@
1 1 % ============================================================================
2 +% Owner: Stage E, Agent 10 (writing). Numbers: tables/macros.tex, sections/macros_stageE.tex.
2 3 \section{Introduction}
3 4 \label{sec:introduction}
4 5 % ============================================================================
5 6
6 −The arrival of generative artificial intelligence has been the defining corporate-narrative shift
7 −of the 2020s. In the three years following the public launch of ChatGPT in late 2022, ``AI'' became
8 −the most frequently invoked theme on earnings calls, and the language of corporate disclosure
9 −filled with references to large language models, generative tools, intelligent automation and the
10 −compute infrastructure that powers them. This paper asks whether the \emph{language} of AI in
11 −firms' mandatory disclosures carried information about their subsequent stock returns, when that
12 −information was priced, and for which firms. It is the second version of a study whose first
13 −version answered the question for the 500 current constituents of the S\&P~500 and concluded that
14 −AI language was barely priced for a decade and strongly priced in one post-launch cohort, mostly
15 −through the technology sector. The present version asks whether that conclusion survives when the
16 −laboratory is enlarged in three directions at once: to some 2,500 firms spanning the capitalization
17 −spectrum of the Russell~3000, to the full text of annual and quarterly reports rather than current
18 −reports alone, and to a second measure of AI engagement that reads what firms say they \emph{do}
19 −rather than how they say it.
7 +Since the public release of ChatGPT on 30~November 2022, references to large language models,
8 +generative tools, intelligent automation and the compute infrastructure behind them have spread
9 +through corporate disclosure. This paper asks whether the \emph{language} of artificial
10 +intelligence (AI) in firms' mandatory SEC filings is associated with their subsequent stock returns,
11 +when that association appears, and for which firms. It is the second version of a study whose
12 +first version \citep{kammoun2026v11} answered the question for the current constituents of the
13 +S\&P~500 and concluded that AI language was barely priced for a decade and strongly priced in one
14 +post-launch cohort, mostly through the technology sector. The present version enlarges the
15 +laboratory in three directions: to some 2,500 firms spanning the capitalization spectrum of the
16 +Russell~3000, to the full text of annual and quarterly reports rather than current reports alone,
17 +and to a second measure of AI engagement that reads what firms say they \emph{do} rather than how
18 +they say it. It also incorporates the results of a peer review of the Russell~3000 version, which
19 +led us to rebuild the market-capitalization variable, to clean the return series against a second
20 +price source, to replace independence-based with dependence-robust inference wherever firms share
21 +a date, and to state every finding at the strength the corrected evidence supports.
20 22
21 −Answering the question requires three ingredients. The first is a large, machine-readable corpus of
22 −corporate disclosure: the UQO EDGAR Research Database, extended in October 2026 to an issuer
23 −universe that merges the S\&P~500 snapshot with the holdings of the iShares Russell~3000 ETF, holds
24 −\mnFilingsCorpus{} Form~10-K, 10-Q and 8-K filings of \mnIssuersCorpus{} issuers filed since January
25 −2010, all of them parsed into sections and 512-word chunks. The second is a measurement technology
26 −that captures the meaning of text rather than the frequency of keywords: every chunk is embedded
27 −with the \texttt{all-MiniLM-L6-v2} sentence transformer \citep{reimers2019sbert} and scored by
28 −cosine similarity to written anchor passages that define three facets of AI engagement---internal
29 −adoption, infrastructure exposure, and product-disruption risk---and five placebo concepts. The
30 −third is a look-ahead-free mapping from disclosure to tradeable returns: the signal of filing year
31 −$Y$ forms portfolios at the end of June of $Y{+}1$, and the forward twelve-month return runs from
32 −July of $Y{+}1$ to June of $Y{+}2$. The resulting estimation sample has \mnObs{} firm-years for
33 −\mnFirms{} firms over formation years 2010--2024, of which \mnSPfy{} firm-years belong to S\&P~500
34 −members, so that every result of the earlier study can be reproduced inside the new panel and
35 −placed next to its broad-universe counterpart.
23 +Three ingredients are needed. The first is a machine-readable corpus of disclosure: the UQO EDGAR
24 +Research Database, extended in October 2026 to an issuer universe that merges the S\&P~500 snapshot
25 +with the holdings of the iShares Russell~3000 ETF, holds \mnFilingsCorpus{} Form~10-K, 10-Q and 8-K
26 +filings of \mnIssuersCorpus{} issuers filed since January 2010, parsed into sections and
27 +\mChunkWordsMedian-word chunks. The second is a measurement technology that targets the meaning of
28 +text rather than the frequency of keywords: every chunk is embedded with the
29 +\texttt{all-MiniLM-L6-v2} sentence transformer \citep{reimers2019sbert} and scored by cosine
30 +similarity to written anchor passages that define three facets of AI engagement---internal
31 +adoption, infrastructure exposure and product-disruption risk---and five placebo concepts. The
32 +third is a mapping from disclosure to tradeable returns whose timing is free of look-ahead: the
33 +signal of filing year $Y$ forms portfolios at the end of June of $Y{+}1$ and the forward
34 +twelve-month return runs from July of $Y{+}1$ to June of $Y{+}2$. The estimation sample has
35 +\mnObs{} firm-years for \mnFirms{} firms over formation years 2010--2024, of which \mnSPfy{}
36 +belong to S\&P~500 members, so that every result of the earlier study can be reproduced inside the
37 +new panel and placed next to its broad-universe counterpart. The universe itself is a
38 +current-membership snapshot: the timing is look-ahead-free, the selection is not, and every level
39 +reported below is conditional on the firm surviving to October 2026 (\cref{subsec:universe,subsec:survivorship}).
36 40
37 −Three findings emerge, and each revises the earlier conclusion. First, in the broad universe AI
38 −language carries a modest but persistent cross-sectional premium over the entire sample rather than
39 −a premium that appears only after 2022. The equal-weighted top-minus-bottom quintile spread on the
40 −AI composite is \mspreadFull{} percentage points per year ($t=\mspreadFullT$) over fifteen
41 −formation cohorts, the Fama--MacBeth slope is \mrkFMcompos{} points per standard deviation
42 −($t=\mrkFMcomposT$), and a within-sector slope of \mfmFullSector{} points ($t=\mfmFullSectorT$)
43 −survives GICS-sector dummies. Among the \mnSmallFirms{} firms outside the largest thousand the
44 −spread is \msmallFull{} points ($t=\msmallFullT$) and it is as large before the launch
45 −(\msmallPre{}) as after (\msmallPost{}). Second, the post-ChatGPT increase documented for the
46 −S\&P~500 replicates within the new panel---the spread among index members rises from
47 −\mspSpreadPre{} to \mspSpreadPost{} points, with the 2024 cohort again dominant---but it does not
48 −generalize to the market. In the Russell~3000 the post-launch spread is \mspreadPost{} points
49 −against \mspreadPre{} before ($t$ of the post mean \mspreadPostT), the 2022 cohort earned a
50 −negative spread, dropping the 2024 cohort leaves a post-period mean below the pre-period one, a
51 −placebo over ten candidate break years ranks 2022 \mplaceboRank{}th of \mplaceboN{} (the largest
52 −break candidate is \mplaceboBest), and the monthly long--short portfolio earns a post-launch
53 −six-factor alpha of \mffsixPost\% per year. The increase is a large-cap phenomenon: among the
54 −largest thousand firms the spread rises from \mlargePre{} to \mlargePost{} points
55 −($t=\mlargePostT$), and in the ChatGPT event study the largest thousand firms in the top tercile of
56 −pre-launch AI language earned \mevLarge{} points more one-year abnormal return than the bottom
57 −tercile ($t=\mevLargeT$) against \mevSmall{} ($t=\mevSmallT$) for the remaining firms. Third, the
58 −two new measurement dimensions matter. The signal built from annual reports alone, which the
59 −earlier study could not compute, is the strongest of the three form-specific signals (full-sample
60 −spread \mtenkPost{} points after the launch, Fama--MacBeth slope \mhrTenK{} per standard deviation
61 −in a horse race against the quarterly- and current-report signals), whereas the count of AI,
62 −data-center, cloud and automation \emph{projects} extracted from the same filings by the SPID
63 −database carries no premium at all (\mspidPre{} and \mspidPost{} points before and after the
64 −launch).
41 +Three findings emerge, each stated with the uncertainty the data allow. First, in the broad
42 +universe AI language carries a modest cross-sectional premium over the whole sample, but its
43 +strength depends on the estimator. On returns winsorized within formation year, the equal-weighted
44 +top-minus-bottom quintile spread on the AI composite is \mspreadFullW{} percentage points per year
45 +over fifteen cohorts (Newey--West $t=\mspreadFullWT$, plain $t=\mspreadFullWTplain$, 95\%
46 +confidence interval \mspreadFullWCI), the median spread within cohort is \mspreadFullMed{} points,
47 +and the raw spread of \mspreadFull{} points owes \mtopOnePctContrib{} points to the top one percent
48 +of returns. The Fama--MacBeth slope is more stable: \mfmComposW{} points per standard deviation
49 +(Newey--West $t=\mfmComposWT$, plain $t=\mfmComposWTplain$, interval \mfmComposWCI), and
50 +\mfmFullSectorW{} points with GICS-sector dummies ($t=\mfmFullSectorWT$). The premium is therefore
51 +robust in regression form and fragile in sort form; within sectors it survives in the regressions
52 +and not in a sector-neutral quintile sort (\msectNeuFullW{} points, $t=\msectNeuFullWT$). Second,
53 +the post-ChatGPT increase documented for the S\&P~500 replicates within the new panel---the
54 +winsorized spread among index members rises from \mspSpreadPreW{} to \mspSpreadPostW{} points,
55 +with the 2024 cohort again dominant---but it does not generalize to the market. In the
56 +Russell~3000 the post-launch spread is \mspreadPostW{} points against \mspreadPreW{} before, on
57 +three cohorts whose interval is \mspreadPostWCI; the 2022 cohort is negative; the pooled
58 +post-minus-pre difference is \mpooledDiffW{} points (firm-clustered $t=\mpooledDiffWT$, two-way
59 +$t=\mpooledDiffWTtw$); the 2022--24 block ranks no higher than a random three-cohort block
60 +(permutation $p=\mpermPW$); and the monthly long--short earns a post-launch six-factor alpha of
61 +\mffsixPost\% per year. An increase is visible only among the largest thousand firms, where the
62 +winsorized spread rises from \mlargePreW{} to \mlargePostW{} points, and only because of the 2024
63 +cohort: without it the large-cap post-launch mean is \mlargePostExTwentyFourW{} points. The formal
64 +test of the large-versus-small difference---a pooled regression with a high-AI $\times$ post
65 +$\times$ large interaction---gives \mcontrastLarge{} points with a firm-clustered $t$ of
66 +\mcontrastLargeT{}, a two-way $t$ of \mcontrastLargeTtw{} and a wild-cluster bootstrap $p$ of
67 +\mcontrastLargePwild{}, against a minimum detectable effect of \mcontrastLargeMDE{} points: the
68 +gradient is visible in the point estimates and not established by the test. The ChatGPT event
69 +study shows the same pattern and the same fragility: under the single-index market model the
70 +largest thousand firms in the top tercile of pre-launch AI language earned \mevLarge{} points more
71 +one-year abnormal return than the bottom tercile (dependence-robust $t=\mevLargeCDA$) against
72 +\mevSmall{} ($t=\mevSmallCDA$) for the remaining firms, but under beta-only, market-adjusted,
73 +size-and-momentum-matched and six-factor calendar-time benchmarks the large-cap statistic falls
74 +to \mevLargeBetaCDA, \mevLargeMadjCDA, \mevLargeSizeMomCDA{} and \mevLargeFFsixT. Third, the two
75 +new measurement dimensions matter, with one caveat each. The signal built from annual reports has
76 +the largest slope in a horse race against the quarterly- and current-report signals (\mhrTenKW{}
77 +points per standard deviation, $t=\mhrTenKWT$), but the ranking of the three forms depends on the
78 +specification: a Wald test rejects equal slopes on winsorized returns ($p=\mformWaldP$) and not on
79 +raw returns ($p=\mformWaldPraw$). And the count of AI,
80 +data-center, cloud and automation \emph{projects} that a large-language-model reader extracts from
81 +the same filings shows no detectable premium: firms with at least one AI-initiative project earn
82 +\mspidAnyFirmFull{} points more than other firms ($t=\mspidAnyFirmFullT$, interval
83 +\mspidAnyFirmFullCI), with a minimum detectable effect of \mspidAnyFirmFullMDE{} points on a
84 +measure whose precision has not yet been validated against human labels.
65 85
66 −Taken together, the evidence is consistent with a persistent and partly within-industry premium on
67 −AI exposure in the broad market, present throughout the 2010s among smaller firms, together with a
68 −re-rating of the largest, most-followed AI-exposed firms after 2022 that the S\&P~500 sample
69 −isolated and that an attention-based account of the post-launch period would not predict. The
86 +Taken together, the evidence is consistent with a modest, regression-robust association between
87 +AI language and subsequent returns in the broad market, conditional on survival, and with a
88 +post-2022 increase that is confined to the largest firms and to one cohort and whose event-study
89 +expression depends on the benchmark; it is not consistent with a market-wide regime change. The
70 90 sector channel that dominated the earlier study is still visible---the long--short loads
71 −\mtechBetaPost{} on a technology-sector factor after the launch, sector-by-year fixed effects
72 −absorb the post-launch cross-sectional slope entirely, and cybersecurity language is rewarded as
73 −much as AI language---but in the broad universe it coexists with a within-sector premium that the
74 −index sample was too small and too homogeneous to detect.
91 +\mtechBetaPost{} on a technology-sector factor after the launch, and sector-by-year fixed effects
92 +reduce the post-launch cross-sectional slope to \msyPost{} points ($t=\msyPostT$)---but over the
93 +full sample it coexists with a within-sector slope that the index sample was too small to detect.
94 +We also report what the measure is and is not: the encoder reads about \mTruncWindowWords{} of
95 +the \mChunkWordsMedian{} words of a typical chunk, a keyword count correlates only \mKwRho{} with
96 +the composite, and two electric utilities sit in the top dozen firms, so the ``AI language''
97 +construct is best read as proximity to a technology-and-infrastructure vocabulary rather than as
98 +validated AI activity (\cref{subsec:construct}).
75 99
76 −\paragraph{Contribution.} The paper makes three contributions. First, substantively, it provides
77 −the first look-ahead-free evidence on the pricing of AI disclosure language across the full
78 −capitalization spectrum of the U.S.\ market, and shows that the headline post-ChatGPT result of the
79 −S\&P~500 literature is a large-cap result. Second, methodologically, it demonstrates that
80 −embedding-based disclosure measurement at the scale of 2,500 firms and 527,000 filings is feasible,
81 −documents how the signal depends on the form of the filing it is computed from, and shows that a
82 −language measure and a project-count measure extracted from the same text can have opposite
83 −pricing implications. Third, it provides a template for disciplining regime claims built on short
84 −post-event samples---cohort-by-cohort and leave-one-out estimates, pooled firm-clustered tests,
85 −placebo break dates, sector-factor spanning, winsorization and dependence-robust event-study
86 −inference---and applies it symmetrically to the index sample and to the broad universe. The
87 −remainder of the paper is organized as follows. Section~\ref{sec:literature} situates the study and
88 −states the hypotheses. Section~\ref{sec:data} describes the corpus, the universe, the semantic
89 −measures and the timing convention. Section~\ref{sec:methodology} lays out the designs.
90 −Section~\ref{sec:results} presents the main results for both universes,
91 −Section~\ref{sec:robustness} the robustness and identification tests, Section~\ref{sec:discussion}
92 −the interpretation and limitations, and Section~\ref{sec:conclusion} concludes.
100 +\paragraph{Contribution.} Relative to the text-based AI-exposure measures of
101 +\citet{eisfeldt2026generative}, \citet{basu2026aiexposure} and \citet{borri2026aifactor}, which
102 +rely on workforce task content, earnings-call language or realized model usage, we measure
103 +exposure from the firm's own mandatory filings across all three forms for the Russell~3000, trace
104 +its association with returns over fifteen look-ahead-free cohorts on a current-membership universe
105 +rather than over an event window, and show that the post-launch premium documented in index
106 +samples is confined to the largest thousand firms; and relative to the AI-washing literature
107 +\citep{erdogan2026aiwashing,wang2026aiwashing} we show that a language-model count of disclosed AI
108 +projects carries no detectable premium while the language does. Methodologically, the paper
109 +documents how an embedding signal depends on the filing form it is computed from and on the
110 +encoder's input window, and it applies to its own regime claim the discipline that short
111 +post-event samples require: cohort-by-cohort and leave-one-out estimates, two-way clustering and
112 +wild-cluster bootstraps, permutation and sup-Wald break tests, multiple-testing adjustments across
113 +families, placebo launches and dependence-robust event-study inference under five benchmarks.
114 +\Cref{sec:literature} positions the study and states the hypotheses; \cref{sec:data} describes
115 +the corpus, the universe, the measures and the timing; \cref{sec:methodology} the designs and the
116 +inference; \cref{sec:results} the main results for both universes; \cref{sec:robustness} the
117 +robustness, construct-validity and survivorship diagnostics; \cref{sec:discussion} the
118 +interpretation and limitations; \cref{sec:conclusion} concludes.
modified paper/sections/literature.tex +173 −97
@@ -1,41 +1,57 @@
1 1 % ============================================================================
2 +% Owner: Stage E, Agent 10 (writing). Numbers: tables/macros.tex, sections/macros_stageE.tex.
2 3 \section{Related Literature and Hypotheses}
3 4 \label{sec:literature}
4 5 % ============================================================================
5 6
6 −Our study sits at the intersection of five strands of research: textual analysis in finance, the
7 −pricing of corporate disclosure, technology and innovation in asset pricing, investor attention and
8 −narrative economics, and the nascent literature on the economics of artificial intelligence. Each
9 −contributes a building block---a measurement tool, an object of study, a candidate risk factor, a
10 −behavioural mechanism, and a comparison case---and a sixth, methodological strand disciplines the
11 −inference. We take the strands in turn and close with the hypotheses they jointly motivate.
7 +The study draws on six strands of research: textual analysis in finance, the pricing of corporate
8 +disclosure, text-based measures of technology and AI exposure, investor attention, the AI-washing
9 +and cybersecurity-disclosure literatures that bear on what the measure captures, and the
10 +methodological literature on short samples, multiple testing and event-study inference. We take
11 +the strands in turn, place the paper against its three closest competitors, and close with the
12 +hypotheses.
12 13
13 14 \subsection{Textual analysis and natural language processing in finance}
15 +\label{subsec:lit_text}
14 16 A now-large literature extracts economic signals from financial text. \citet{tetlock2007} showed
15 17 that the tone of a \textit{Wall Street Journal} column forecasts aggregate returns, and
16 18 \citet{tetlock2008} extended this to firm-level news. \citet{lm2011} demonstrated that
17 19 finance-specific sentiment dictionaries materially outperform general-purpose word lists in 10-K
18 −analysis, launching a wave of dictionary-based studies; \citet{lm2016} and \citet{lm2020} survey
19 −the field and its pitfalls, including the sensitivity of results to document selection and
20 −pre-processing. A second generation moved from counting words to modelling them: \citet{gentzkow2019}
21 −review text-as-data methods in economics, \citet{kekellyxiu2019} develop a supervised topic model
22 −that predicts returns from news, \citet{kelly2019textregression} propose a sparse text-regression
23 −framework, and \citet{bybee2024narrative} recover a ``state of the economy'' from business news.
24 −A third generation uses distributed representations of meaning. \citet{hanleyhoberg2019} apply word
20 +analysis; \citet{lm2016} and \citet{lm2020} survey the field and its pitfalls, including the
21 +sensitivity of results to document selection and pre-processing. A second generation moved from
22 +counting words to modelling them: \citet{gentzkow2019} review text-as-data methods in economics,
23 +\citet{kekellyxiu2019} develop a supervised topic model that predicts returns from news,
24 +\citet{kelly2019textregression} propose a sparse text-regression framework, and
25 +\citet{bybee2024narrative} recover a ``state of the economy'' from business news. A third
26 +generation uses distributed representations of meaning. \citet{hanleyhoberg2019} apply word
25 27 embeddings to bank 10-Ks to identify emerging risks before they are named in dictionaries;
26 28 \citet{limai2021} measure corporate culture from earnings calls with word embeddings;
27 29 \citet{huang2023finbert} show that a domain-adapted transformer outperforms dictionaries and
28 30 bag-of-words classifiers on financial text; and \citet{cong2022textfactor} build interpretable
29 −``textual factors'' from embeddings. Our paper belongs to this embedding-based generation. The
30 −limitations of keyword counting are acute for a theme like AI: a count cannot distinguish ``we do
31 −not currently use artificial intelligence'' from ``we are an artificial-intelligence leader,'' it
32 −misses the paraphrase of a firm that describes machine learning without writing ``AI,'' and it
33 −treats boilerplate and substance identically. We instead measure the semantic similarity of each
34 −passage to anchor concepts with a sentence transformer \citep{reimers2019sbert}, a design close in
35 −spirit to the similarity-to-reference-text constructions of \citet{hanleyhoberg2019}. The
36 −literature also warns, however, that text-based measures inherit the composition of the corpus they
37 −are computed on \citep{lm2016,lm2020}; Section~\ref{sec:data} therefore documents that composition
38 −in detail.
31 +``textual factors'' from embeddings. The firm-level exposure measures of \citet{hassan2019firmlevel}
32 +(political risk from earnings-call text), \citet{davis2020exposures} (COVID-19 exposures),
33 +\citet{bloom2021diffusion} (technology diffusion from patents, calls and job postings) and
34 +\citet{kim2023llmrisks} (risk assessments produced by a generative model) are the direct
35 +methodological lineage of a construct-specific score computed from a firm's own text.
36 +
37 +Our paper belongs to the embedding-based generation. Its measure is the semantic similarity of
38 +each passage to anchor paragraphs, computed with a sentence transformer
39 +\citep{reimers2019sbert}---a design close in spirit to the similarity-to-reference-text
40 +constructions of \citet{hanleyhoberg2019} and \citet{celenymarechal2024cyber}. Embeddings have
41 +two advantages over a keyword count: a passage that describes machine learning without writing
42 +``AI'' is scored, and a count of the acronym does not distinguish a risk-factor boilerplate from
43 +an operating description. They also have limits that \cref{subsec:construct} measures rather
44 +than assumes. The first version of this paper claimed that the embedding separates ``we do not
45 +currently use artificial intelligence'' from ``we are an artificial-intelligence leader''; a
46 +direct probe shows that it does so only in part---negated sentences score lower than affirmative
47 +ones (\mProbeNegated{} against \mProbeAffirm{} on the adoption anchor) but still above the
48 +\mProbeNegatedPct{}th percentile of corpus chunks---so the claim is withdrawn. Sentence encoders
49 +of the MiniLM class are known to be weakly sensitive to negation, and the measure should be read
50 +as topical proximity to AI vocabulary rather than as a stance. The literature also warns that
51 +text-based measures inherit the composition of the corpus they are computed on
52 +\citep{lm2016,lm2020} and that firms adapt their language to the machines that read it
53 +\citep{cao2023machinelistening}; \cref{sec:data,subsec:construct} therefore document the corpus
54 +composition, the encoder's input window and the keyword baseline in detail.
39 55
40 56 \subsection{The pricing of corporate disclosure}
41 57 A parallel literature asks whether the \emph{content} and \emph{changes} of mandatory disclosure
@@ -43,35 +59,78 @@ predict returns. \citet{cohen2020lazy} show that firms that change the language
43 59 reports relative to the prior year experience predictable future returns and operating outcomes.
44 60 \citet{hoberg2010product} and \citet{hobergphillips2016} use 10-K product descriptions to build
45 61 text-based industry classifications that carry asset-pricing content. \citet{lopezlira2023risk}
46 −extracts priced risk exposures from risk-factor sections. The general lesson is that the
47 −qualitative portions of filings contain value-relevant information that markets process
48 −incompletely and with delay. We contribute a concept-level measure---AI engagement---computed from the annual, quarterly
49 −and current reports of a broad universe of U.S.\ listed firms, and we characterize how its relation to
50 −returns varies over time and across firm size, dimensions the disclosure-pricing literature has largely left implicit.
62 +extracts priced risk exposures from risk-factor sections. The lesson is that the qualitative
63 +portions of filings contain value-relevant information that markets process incompletely and with
64 +delay. We contribute a concept-level measure---AI engagement---computed from the annual,
65 +quarterly and current reports of a broad universe of U.S.\ listed firms, and we characterize how
66 +its association with returns varies over time and across firm size, two dimensions the
67 +disclosure-pricing literature has largely left implicit.
68 +
69 +\subsection{Text-based AI exposure and the cross-section of returns: the closest papers}
70 +\label{subsec:lit_ai}
71 +Asset pricing has long sought to connect technological change to risk and return
72 +\citep{hirshleifer2013innovative,kogan2017technological,koganpapanikolaou2014}, and a recent
73 +literature measures AI exposure directly. \citet{babina2024aiinvestment} measure firm-level AI
74 +investment from job postings and r\'esum\'es and document faster growth among AI-adopting firms;
75 +\citet{acemoglu2022aijobs}, \citet{webb2020} and \citet{hampole2025ai} measure the exposure of
76 +occupations, establishments and firms' workforces to AI; \citet{acemoglu2020wrong} study the
77 +labour-market effects of industrial robots. Three papers of 2026 are the closest to ours and were
78 +absent from the first version. \citet{eisfeldt2026generative} construct a firm-level measure of
79 +\emph{generative}-AI exposure from the task content of occupations, show that a long--short
80 +portfolio on that measure earned large returns in the two weeks after ChatGPT's release, and
81 +separate this workforce exposure from firms' \emph{product} exposure to AI. \citet{basu2026aiexposure}
82 +build text-based AI-exposure measures from the earnings-call transcripts of some 13,800 firms in
83 +44 countries over 2010--2025, find that a one-standard-deviation change in exposure is associated
84 +with higher annual returns concentrated after the ChatGPT launch, and report that a U.S.\
85 +long--short portfolio of high-exposure firms earns no abnormal return before the launch, once its
86 +large-cap growth tilts are accounted for, and a large one afterward. \citet{borri2026aifactor}
87 +measure exposure from realized consumption of language-model tokens and find a positive premium
88 +in developed markets. Three further papers bear on specific constructs: \citet{song2026productexposure}
89 +builds a product-exposure measure from 10-K business descriptions, which is the object our
90 +disruption-risk anchor targets; \citet{wulam2026genai} show that short-window reactions to AI
91 +discussions on earnings calls depend on whether the discussion conveys new or concrete
92 +information, a result that parallels our language-versus-projects contrast; and
93 +\citet{jha2024chatgpt} and \citet{lopezliratang2026} use a language model itself as the reader of
94 +corporate or news text. \citet{yilmazkuday2026chatgpt} examines the aggregate return response to
95 +ChatGPT.
96 +
97 +\Cref{tab:positioning} places the paper against the three closest papers. Our measure is built
98 +from the firm's own mandatory filings across all three forms, which none of them uses; our return
99 +profile covers fifteen look-ahead-free formation cohorts on a current-membership universe rather
100 +than an event window; and our second measure---a language-model count of disclosed AI
101 +projects---has no counterpart. Where \citet{basu2026aiexposure} find the association concentrated
102 +after the launch, we find a modest association throughout the sample in regression form and a
103 +post-launch increase confined to the largest firms; the two readings are not incompatible, because
104 +their sample is dominated by the large firms in which we also find the increase, but the
105 +difference in universe, text source and horizon should be kept in mind. Where
106 +\citet{eisfeldt2026generative} and \citet{song2026productexposure} distinguish labour from product
107 +exposure, our three anchors map imperfectly onto that distinction: the adoption anchor describes
108 +the firm's own use of AI (closest to an operating, partly labour-saving exposure), the
109 +infrastructure anchor describes compute and data-center investment (a supplier or user position
110 +in the AI build-out, which neither measure targets), and the disruption-risk anchor describes the
111 +threat of AI to the firm's products (the product exposure of Song). The sign and timing of our
112 +disruption-risk spread are compared with their results in \cref{sec:discussion}.
51 113
52 −\subsection{Technology, innovation, and the cross-section of returns}
53 −Asset pricing has long sought to connect technological change to risk and return.
54 −\citet{hirshleifer2013innovative} show that innovative efficiency predicts returns;
55 −\citet{kogan2017technological} build market-based measures of innovation from patents and link them
56 −to growth and reallocation; \citet{koganpapanikolaou2014} show that investment-specific technology
57 −shocks shape the cross-section. A more recent literature studies AI exposure directly.
58 −\citet{babina2024aiinvestment} measure firm-level AI investment from job postings and r\'esum\'es
59 −and document faster growth among AI-adopting firms; \citet{acemoglu2022aijobs} and \citet{webb2020}
60 −measure the exposure of occupations and establishments to AI; \citet{acemoglu2020wrong} study the
61 −labour-market effects of industrial robots. Closest to us, \citet{eisfeldt2023generative} construct
62 −a firm-level measure of \emph{generative}-AI exposure from the task content of occupations and show
63 −that a long--short portfolio on that measure earned large returns in the weeks after ChatGPT's
64 −release; \citet{jha2024chatgpt} use ChatGPT itself to extract expected investment and employment
65 −policies from earnings calls; and \citet{lopezliratang2026} show that a large language model can
66 −rank news headlines in a way that predicts next-day returns. Our contribution relative to this work
67 −is to measure AI exposure from the firm's own disclosure language, available for some 2,500 firms
68 −at annual frequency, and to characterize the full time profile of its pricing across fifteen
69 −cohorts rather than an event window. Where \citet{eisfeldt2023generative} document a sharp
70 −repricing in the weeks around the launch, we ask whether a persistent, text-based exposure measure
71 −carried a premium before the launch and whether the premium after it is distinguishable from a
72 −technology-sector effect.
114 +\begin{table}[htbp]\centering\footnotesize
115 +\caption{\textbf{Positioning relative to the closest papers.} Qualitative summary of each paper's
116 +question, measure, sample and headline finding, taken from the published abstracts; no number of
117 +this paper is compared here.}\label{tab:positioning}
118 +\begin{tabular}{@{}p{3.1cm}p{3.6cm}p{3.1cm}p{5.6cm}@{}}
119 +\toprule
120 +Paper & Exposure measure & Sample & Headline finding and relation to this paper \\
121 +\midrule
122 +\citet{eisfeldt2026generative} & Occupational task content of the workforce (generative-AI exposure); product exposure separated & U.S.\ listed firms, event window around 30 Nov 2022 & Long--short on workforce exposure earns large returns in the two weeks after the launch; labour and product exposures distinct. We use filing language, fifteen annual cohorts and a size split; our disruption-risk anchor is a product-exposure construct. \\
123 +\addlinespace
124 +\citet{basu2026aiexposure} & Hybrid NLP on earnings-call transcripts; topic-specific AI measures & $\approx$13,800 firms, 44 countries, 2010--2025 & Return association concentrated after ChatGPT; U.S.\ long--short has no pre-launch alpha once large-cap growth tilts are controlled. Same question and event; we use mandatory filings, a U.S.\ Russell~3000 universe and a language-versus-projects contrast, and find the post-launch increase only among the largest firms. \\
125 +\addlinespace
126 +\citet{borri2026aifactor} & Realized consumption of language-model tokens & U.S.\ and international equities & Positive premium on usage-based exposure, significant in developed markets. A third exposure construct; usage rather than disclosure. \\
127 +\addlinespace
128 +This paper & Sentence-embedding similarity of 10-K, 10-Q and 8-K text to anchor passages; LLM count of disclosed AI projects & \mnFirms{} firms (current S\&P~500 $\cup$ IWV holdings), formation cohorts 2010--2024 & Modest regression-robust, sort-fragile association over the full sample, conditional on survival; post-launch increase confined to the largest thousand firms and the 2024 cohort; no detectable premium on disclosed projects. \\
129 +\bottomrule
130 +\end{tabular}
131 +\end{table}
73 132
74 −\subsection{Investor attention, salience, and narrative economics}
133 +\subsection{Investor attention, salience, and the size gradient}
75 134 Why might value-relevant disclosure be priced only after a catalyst? The behavioural literature
76 135 offers a mechanism: limited investor attention. \citet{barberodean2008} show that individual
77 136 investors are drawn to attention-grabbing stocks; \citet{da2011} construct an attention proxy from
@@ -79,60 +138,77 @@ Google search and link it to prices; \citet{dellavigna2009} document under-react
79 138 released when attention is low; and \citet{hirshleifer2009driven} show that earnings news is
80 139 underweighted when investors are distracted. \citet{shiller2017narrative} argues that contagious
81 140 narratives drive major asset-price movements, and \citet{bakerwurgler2006} show that sentiment
82 −differentially affects hard-to-value, speculative stocks. These mechanisms predict that a salient
83 −consumer event such as ChatGPT could draw attention to information that had been sitting in filings
84 −for years. They do not, however, distinguish a firm-level attention shock from a sector-level
85 −revision of expected cash flows, and our tests are designed to separate the two.
141 +differentially affects hard-to-value, speculative stocks. \citet{honglimstein2000} show that
142 +firm-specific information diffuses more slowly among small, lightly covered firms, which is the
143 +natural prediction for where a disclosure-based signal should be priced with the longest delay.
144 +These mechanisms predict that a salient consumer event such as ChatGPT could draw attention to
145 +information that had been sitting in filings for years, and that the effect should be larger
146 +where attention frictions are larger. They do not distinguish a firm-level attention shock from a
147 +sector-level revision of expected cash flows, and our designs separate the two only partially:
148 +no attention proxy is in the panel, so the size gradient is the only attention-related test.
86 149
87 −\subsection{ESG and climate disclosure as a comparison case}
88 −Our placebo design connects to the literature on the pricing of ESG and climate disclosure.
89 −\citet{pastor2021sustainable} model how sustainability preferences depress the expected returns of
90 −green assets in equilibrium, and \citet{pastor2022dissecting} show that green stocks' strong
91 −realized returns through 2021 reflected unexpected increases in climate concern rather than high
92 −expected returns. \citet{bolton2021risk} document a carbon premium, and \citet{sautner2023} build a
93 −firm-level climate-exposure measure from earnings-call language that is priced in options and
94 −returns. Consistent with the post-2021 reversal in this literature, we find that high-ESG and
95 −high-climate disclosure earned negative spreads after 2022, which helps to rule out a generic
96 −``more disclosure earns more'' explanation of the AI result.
150 +\subsection{AI washing, cybersecurity disclosure, and what a language measure captures}
151 +\label{subsec:lit_washing}
152 +Two literatures bear on the interpretation of a premium on AI \emph{language} that is absent on
153 +AI \emph{projects}. The first names the phenomenon ``AI washing'': \citet{erdogan2026aiwashing}
154 +study the disclosure response of U.S.\ firms to the SEC's warnings about unsupported AI claims,
155 +and \citet{wang2026aiwashing} find that AI washing earns a short-lived post-disclosure
156 +abnormal-return premium that later reverses. Our reading of the language premium as a premium on
157 +``sounding like an AI firm'' is the AI-washing hypothesis, and it carries a testable
158 +implication---reversal at longer horizons---that \cref{subsec:spidres} examines. The second
159 +documents a cybersecurity premium: \citet{celenymarechal2024cyber} measure the proximity of 10-K
160 +text to a cyber corpus with a construction like ours and find that high-cyber-risk stocks earn
161 +higher returns, and \citet{kearney2025cyberchatgpt} study cyber risk around the ChatGPT launch.
162 +The cybersecurity spread that our placebo design produces is therefore a documented premium, not
163 +an unexplained artefact, and \cref{subsec:placebo,subsec:spidres} report the correlation of the
164 +two measures and a cyber-neutral AI sort. The ESG and climate placebos connect to
165 +\citet{pastor2021sustainable}, \citet{pastor2022dissecting}, \citet{bolton2021risk} and
166 +\citet{sautner2023}; consistent with the post-2021 reversal in that literature, high-ESG and
167 +high-climate language earned negative spreads after 2022.
97 168
98 −\subsection{Short samples, data snooping, and event-study inference}
99 −Two methodological literatures bear directly on how much weight our post-2022 estimates can carry.
100 −The first concerns the credibility of cross-sectional return patterns. \citet{harveyliuzhu2016}
101 −argue that the multiplicity of tested predictors requires $t$-statistics well above conventional
102 −thresholds; \citet{houxuezhang2020} find that most published anomalies fail to replicate under
103 −standard procedures; and \citet{mcleanpontiff2016} document post-publication decay. A spread
104 −estimated on three annual observations, one of which is extreme, must be read in that light. The
105 −second concerns inference in event studies when all firms share a single event date, so that
106 −abnormal returns are cross-sectionally correlated \citep{brownwarner1985,kolari2010} and
107 −event-induced variance may differ from estimation-window variance \citep{boehmer1991}, and in
108 −long-horizon settings, where the choice of benchmark and the compounding of abnormal returns
109 −matter \citep{barberlyon1997,fama1998,mitchellstafford2000,kothariwarner2007}. Our event-study
110 −inference follows the portfolio approach recommended in that literature. Finally, the use of a
111 −current-membership universe raises the survivorship concerns analysed by
112 −\citet{browngoetzmannross1995}, which we address with restrictions on listing history.
169 +\subsection{Short samples, multiple testing, survivorship, and event-study inference}
170 +Four methodological literatures govern how much weight the estimates can carry.
171 +\citet{harveyliuzhu2016} argue that the multiplicity of tested predictors requires $t$-statistics
172 +well above conventional thresholds; \citet{houxuezhang2020} find that most published anomalies
173 +fail to replicate under standard procedures; and \citet{mcleanpontiff2016} document
174 +post-publication decay. A spread estimated on three annual observations, one of which is extreme,
175 +must be read in that light, and \cref{subsec:mht} reports family sizes, adjusted $p$-values, a
176 +deflated Sharpe ratio and a specification curve. The use of a current-membership universe raises
177 +the survivorship and look-back selection analysed by \citet{browngoetzmannross1995}; the first
178 +version addressed it with restrictions on listing history, which condition on entry, not on exit,
179 +and \cref{subsec:survivorship} replaces them with point-in-time holdings, attrition rates and a
180 +bounding exercise. Event studies in which all firms share a single date have cross-sectionally
181 +correlated abnormal returns \citep{brownwarner1985,kolari2010}, event-induced variance
182 +\citep{boehmer1991} and, at long horizons, a dependence on the benchmark and the compounding
183 +convention \citep{barberlyon1997,fama1998,mitchellstafford2000,kothariwarner2007}; our inference
184 +follows the portfolio approach of that literature and reports every statistic under several
185 +benchmarks. Finally, inference in panels with few time clusters follows
186 +\citet{petersen2009}, \citet{thompson2011} and \citet{cameron2011}.
113 187
114 188 \subsection{Hypotheses}
115 189 \label{subsec:hypotheses}
116 −Our analysis is organized around four hypotheses.
190 +The analysis is organized around five hypotheses, each stated as a prediction about an
191 +association and each with the clause that would falsify it.
117 192 \begin{description}[leftmargin=2.4em,style=nextline]
118 −\item[H1 (Unconditional pricing).] If AI exposure carries a stable risk premium or persistent
119 −mispricing, then sorting on AI-disclosure intensity yields a positive and significant risk-adjusted
120 −return spread over the full 2010--2024 formation sample.
193 +\item[H1 (Unconditional association).] If AI exposure carries a stable premium or persistent
194 +mispricing, sorting on AI-disclosure intensity yields a positive risk-adjusted return spread over
195 +the full 2010--2024 formation sample, in sort and in regression form, that survives winsorization.
121 196 \item[H2 (Regime dependence).] If instead AI exposure is repriced only once AI becomes salient, the
122 −return spread is small before and large after the late-2022 generative-AI catalyst, and the
123 −increase is visible in each post-launch cohort rather than in a single one.
197 +return spread is small before and large after the late-2022 generative-AI catalyst, the increase
198 +is visible in each post-launch cohort rather than in a single one, and the 2022 break date stands
199 +out among placebo break dates.
124 200 \item[H3 (AI-specificity).] If the post-2022 spread reflects an AI repricing rather than a
125 −mechanical or market-wide artefact, it appears for AI-related semantic factors but not for unrelated
126 −factors such as ESG, climate, or macro risk, and it is not spanned by a technology-sector factor.
201 +mechanical or market-wide artefact, (a) it appears for AI-related semantic factors but not for
202 +unrelated factors such as ESG, climate or macro risk, and (b) it is not spanned by a
203 +technology-sector factor.
127 204 \item[H4 (Cross-sectional, not within-firm).] If the market repriced \emph{which} firms are
128 −AI-exposed rather than rewarding incremental wording, the effect is concentrated in the
129 −cross-section, attenuates under firm fixed effects, and displays no pre-trend when exposure is held
130 −fixed at its pre-launch level.
131 −\end{description}
132 −\noindent We add a fifth hypothesis that the broad universe makes testable.
133 −\begin{description}[leftmargin=2.4em,style=nextline]
134 −\item[H5 (Size and attention).] If the post-launch repricing reflects limited attention rather than a
135 −revision of sector cash flows, it is stronger among smaller, less-followed firms than among the largest
136 −thousand, and it is visible in the firm's own disclosed AI projects as well as in its language.
205 +AI-exposed rather than rewarding incremental wording, the association is concentrated in the
206 +cross-section, attenuates under firm fixed effects, and the event-time profile of a pre-determined
207 +exposure shows no differential trend before the launch and a positive shift after it.
208 +\item[H5 (Size and projects).] (a) If the post-launch repricing reflects limited attention rather
209 +than a revision of sector cash flows, it is stronger among smaller, less-followed firms than among
210 +the largest thousand. (b) If the language premium reflects concrete AI activity, it is visible in
211 +the firm's own disclosed AI projects as well as in its language.
137 212 \end{description}
138 −\noindent Section~\ref{sec:discussion} summarizes what the Russell~3000 evidence says about each hypothesis.
213 +\noindent \Cref{subsec:verdicts} records, clause by clause, what the Russell~3000 evidence says
214 +about each hypothesis.
139 215
Menu

Theme
↑↓ navigate ↵ open esc close / or ⌘K anywhere