\documentclass[11pt]{article}
\usepackage[margin=1in]{geometry}
\usepackage{graphicx,booktabs,amsmath,amssymb,siunitx,caption,float,multirow}
\usepackage[round,authoryear]{natbib}
\usepackage[hidelinks]{hyperref}
\usepackage{xcolor,sectsty,setspace,enumitem}
\usepackage{titling}
\definecolor{navy}{HTML}{1b3a6b}
\sectionfont{\color{navy}\large}\subsectionfont{\color{navy}\normalsize}
\captionsetup{font=small,labelfont={bf,color=navy}}
\graphicspath{{../figures/}}
\setlength{\parskip}{0.4em}
\sisetup{group-minimum-digits=4,group-separator={,}}
\newcommand{\tstat}[1]{{\footnotesize(\textit{t}=#1)}}
\doublespacing
\hypersetup{linkcolor=navy,citecolor=navy}

\title{\color{navy}\bfseries Does the Language of Artificial Intelligence Move Markets?\\[3mm]
\large A Semantic Analysis of S\&P~500 SEC Filings and the Cross-Section of Stock Returns, 2010--2026}
\author{%
\begin{tabular}{c}
\textbf{Manel Kammoun}\\
\texttt{manel.kammoun@uqo.ca}\\[0.5em]
\textbf{Djerrycharli Tandjambianda}\\
\texttt{djerrycharli.tandjambianda@uqo.ca}\\[0.5em]
\textbf{Simon-Pierre Boucher}\\
\texttt{simon-pierre.boucher@uqo.ca}\\[0.8em]
{\large\textbf{Universit\'e du Qu\'ebec en Outaouais (UQO)}}\\
\end{tabular}%
}
\date{June 2026}

\begin{document}
\maketitle
\thispagestyle{empty}

\begin{abstract}
\noindent\normalsize
We construct a \emph{semantic} measure of artificial-intelligence (AI) engagement for every
S\&P~500 firm by embedding the text of its annual SEC filings (Form~10-K) with a sentence
transformer and scoring each passage by cosine similarity to professionally curated AI
``anchor'' concepts---internal AI adoption, AI infrastructure exposure, and AI product-disruption
risk. Linking these scores to monthly equity prices from Yahoo Finance and to the Fama--French
factor data, we assemble a panel of \num{7869} firm-years (500 firms, fiscal years 2010--2026)
and ask whether the intensity of AI language predicts the cross-section of subsequent returns.
Our central result is \textbf{conditional and economically large}. Over the full sample the
relation between AI-disclosure intensity and future returns is positive but modest and only
marginally significant; a risk-adjusted long--short portfolio earns an insignificant
$\sim$2\%/year. After the public launch of ChatGPT (30 November 2022), however, the relation
\emph{more than sextuples}: the spread between the top and bottom quintile of AI disclosure rises
from $+2.4$ percentage points per year before 2022 to $+14.6$ points per year afterward. The
break survives section-weighting choices, alternative signal definitions and horizons, dimension
reduction, and---crucially---the exclusion of the ``Magnificent~7'' mega-caps. A battery of
identification tests sharpens the interpretation: (i) a \emph{placebo} across five non-AI
semantic factors shows the post-2022 jump is specific to AI-related language (ESG and climate
disclosure, by contrast, earn \emph{negative} post-2022 spreads); (ii) a market-model
\emph{event study} finds that firms with higher pre-event AI disclosure earned $+10.1\%$ higher
cumulative abnormal returns over the year after ChatGPT \tstat{1.97}; and (iii) the effect is
absorbed by firm fixed effects and roughly halved by sector fixed effects, indicating a
\emph{cross-sectional, regime-dependent re-rating}---the market repriced \emph{which firms and
sectors} are AI-exposed once AI became salient, rather than rewarding year-to-year changes in any
single firm's wording. The findings provide direct, text-based evidence for an attention- and
narrative-driven repricing of AI exposure, and illustrate the value of embedding-based disclosure
measurement for asset pricing.

\vspace{2mm}
\noindent\textbf{Keywords:} textual analysis, sentence embeddings, artificial intelligence,
disclosure, cross-section of returns, investor attention, event study, generative AI.\\
\noindent\textbf{JEL:} G12, G14, G17, M41, O33.
\end{abstract}

\clearpage
\setcounter{page}{1}
\tableofcontents
\clearpage

\section{Introduction}
The arrival of generative artificial intelligence has been the defining corporate-narrative shift
of the 2020s. Between the public launch of ChatGPT in late 2022 and 2025, the market
capitalization of AI-linked firms swelled by trillions of dollars, ``AI'' became the most
frequently invoked theme on earnings calls, and the language of corporate disclosure filled with
references to large language models, generative tools, intelligent automation, and the compute
infrastructure that powers them. This paper asks a deceptively simple question: did the
\emph{language} of AI in firms' mandatory disclosures carry information about their subsequent
stock returns---and if so, when and through what channel?

Answering this question well requires three ingredients that have only recently become jointly
feasible: a large, machine-readable corpus of corporate disclosure; a measurement technology that
captures the \emph{meaning} of text rather than the mere frequency of keywords; and a clean
mapping from disclosure to tradeable returns. We bring all three to bear. Our laboratory is the
universe of S\&P~500 constituents over fiscal years 2010--2026. Our raw material is the full text
of their annual reports (Form~10-K), parsed section by section. Our measurement engine is the
EDGAR Research Database, which embeds every text chunk with the \texttt{all-MiniLM-L6-v2} sentence
transformer and scores it by cosine similarity to a set of carefully written ``anchor'' passages
that define eight economic concepts. Three of these anchors isolate distinct facets of AI
engagement: \emph{internal AI adoption} (using AI to run the business), \emph{AI infrastructure
exposure} (capital tied to AI compute), and \emph{AI product-disruption risk} (the danger that AI
commoditizes the firm's offerings). The result is a continuous, comparable, firm-year measure of
how intensively---and in what sense---each firm engages with AI in its own words.

We pair these semantic scores with monthly adjusted prices for all 500 tickers from Yahoo Finance
and with the Fama--French five factors plus momentum. To avoid look-ahead bias we adopt the
standard portfolio-formation convention: because a 10-K covering fiscal year $Y$ is public by
early in year $Y{+}1$, we form signals at the end of June of $Y{+}1$ and measure the forward
twelve-month return from July of $Y{+}1$ to June of $Y{+}2$. The signal is therefore always known
before the holding period begins. This yields \num{6852} firm-years with a strictly tradeable
forward return, spanning formation years 2010 through 2024.

Our headline finding is that the pricing of AI language is \emph{not} a stable factor premium but
a \textbf{regime-dependent re-rating that switched on with the generative-AI boom}. For most of
the 2010s, sorting S\&P~500 firms on the intensity of their AI disclosure produced only a faint
return pattern: the equal-weighted spread between the highest and lowest quintiles averaged about
$+2.4$ percentage points per year, and a risk-adjusted long--short portfolio earned an
economically small and statistically insignificant alpha. After the launch of ChatGPT in November
2022, the same sort produced a spread of $+14.6$ percentage points per year. The cross-sectional
re-rating is visible in every post-2022 formation cohort, in a cumulative long--short wealth
curve that inflects sharply upward at the end of 2022, and in a year-by-year Fama--MacBeth slope
that jumps from near zero to large and positive.

A natural worry is that this pattern simply reflects the well-known rally of a handful of
technology mega-caps. We address this concern head-on and reject it. The post-2022 spread is
$+14.1$ percentage points (\textit{t}=1.9) when the ``Magnificent~7'' are excluded and $+14.5$
points when the ten largest firms by market capitalization are dropped---the effect is
broad-based across the index, not an artifact of Nvidia, Microsoft, or their peers. It is robust
to using the simple mean, the section-weighted mean, or the top-5\% tail of similarity scores; to
six-, twelve-, and twenty-four-month horizons; and to replacing the three AI signals with their
first principal component (which alone explains 87\% of their common variation).

We then deploy three complementary tests to pin down what the result is---and is not. First, a
\emph{placebo} exercise re-runs the entire design on the five non-AI semantic factors in the
database (ESG, climate, cybersecurity, automation/labor, and macro risk). If the post-2022 jump
merely reflected a generic ``more disclosure earns more'' phenomenon, or a mechanical artifact of
our pipeline, every factor should display it. They do not. The post-2022 spread is large and
positive for the AI factors, essentially zero for macro-risk language, and \emph{negative} for
ESG and climate disclosure---high-ESG firms \emph{underperformed} after 2022. The post-ChatGPT
re-rating is specific to AI-related content. Second, we run a market-model \emph{event study}
around the 30 November 2022 ChatGPT launch. Firms in the top tercile of pre-event AI disclosure
(measured from their FY2021 10-K, the last report filed before the event) earned cumulative
abnormal returns $+10.1$ percentage points higher than the bottom tercile over the following year
\tstat{1.97} and $+14.6$ points higher over eighteen months. Because abnormal returns are
purged of market-beta exposure, this is a clean cross-sectional response to an unanticipated
narrative shock. Third, we use two-way fixed effects and difference-in-differences to localize the
mechanism. The premium is absorbed by firm fixed effects and roughly halved by sector fixed
effects, telling us that the durable signal is the persistent \emph{level} of a firm's (and its
sector's) AI exposure relative to peers, not the year-to-year wiggle in its wording.

Taken together, the evidence supports an \textbf{attention- and narrative-driven repricing}
interpretation. The textual evidence of firms' AI engagement accumulated quietly for a decade and
was barely priced. A salient catalyst---the consumer breakthrough of generative AI---then drew
investor attention to that pre-existing information, and capital flowed toward the firms and
sectors whose disclosures revealed the greatest AI exposure. This is the asset-pricing signature
of limited attention and narrative economics, observed not through a sentiment proxy but directly
in the semantic content of mandatory disclosure.

\paragraph{Contribution.} We make four contributions. \emph{First}, methodologically, we
demonstrate the value of embedding-based semantic measurement of disclosure for asset pricing: a
cosine-similarity-to-anchor design captures economically meaningful, multi-faceted exposure
(adoption vs.\ infrastructure vs.\ risk) that keyword counts cannot, and it does so at scale across
the entire S\&P~500 and a decade and a half of filings. \emph{Second}, substantively, we provide
the first text-based, look-ahead-free evidence that the cross-sectional pricing of AI exposure is
a post-2022 regime phenomenon rather than a long-standing premium. \emph{Third}, we discipline the
finding with a placebo across non-AI factors and an event study, distinguishing a genuine
AI-specific repricing from a mechanical or market-wide artifact. \emph{Fourth}, we document the
\emph{anatomy} of the effect---its concentration in smaller firms, its operation substantially
through a sector channel, and its disappearance under firm fixed effects---which sharpens the
economic interpretation and warns against naive trading strategies.

The remainder of the paper proceeds as follows. Section~\ref{sec:lit} situates the study in five
literatures. Section~\ref{sec:hyp} states our hypotheses. Section~\ref{sec:data} describes the
data and the construction of the semantic measures. Section~\ref{sec:method} lays out the
empirical methodology. Section~\ref{sec:results} presents the main results, and
Section~\ref{sec:robust} the robustness and identification tests. Section~\ref{sec:disc} discusses
mechanisms, Section~\ref{sec:limits} the limitations, and Section~\ref{sec:concl} concludes.

\section{Related Literature}\label{sec:lit}
Our study sits at the intersection of five strands of research: textual analysis in finance, the
pricing of corporate disclosure, technology and innovation in asset pricing, investor attention
and narrative economics, and the nascent literature on the economics of artificial intelligence.

\subsection{Textual analysis and natural language processing in finance}
A now-large literature extracts economic signals from financial text. The foundational work of
\citet{tetlock2007} showed that the tone of a Wall Street Journal column forecasts aggregate
returns, and \citet{tetlock2008} extended this to firm-level news. \citet{lm2011} demonstrated
that finance-specific sentiment dictionaries materially outperform general-purpose word lists in
10-K analysis, launching a wave of dictionary-based studies and the widely used Loughran--McDonald
word lists. \citet{lm2016} survey this field. A second generation moved from counting words to
modeling them: \citet{gentzkow2019} review text-as-data methods in economics, while
\citet{kelly2019textregression} and \citet{bybee2024narrative} use machine-learning
representations of news to construct interpretable factors and a ``state of the economy'' from
text. Our paper belongs to the most recent, embedding-based generation. Rather than counting AI
keywords---which conflate ``we use no AI'' with ``we are an AI leader'' and miss paraphrase---we
measure the \emph{semantic similarity} of disclosure to anchor concepts using a sentence
transformer \citep{reimers2019sbert}. This follows a growing practice of using contextual
embeddings to quantify disclosure content \citep{cong2022textfactor}, and is closest in spirit to
similarity-to-concept designs that score documents against reference passages.

\subsection{The pricing of corporate disclosure}
A parallel literature asks whether the \emph{content} and \emph{changes} of mandatory disclosure
predict returns. \citet{cohen2020lazy} show that firms that change the language of their 10-K
relative to the prior year (``lazy prices'' when they do not) experience predictable future
returns and operating outcomes. \citet{hoberg2010product} and \citet{hobergphillips2016} use 10-K
product descriptions to build text-based industry classifications and measures of product-market
fluidity that carry asset-pricing content. \citet{lopezlira2023risk} extract risk exposures from
10-K risk-factor sections. The general lesson is that the qualitative portions of filings contain
value-relevant information that markets process incompletely and with delay. We contribute a
focused, concept-level measure---AI engagement---and show that its pricing is strongly
state-dependent, a dimension the disclosure-pricing literature has largely left implicit.

\subsection{Technology, innovation, and the cross-section of returns}
Asset pricing has long sought to connect technological change to risk and return.
\citet{hirshleifer2013innovative} show that innovative efficiency predicts returns;
\citet{kogan2017technological} build market-based measures of innovation from patents and link
them to growth and reallocation. \citet{koganpapanikolaou2014} show that investment-specific
technology shocks shape the cross-section. A more recent literature studies AI exposure directly:
\citet{babina2024aiinvestment} measure firm-level AI investment from job-posting and resume data
and document faster growth among AI-adopting firms; \citet{eisfeldt2023generative} construct a
firm-level measure of \emph{generative}-AI exposure from the task content of occupations and show
that a generative-AI factor was priced precisely in the weeks around ChatGPT's release.
\citet{acemoglu2020wrong} and \citet{webb2020} study the labor-market exposure of firms to AI and
automation. Our contribution relative to this work is to measure AI exposure from the firm's
\emph{own disclosure}---a self-reported, narrative signal available for the entire index at annual
frequency---and to characterize the full time profile of its pricing across a decade and a half,
not only the event window. Where \citet{eisfeldt2023generative} use occupation-task exposure, we
use the language firms choose, providing a complementary and highly granular measure.

\subsection{Investor attention, salience, and narrative economics}
Why might value-relevant disclosure be priced only after a catalyst? The behavioral literature
offers a clear mechanism: limited investor attention. \citet{barberodean2008} show that retail
investors are drawn to attention-grabbing stocks; \citet{da2011} construct a direct attention
proxy from Google search and link it to prices; \citet{dellavigna2009} document under-reaction to
information released when attention is low; and \citet{hirshleifer2009driven} show that earnings
news is underweighted when investors are distracted. \citet{shiller2017narrative} argues that
contagious narratives drive major asset-price movements, and \citet{bakerwurgler2006} show that
sentiment differentially affects hard-to-value, speculative stocks. Our results provide a sharp
illustration: the information about firms' AI engagement existed in filings for years but was
repriced abruptly when a salient consumer event---ChatGPT---focused investor attention on AI. The
asymmetry we document (AI rewarded, ESG penalized after 2022) is itself consistent with a shift in
the prevailing market narrative.

\subsection{ESG and climate disclosure as a comparison case}
Finally, our placebo design connects to the literature on the pricing of ESG and climate
disclosure. \citet{pastor2021sustainable} model how sustainability preferences depress the
expected returns of green assets in equilibrium, and \citet{pastor2022dissecting} show that green
stocks' strong \emph{realized} returns through 2021 reflected unexpected increases in climate
concern rather than high expected returns. \citet{bolton2021risk} document a carbon premium.
Consistent with the post-2021 reversal in this literature, we find that high-ESG and high-climate
disclosure earned \emph{negative} spreads after 2022, exactly when AI disclosure began to earn
large positive ones---underscoring that the post-2022 re-rating reflects a thematic rotation
toward AI rather than a generic disclosure effect.

\section{Hypotheses}\label{sec:hyp}
Our analysis is organized around four hypotheses.

\begin{description}[leftmargin=2.4em,style=nextline]
\item[H1 (Unconditional pricing).] If AI exposure carries a stable risk premium or persistent
mispricing, then sorting on AI-disclosure intensity yields a positive and significant
risk-adjusted return spread over the full 2010--2026 sample.
\item[H2 (Regime dependence).] If instead AI exposure is repriced only once AI becomes salient,
the return spread is small before and large after the late-2022 generative-AI catalyst, with a
detectable structural break.
\item[H3 (AI-specificity).] If the post-2022 spread reflects a genuine AI repricing rather than a
mechanical or market-wide artifact, it appears for AI-related semantic factors but not for
unrelated factors (e.g.\ ESG, climate, macro risk).
\item[H4 (Cross-sectional, not within-firm).] If the market repriced \emph{which} firms are
AI-exposed rather than rewarding incremental wording, the effect is concentrated in the
cross-section and attenuates under firm and sector fixed effects.
\end{description}

\noindent The data, as we show, decisively favor H2--H4 over H1.

\section{Data and Variable Construction}\label{sec:data}
\subsection{The semantic factor panel}
The backbone of our study is the EDGAR Research Database, an infrastructure that downloads,
parses, and semantically scores SEC filings for S\&P~500 firms. For each filing the pipeline
(i)~extracts the canonical sections (Risk Factors, MD\&A, legal proceedings, financial statements,
and so on); (ii)~splits each section into overlapping text chunks; (iii)~embeds every chunk into a
384-dimensional vector with the \texttt{all-MiniLM-L6-v2} sentence transformer; and (iv)~scores
each chunk by cosine similarity to a set of \emph{anchor} passages defining eight economic
concepts. The anchors are short, deliberately written paragraphs that articulate the concept in
natural language; for example, the AI-adoption anchor describes a firm ``internally using
artificial intelligence, machine learning, generative AI, large language models, automation,
predictive analytics \dots to improve productivity, operations, efficiency, decision-making.''
Cosine similarity to such an anchor measures how closely a passage's \emph{meaning} matches the
concept, regardless of the exact words used---capturing paraphrase and context in a way keyword
counts cannot.

Chunk-level scores are aggregated to the filing level using section weights that emphasize the
parts of the document where the concept is most likely to be discussed substantively (e.g.\ MD\&A
and Risk Factors), and then to the firm-fiscal-year level. For each concept the database reports
the mean, maximum, standard deviation, top-5\% tail, and section-weighted mean (\texttt{wmean}) of
chunk similarities, along with first and second differences, percentage changes, and three-year
rolling statistics. The eight concepts comprise three AI facets---\emph{AI adoption}, \emph{AI
infrastructure}, \emph{AI disruption risk}---and five others used as placebos: \emph{ESG
commitment}, \emph{climate risk}, \emph{cybersecurity}, \emph{automation/labor}, and \emph{macro
risk}.

The resulting panel covers all 500 current S\&P~500 members across fiscal years 2010--2026, a
total of \num{7869} firm-years. We take the section-weighted mean (\texttt{wmean}) as our headline
signal because it up-weights the forward-looking, strategy-rich sections; we verify robustness to
the simple mean and the top-5\% tail in Section~\ref{sec:robust}. Our three primary signals are
therefore \texttt{ai\_adoption\_wmean} (\textbf{AI\,Adopt}), \texttt{ai\_infrastructure\_wmean}
(\textbf{AI\,Infra}), and \texttt{ai\_disruption\_risk\_wmean} (\textbf{AI\,Risk}); we also form a
standardized opportunity composite, $\text{AI\,Compos}=\tfrac12[z(\text{AI\,Adopt})+z(\text{AI\,Infra})]$.

\subsection{Price and factor data}
We download monthly adjusted-close prices for all 500 tickers from Yahoo Finance over 2009--2026
(handling share-class tickers such as BRK.B and BF.B), the SPDR S\&P~500 ETF (SPY) as the market
benchmark, and the Fama--French five factors, the momentum factor, and the risk-free rate from the
Kenneth R.\ French data library. Monthly simple returns are computed from adjusted prices, which
incorporate dividends and splits. A market-capitalization proxy is formed as shares outstanding
(from Yahoo Finance) times the historical price; we use it for value-weighting and as a size
control, acknowledging the shares-held-constant approximation.

\subsection{Timing and the construction of tradeable returns}
The single most important design choice in a disclosure--return study is timing. A 10-K covering
fiscal year $Y$ becomes public only when filed, typically in February or March of year $Y{+}1$. To
guarantee that our signal is in the public information set before any return is measured, we follow
the Fama--French portfolio convention: signals from fiscal year $Y$ are used to form portfolios at
the \textbf{end of June of $Y{+}1$}, and we measure the \textbf{forward twelve-month return from
July of $Y{+}1$ through June of $Y{+}2$}. The roughly three-month gap between the typical filing
date and June~30 ensures no look-ahead bias even for late filers. After requiring a valid forward
return with adequate monthly coverage, we obtain \num{6852} firm-years spanning formation years
2010--2024, with return realizations extending through mid-2026.

\subsection{Controls and standardization}
For each formation year we compute cross-sectional $z$-scores of every signal, so that all
coefficients are expressed per cross-sectional standard deviation and the powerful secular rise in
AI language (documented below) is differenced out---in any given year we compare firms to their
contemporaneous peers. We control for the three workhorse return determinants: \emph{size}
(log market-cap proxy), \emph{momentum} (the firm's own return over the twelve months ending at
formation, skipping the last month), and \emph{idiosyncratic volatility} (the annualized standard
deviation of monthly returns over the trailing year). In regressions, returns and continuous
controls are winsorized at the 1st and 99th percentiles to limit the influence of outliers; in
portfolio sorts we use raw returns.

\subsection{Summary statistics}
Table~\ref{tab:summ} reports summary statistics for the estimation sample. The average forward
twelve-month return is 18.6\% (reflecting the strong 2011--2026 bull market for U.S.\ large caps),
with a wide cross-sectional dispersion (standard deviation 41.5\%). The AI signals are small in
level---cosine similarities of disclosure to narrow anchors---but exhibit meaningful
cross-sectional variation. Table~\ref{tab:corr} reports correlations. Two features stand out.
First, the three AI signals are highly collinear (pairwise $\rho$ between 0.75 and 0.86): firms
that write more about adopting AI also write more about AI infrastructure and AI risk. They are
best understood as three lenses on a single latent construct---\emph{AI salience}---a point we
formalize with principal-components analysis in Section~\ref{sec:robust}. Second, the raw
(unconditional) correlation between each AI signal and the forward return is positive but small
($\approx 0.04$), foreshadowing that the unconditional relation is weak and that the action lies in
its time variation.

\input{tab_summary}
\input{tab_corr}

\subsection{The secular rise of AI language}
Before turning to returns, Figure~\ref{fig:trend} documents the raw phenomenon that motivates our
within-year design. The cross-sectional mean similarity to all three AI concepts trended steadily
upward through the 2010s and inflected sharply after 2022; by 2025--2026 AI language saturates
S\&P~500 disclosure. This is precisely why our tests rely on within-year cross-sectional ranking:
the interesting question is not whether the index as a whole talks more about AI over time (it
emphatically does), but whether, \emph{in a given year}, firms that talk about AI more than their
peers subsequently earn different returns. Figure~\ref{fig:sectorbox} shows that AI disclosure is
also unevenly distributed across sectors, concentrated in Information Technology, Communication
Services, and Industrials---a fact that will matter for interpreting the sector channel.

\begin{figure}[H]\centering
\includegraphics[width=0.9\textwidth]{fig4_ai_language_trend.png}
\caption{Cross-sectional mean of section-weighted cosine similarity to each AI anchor, by fiscal
year. All three measures trend upward and inflect after the 2022 launch of ChatGPT (dashed
line).}\label{fig:trend}
\end{figure}

\begin{figure}[H]\centering
\includegraphics[width=0.9\textwidth]{fig12_sector_box.png}
\caption{Distribution of the AI-adoption score by GICS sector (box plots, outliers suppressed; red
line = median). AI disclosure is concentrated in technology-intensive sectors.}\label{fig:sectorbox}
\end{figure}

\subsection{Which firms disclose the most AI? Descriptive evidence}
Table~\ref{tab:topfirms} lists the ten most AI-intensive S\&P~500 firms by average AI-adoption
score over the full sample. The list is instructive in two ways. First, it has strong face
validity: it features enterprise-technology and software firms (Hewlett Packard Enterprise, Intel,
IBM, Tyler Technologies, Autodesk), an algorithm-centric platform (Uber), and data-intensive
financials (BNY Mellon, Berkshire Hathaway), exactly the kinds of firms one would expect to write
substantively about deploying AI. Second, the list is \emph{not} confined to the technology sector:
industrials (Cummins) and energy (EOG Resources) appear, confirming that intensive AI disclosure
cuts across the GICS taxonomy and that our continuous, text-based measure captures within- and
cross-sector variation that a sector dummy would miss.

A further descriptive fact is central to our interpretation: the AI signal is highly
\emph{persistent}. The year-over-year cross-sectional rank autocorrelation of the AI composite is
0.88, and even at a five-year horizon it remains 0.73. Firms that are AI-intensive disclosers stay
AI-intensive; the measure captures a slow-moving structural characteristic of the firm rather than
transient wording. This persistence has two implications we exploit later: it explains why firm
fixed effects absorb most of the signal (there is little within-firm variation to identify), and it
underpins the difference-in-differences design, in which we fix each firm's exposure at its
pre-event level precisely because that level is stable.

\input{tab_topfirms}

\section{Methodology}\label{sec:method}
We bring six complementary research designs to bear, ranging from simple portfolio sorts to a
market-model event study. Using multiple designs with different identifying assumptions guards
against any single method's weaknesses and lets us triangulate the economic interpretation.

\subsection{Portfolio sorts}
Each formation year we sort firms into quintiles on a given (within-year standardized) AI signal
and compute the equal-weighted (EW) and value-weighted (VW) forward return of each quintile, plus
the high-minus-low (Q5$-$Q1) spread. We report the time-series average of the annual spreads and
test its significance with Newey--West \citep{neweywest1987} standard errors (three lags) to
account for any serial correlation in the overlapping return windows. Sorts are model-free and
transparent, and the EW--VW comparison reveals whether an effect is concentrated in larger or
smaller firms.

\subsection{Factor-model spanning of a long--short portfolio}
To ask whether the AI spread survives risk adjustment, we build an annually rebalanced,
equal-weighted, self-financing long--short portfolio (top minus bottom quintile), record its
monthly return, and regress it on (i)~the CAPM market factor and (ii)~the Fama--French five
factors plus momentum (FF6). The intercept (alpha) measures the average return unexplained by
known risk factors; we use heteroskedasticity- and autocorrelation-consistent (HAC) standard
errors. A significant alpha would indicate pricing beyond standard exposures.

\subsection{Two-way fixed-effects panel regressions}
We estimate
\begin{equation}
r_{i,t\to t+12} = \beta\, \text{AI}_{i,t} + \gamma' X_{i,t} + \alpha_i + \delta_t + \varepsilon_{i,t},
\end{equation}
where $r$ is the winsorized forward return, $\text{AI}_{i,t}$ the standardized signal, $X$ the
controls, and $\alpha_i,\delta_t$ firm and year fixed effects. Standard errors are double-clustered
by firm and year \citep{petersen2009,cameron2011}. Firm fixed effects absorb all
time-invariant cross-firm differences, so $\beta$ identifies whether \emph{within-firm} changes in
AI language predict returns; year fixed effects absorb common shocks. Contrasting this estimate
with the cross-sectional ones isolates whether the signal lives across or within firms (H4).

\subsection{Fama--MacBeth regressions}
We also run annual cross-sectional regressions of the forward return on the AI signal and controls
and average the slopes \`a la \citet{famamacbeth1973}, with Newey--West errors on the time series of
coefficients. This estimator weights each year equally and yields an interpretable
per-standard-deviation premium; plotting the annual slopes reveals the time profile of the premium
directly.

\subsection{Structural-break and difference-in-differences designs}
To test regime dependence (H2) we split the sample at the 30~November 2022 ChatGPT launch and
compare sub-period spreads, and we estimate pooled interaction models of the form
$r = \beta_1 \text{AI} + \beta_2 (\text{AI}\times \text{Post}) + \gamma' X + \varepsilon$, where
$\text{Post}=\mathbb{1}[\text{formation year}\ge 2022]$ and $\beta_2$ captures the change in the
premium. We complement this with a difference-in-differences design in which treatment is the
firm's \emph{pre-determined} AI exposure---fixed at its 2019--2021 average so it cannot respond to
the event---interacted with Post, within firm and year fixed effects. An event-time version
interacts pre-determined exposure with each year dummy (relative to 2021) to trace the dynamic
response.

\subsection{Market-model event study}
Finally, we run a classic event study \citep{mackinlay1997} around the ChatGPT launch. For each
stock we estimate a market model ($r_{i,d}=a_i+b_i r^{m}_{d}+e_{i,d}$) over the estimation window
January--October 2022 and compute daily abnormal returns $\text{AR}_{i,d}=r_{i,d}-(\hat a_i+\hat
b_i r^m_d)$ over the event window November 2022--June 2024. Cumulative abnormal returns (CAR) are
compared across terciles of \emph{pre-event} AI disclosure (from the FY2021 10-K), and we regress
firm-level CAR on pre-event exposure with size, beta, and sector controls. Because abnormal
returns net out market-beta exposure, the event study isolates the cross-sectional response to the
AI-narrative shock from the contemporaneous (and initially negative) market drift of 2022--2023.

\section{Results}\label{sec:results}
\subsection{Portfolio sorts: a weak unconditional relation}
Table~\ref{tab:sorts} reports the quintile sorts over the full sample. The relation between AI
disclosure and future returns is positive and broadly monotone but economically modest. The
equal-weighted Q5$-$Q1 spread is $+4.6$ percentage points per year for AI adoption \tstat{1.83},
$+3.9$ for AI infrastructure \tstat{1.58}, $+5.3$ for AI disruption risk \tstat{2.27}, and $+4.9$
for the composite \tstat{1.88}. Value-weighting attenuates the adoption and composite spreads,
indicating the effect is stronger among smaller index members. The left panel of
Figure~\ref{fig:quint} displays the full-sample pattern: returns rise gently from the lowest to
the highest AI quintile. On its own, this is suggestive but far from decisive evidence---and, as
we now show, it averages over two very different regimes.

\input{tab_sorts}

\subsection{Does AI disclosure earn risk-adjusted alpha unconditionally? (H1)}
Table~\ref{tab:alpha} reports CAPM and FF6 alphas of the long--short portfolios. The
AI-opportunity composite earns a raw spread of about $+2.3\%$ per year, of which the FF6 alpha is
$+1.9\%$---but statistically insignificant \tstat{1.24}. The AI-adoption and AI-infrastructure
long--shorts are similarly insignificant after risk adjustment. The one exception is the
AI-disruption-risk spread, whose FF6 alpha of $+2.8\%$ per year is marginally significant
\tstat{2.10}, hinting at a small risk premium for firms that flag AI as a competitive threat. The
factor loadings (not the focus here) show the AI long--short is roughly market-neutral with modest
negative profitability (RMW) and investment (CMA) tilts, consistent with a growth-oriented long
leg. The verdict on H1 is clear: \textbf{unconditionally, AI disclosure is at best weakly
priced.} This is not the end of the story but the beginning of it.

\input{tab_alpha}

\subsection{Panel and Fama--MacBeth estimates}
Table~\ref{tab:panel} juxtaposes the two-way fixed-effects panel with the Fama--MacBeth estimator.
The contrast is instructive. In the Fama--MacBeth (cross-sectional) regressions, the AI-adoption
and AI-infrastructure premia are positive---$+1.3$ and $+1.6$ percentage points per
standard-deviation---with \textit{t}-statistics near 1.6--1.7, indicating a weak but positive
cross-sectional premium on average. In the two-way fixed-effects panel, however, every AI
coefficient collapses to essentially zero. Because firm fixed effects remove all permanent
cross-firm differences, this tells us the signal lives \emph{across} firms, not \emph{within}
them: a given firm writing relatively more about AI from one year to the next does not predict its
own subsequent return (anticipating H4). Throughout, size carries the familiar negative premium
($t\approx-8$) and volatility a positive one, consistent with the broader cross-sectional
literature in this large-cap universe.

\input{tab_panel}

\subsection{The structural break: AI pricing switched on after ChatGPT (H2)}
The full-sample averages mask a sharp regime change that is the paper's central result. Splitting
the sample at the November 2022 ChatGPT launch, the center and right panels of
Figure~\ref{fig:quint} reveal a striking transformation: the quintile pattern is nearly flat
before 2022 and steeply increasing afterward. The high-minus-low composite spread is $+2.4$
percentage points per year pre-ChatGPT but $+14.6$ points per year post-ChatGPT---a more than
sixfold increase.

\begin{figure}[H]\centering
\includegraphics[width=\textwidth]{fig1_quintile_returns.png}
\caption{Mean forward twelve-month return by AI-disclosure composite quintile, full sample and by
sub-period. The high-minus-low spread is $+2.4$ percentage points before ChatGPT and $+14.6$ points
after.}\label{fig:quint}
\end{figure}

Table~\ref{tab:break} quantifies the break two ways. In sub-period sorts, the AI-adoption Q5$-$Q1
spread rises from $+2.2\%$ per year \tstat{2.04} pre-2022 to $+14.3\%$ \tstat{2.15} post-2022; the
composite rises from $+2.5\%$ to $+14.6\%$; AI infrastructure from $+1.4\%$ to $+13.6\%$. In the
pooled interaction regressions, the post-2022 incremental slope is positive for all three signals
and statistically significant for AI infrastructure \tstat{2.63}---the facet most directly tied to
the chip, cloud, data-center, and power firms that led the 2023--2025 rally. Figure~\ref{fig:spread}
shows the annual spreads: small and mixed-sign through 2021, uniformly large and positive for every
post-2022 cohort.

\input{tab_break}

\begin{figure}[H]\centering
\includegraphics[width=0.9\textwidth]{fig3_annual_spread.png}
\caption{Annual high-minus-low (Q5$-$Q1) forward-return spread on the AI-disclosure composite, by
10-K fiscal year. Post-ChatGPT cohorts (red) are uniformly and substantially positive.}\label{fig:spread}
\end{figure}

The same regime change is visible in two further lenses. Figure~\ref{fig:cum} plots the cumulative
value of one dollar invested in the equal-weighted, annually rebalanced, self-financing
high-minus-low portfolio: it drifts sideways-to-up for a decade and then inflects upward after the
dashed ChatGPT line. Figure~\ref{fig:fmyear} plots the year-by-year Fama--MacBeth slope on the AI
composite (with controls): it hovers near zero for the entire 2010s and then turns sharply
positive, peaking in the 2024 cohort. The latter figure also delivers an important nuance: the
size-controlled premium is largest in the most recent cohort, consistent with the AI
\emph{infrastructure} super-cycle of 2024--2025 (the build-out of data centers and accelerated
compute). We return to this concentration when discussing limitations.

\begin{figure}[H]\centering
\includegraphics[width=0.86\textwidth]{fig2_cumulative_ls.png}
\caption{Cumulative value of \$1 invested in the equal-weighted, annually rebalanced, self-financing
high-minus-low AI-disclosure portfolio (AI opportunity composite and AI disruption-risk).}\label{fig:cum}
\end{figure}

\begin{figure}[H]\centering
\includegraphics[width=0.9\textwidth]{fig6_fm_byyear.png}
\caption{Year-by-year Fama--MacBeth slope on the AI-disclosure composite (forward return per
cross-sectional standard deviation), controlling for size, momentum, and volatility. Bars are
$\pm1$ standard error.}\label{fig:fmyear}
\end{figure}

Finally, Figure~\ref{fig:bin} confirms the pattern is not an artifact of a few influential
observations: the binned relation between the AI composite and forward returns is essentially flat
in the pre-ChatGPT sub-period and distinctly upward-sloping afterward, across the entire
distribution of AI disclosure rather than only at the extreme quintiles.

\begin{figure}[H]\centering
\includegraphics[width=\textwidth]{fig5_binscatter.png}
\caption{Binned scatter of forward twelve-month return against the AI-disclosure composite, by
sub-period. Each point is a portfolio of firms in a bin of the AI signal; the line is an OLS fit.
The conditional slope steepens markedly after ChatGPT.}\label{fig:bin}
\end{figure}

\subsection{The anatomy of the premium: size and sector}
Where in the cross-section does the post-2022 premium concentrate? Table~\ref{tab:doublesort} and
Figure~\ref{fig:double} report an AI~$\times$~size double sort. Before 2022 the AI gradient is flat
within both size halves. After 2022 it is steep, and dramatically so among smaller firms: small,
high-AI firms earned a mean forward return of 34.6\%, versus 19.5\% for small, low-AI firms---a
15-point gap---while the corresponding gap among large firms is 7 points. The premium is thus
\emph{not} merely a mega-cap phenomenon (a point we reinforce in Section~\ref{sec:robust}); if
anything it is strongest among the smaller, less-followed names, consistent with an
attention-based mechanism in which information diffuses more slowly into smaller stocks.

\input{tab_doublesort}

\begin{figure}[H]\centering
\includegraphics[width=\textwidth]{fig9_double_sort.png}
\caption{AI-disclosure $\times$ firm-size double sort: mean forward twelve-month return by AI
tercile within small and large firms, pre- and post-ChatGPT.}\label{fig:double}
\end{figure}

\subsection{Economic significance}
The statistical break is matched by an economic one. The Sharpe ratio of the equal-weighted
AI-composite long--short portfolio---a direct gauge of risk-adjusted attractiveness---rises from
0.15 in the pre-ChatGPT decade to 1.33 afterward; for the AI-disruption-risk spread it rises from
0.33 to 1.15. A Sharpe ratio above one is exceptional for a single-characteristic, self-financing
equity strategy, and its emergence precisely at the ChatGPT boundary is hard to reconcile with a
constant risk premium. The point estimates in the quintile sorts are equally vivid: a 14-point
annual spread between the top and bottom AI quintiles is roughly three times the unconditional
equity risk premium and an order of magnitude larger than the pre-2022 spread. Even acknowledging
that a portion of this reflects the higher realized volatility of AI-exposed firms (documented in
Section~\ref{sec:robust}) and the broad technology rally, the magnitude of the post-2022 re-rating
is among the largest single-theme cross-sectional effects in the recent asset-pricing record. The
contrast between an essentially flat pre-period and a near-vertical post-period is, to us, the most
compelling single piece of evidence that a genuine regime change occurred.

\section{Robustness and Identification}\label{sec:robust}
The structural break is economically large; this section asks how much we should believe it. We
subject it to a placebo across non-AI factors, an event study, a difference-in-differences design,
and a battery of construction and universe robustness checks.

\subsection{Placebo: is the effect specific to AI? (H3)}
If the post-2022 spread were a mechanical artifact of our pipeline---or a generic ``high-disclosure
firms did well'' phenomenon---it should appear for any semantic factor. Table~\ref{tab:placebo} and
Figure~\ref{fig:placebo} re-run the entire sub-period and interaction design on all eight factors.
The result strongly supports AI-specificity. The three AI factors show large positive post-2022
spreads ($+13.6$ to $+14.3$ points). Macro-risk language shows essentially no spread ($-0.5$,
$t=-0.1$). And, strikingly, ESG-commitment and climate-risk disclosure earn \emph{negative}
post-2022 spreads ($-8.8$ and $-5.0$ points, both significant)---high-ESG firms \emph{underperformed}
after 2022, in line with the post-2021 reversal in the sustainable-investing literature
\citep{pastor2022dissecting}. The one non-AI factor that does show a large positive post-2022
spread is cybersecurity ($+13.5$); we read this not as a violation but as confirmation, since
cybersecurity disclosure is tightly bundled with technology exposure and is itself increasingly
AI-related. The cleanest contrast---AI rewarded, ESG and climate penalized---makes a powerful case
that the post-ChatGPT re-rating reflects a thematic rotation \emph{toward AI}, not a generic
disclosure effect.

\input{tab_placebo}

\begin{figure}[H]\centering
\includegraphics[width=0.92\textwidth]{fig7_placebo.png}
\caption{Placebo test: post-ChatGPT high-minus-low return spread for each of the eight semantic
factors. AI-related factors (red) cluster at the top; ESG and climate disclosure earn negative
spreads.}\label{fig:placebo}
\end{figure}

\subsection{Event study around ChatGPT}
The portfolio results use annual formation and forward returns; the event study provides a
higher-frequency, cleanly identified complement. Table~\ref{tab:event} and Figure~\ref{fig:event}
report market-model cumulative abnormal returns by pre-event AI tercile. The ordering is monotone
at every horizon and widens steadily over time. The High-minus-Low CAR difference is $+2.9$ points
at 60 trading days \tstat{1.77}, $+10.1$ points at one year \tstat{1.97}, and $+14.6$ points at
eighteen months \tstat{1.89}. Because these are abnormal returns---purged of each firm's market
beta---the spread cannot be attributed to high-AI firms simply being higher-beta in a recovering
market. A cross-sectional regression of one-year CAR on pre-event AI exposure yields $+3.1$ points
per standard deviation \tstat{1.39}; adding sector fixed effects halves it to $+2.1$ points
\tstat{0.99}, again locating much---but not all---of the effect in a sector channel.
Figure~\ref{fig:eventbin} shows the relation is smooth across the exposure distribution, not driven
by a few outliers.

\input{tab_event}

\begin{figure}[H]\centering
\includegraphics[width=\textwidth]{fig8_event_car.png}
\caption{ChatGPT event study. Left: mean cumulative abnormal return (market model) by pre-event AI
tercile. Right: the High$-$Low CAR spread.}\label{fig:event}
\end{figure}

\begin{figure}[H]\centering
\includegraphics[width=0.66\textwidth]{fig13_event_binscatter.png}
\caption{One-year cumulative abnormal return after ChatGPT versus pre-event AI exposure (binned
scatter with fitted line).}\label{fig:eventbin}
\end{figure}

\subsection{Difference-in-differences and the cross-sectional nature of the effect (H4)}
Table~\ref{tab:did} reports difference-in-differences estimates in which treatment is the firm's
pre-determined (2019--2021) AI exposure interacted with the post-2022 indicator, within firm and
year fixed effects. The continuous-exposure DiD coefficient is $+2.7$ percentage points per
standard deviation \tstat{1.51}; the binary version is $+1.7$ points \tstat{0.78}. These are
positive and economically meaningful but weaker than the cross-sectional estimates---exactly the
pattern predicted by H4. Once firm fixed effects absorb the permanent component of AI exposure,
only the residual within-firm variation identifies the effect, and that variation is small. The
event-time profile in Figure~\ref{fig:eventtime} reinforces this reading: the return premium on a
firm's \emph{fixed} pre-event AI exposure is flat and near zero through 2021 and jumps after 2022,
peaking in 2024. The market did not reward firms for adding AI sentences to their filings; it
repriced the firms whose pre-existing, persistent exposure made them AI winners once the narrative
turned.

\input{tab_did}

\begin{figure}[H]\centering
\includegraphics[width=0.82\textwidth]{fig11_eventtime.png}
\caption{Event-study coefficients: the return premium on pre-determined (2019--2021) AI exposure by
year (95\% confidence intervals), relative to 2021. The premium activates after ChatGPT.}\label{fig:eventtime}
\end{figure}

\subsection{Construction, horizon, and universe robustness}
Table~\ref{tab:robust} collects the remaining checks. The post-2022 spread is essentially unchanged
whether the signal is the section-weighted mean (base, $+14.6$), the simple mean ($+13.4$), or the
top-5\% tail ($+14.9$). It holds at six-, twelve-, and twenty-four-month horizons (growing with the
horizon, as a persistent premium should). It survives replacing the three AI signals with their
first principal component---which alone explains 86.8\% of their common variance, confirming the
``single latent AI construct'' interpretation---yielding a post-2022 PC1 spread of $+13.4$ points
\tstat{1.97}. Most importantly for the mega-cap concern, the spread is $+14.1$ points \tstat{1.92}
when the Magnificent~7 are excluded and $+14.5$ points \tstat{1.93} when the ten largest firms are
dropped (Figure~\ref{fig:mag7}). The AI re-rating is a broad-based feature of the index, not the
mechanical footprint of a handful of trillion-dollar winners. Finally, the sector-neutral
composite---which compares firms only to others \emph{within} their GICS sector---produces a much
smaller and insignificant post-2022 spread ($+3.8$ points, $t=0.84$). This is not a failure of the
result but a sharpening of it: a large share of the AI premium operates through a
\emph{between-sector} channel, as the market re-rated AI-exposed sectors (technology above all)
relative to others, with a smaller within-sector residual.

\input{tab_robust}

\begin{figure}[H]\centering
\includegraphics[width=0.78\textwidth]{fig10_mag7.png}
\caption{Robustness of the AI return spread to excluding mega-cap winners. The post-2022 composite
spread is stable when the Magnificent~7 or the ten largest firms are removed.}\label{fig:mag7}
\end{figure}

\subsection{An alternative performance outcome: risk}
Performance is not only about mean returns. We also relate AI disclosure to subsequent realized
volatility. High-AI firms exhibit modestly higher forward volatility, and the gap widens after
2022 (a Q5$-$Q1 difference of about 2.9 annualized volatility points, $t=3.1$). The post-2022 AI
premium therefore came bundled with somewhat higher risk---part, though by no means all, of the
elevated returns is compensation for the greater volatility of AI-exposed firms during a period of
rapid technological repricing.

\section{Discussion: Mechanism and Interpretation}\label{sec:disc}
What economic mechanism best fits this constellation of facts? We find an \emph{attention- and
narrative-driven repricing} the most parsimonious explanation, and we read the evidence as follows.

\paragraph{The information pre-existed; the attention did not.} Firms disclosed their AI engagement
for years before it was priced. The semantic content of those disclosures was, in cross-section,
only faintly related to returns through 2021. What changed in late 2022 was not the information but
its \emph{salience}: a consumer breakthrough (ChatGPT) made AI the dominant market narrative
\citep{shiller2017narrative}, drawing investor attention \citep{da2011,barberodean2008} to a body
of disclosure that had been hiding in plain sight. The abrupt, near-simultaneous repricing across
the cross-section---visible in the event study within months of the launch---is the signature of an
attention shock rather than the gradual accrual of a risk premium.

\paragraph{It is a re-rating of levels, not of changes.} The disappearance of the effect under firm
fixed effects, the muted difference-in-differences estimates, and the event-time profile all point
the same way: the market repriced the persistent \emph{level} of a firm's AI exposure, not the
year-to-year change in its wording. This has a clear practical corollary---firms cannot
mechanically earn the premium by adding AI boilerplate to their filings (and our top-5\% and
simple-mean robustness suggests the measure is not easily gamed)---and a clear methodological one:
studies that difference out firm effects will understate the cross-sectional importance of
disclosure content.

\paragraph{It operates substantially, but not entirely, through sectors.} The sector-neutral and
sector-fixed-effects results show that much of the AI premium reflects the re-rating of AI-exposed
\emph{sectors}---technology, communication services, and the industrial firms supplying the AI
build-out---relative to the rest of the index. This is economically sensible: the generative-AI
boom was, to first order, a sectoral reallocation of expected cash flows and valuations. Yet a
positive within-sector residual remains in the event study, indicating that the firm's own
disclosure carried incremental information even among industry peers.

\paragraph{The thematic asymmetry is telling.} Perhaps the most revealing single fact is the
placebo: in the very same post-2022 window, high-AI firms were rewarded while high-ESG and
high-climate firms were penalized. Markets did not simply reward verbose or virtuous-sounding
disclosure; they rotated toward one theme (AI) and away from another (sustainability). This is
difficult to reconcile with a purely mechanical or risk-based account and natural under a
narrative-rotation view of investor attention.

\paragraph{Relation to the AI-asset-pricing literature.} Our findings complement
\citet{eisfeldt2023generative}, who show that an occupation-task measure of generative-AI exposure
was priced around ChatGPT. We reach a convergent conclusion from an entirely different
measurement---firms' own disclosure language---and extend it along the time dimension, showing that
the same exposure was \emph{un}priced for a decade beforehand. The agreement of two independent
measures on the timing and sign of the effect strengthens confidence that a real AI repricing
occurred.

\subsection{Implications}
Our findings carry distinct implications for three audiences. \emph{For investors and quantitative
strategists}, the results are a caution as much as an opportunity. The AI-disclosure signal
delivered a striking post-2022 Sharpe ratio, but the same signal was nearly inert for the preceding
decade, and the effect is absorbed by firm fixed effects---so a strategy that traded on
year-to-year changes in AI verbosity would have captured little. The tradeable content lies in the
persistent cross-sectional \emph{level} of AI exposure, and even that is partly a sector bet whose
durability is unproven. The honest reading is that disclosure-based AI signals are best viewed as a
lens on a regime-dependent theme rather than as a stable alpha source. \emph{For corporate managers
and disclosure policy}, the asymmetry between AI (rewarded) and ESG/climate (penalized) after 2022
illustrates how powerfully the prevailing market narrative conditions the valuation consequences of
what firms choose to emphasize. Because the premium attaches to persistent, hard-to-fake exposure
rather than to boilerplate, the result does not invite cheap disclosure manipulation; but it does
suggest that the market reads filings for genuine strategic content and reprices it when attention
turns. \emph{For researchers}, the study demonstrates that embedding-based semantic measurement
uncovers pricing dynamics invisible to both keyword counts (which cannot distinguish meaning) and
firm-differenced panels (which discard the persistent cross-sectional variation that turns out to
matter). The broader agenda---measuring multi-faceted, concept-level exposure from disclosure at
scale and studying its state-dependent pricing---appears fertile.

\subsection{Summary of the evidence across designs}
It is worth stepping back to see how the designs corroborate one another, since each has different
weaknesses. The portfolio sorts establish the raw magnitude and its post-2022 timing but are
model-free and could in principle reflect risk; the factor-model spanning regressions show the
unconditional spread is not explained by standard factors yet is statistically weak over the full
sample; the sub-period and interaction tests localize the effect to the post-ChatGPT regime; the
placebo rules out a mechanical or generic-disclosure explanation by showing the opposite sign for
ESG and climate; the event study, using daily abnormal returns purged of beta, provides
cleanly-timed, cross-sectional confirmation within months of the catalyst; and the fixed-effects
and difference-in-differences estimates pin down that the priced object is persistent firm- and
sector-level exposure, not incremental wording. No single design is decisive, but their
convergence---on the same sign, the same timing, and a consistent economic story---is what gives us
confidence in the conclusion. A spurious result would have to survive all six designs
simultaneously while also producing the precise ESG/climate sign reversal, which strains credulity.

\section{Limitations}\label{sec:limits}
Several caveats temper our conclusions. \emph{First}, the post-2022 window contains only three
formation cohorts, and the size-controlled premium is concentrated in the most recent (2024)
cohort, which captures the 2024--2025 AI-infrastructure capital-expenditure boom. The regime result
is economically large and consistent across designs (sorts, event study, DiD), but its post-period
rests on a short sample and could in principle reflect a single extraordinary episode rather than a
durable new regime; only future data can settle this. \emph{Second}, our universe is current
S\&P~500 membership, which introduces survivorship and look-ahead-at-the-index-level bias: firms
that exited the index are absent, and current members were selected partly on past success. This
biases the level of returns upward but is less likely to drive the \emph{cross-sectional}
AI--return relation that is our focus. \emph{Third}, the market-capitalization proxy holds shares
outstanding at their latest value, approximating historical size with error. \emph{Fourth}, the
semantic anchors, though carefully written, embed the authors' definitional choices; different
anchor wording would yield somewhat different scores, although the robustness to mean/tail
aggregations and to PCA mitigates this concern. \emph{Fifth}, disclosure intensity is correlated
with sector, and while our within-year, size-, momentum-, and volatility-controlled tests and the
sector-FE event study address this, we cannot fully separate ``AI exposure'' from ``technology
sector'' as the priced characteristic---indeed, our sector-channel results suggest the two are
deeply intertwined. \emph{Finally}, our designs establish a robust association and an
event-study-based timing argument, but not strict causality; the attention-repricing interpretation
is the most consistent with the evidence rather than the only logically possible one.

\section{Conclusion}\label{sec:concl}
Using a semantic, embedding-based measure of artificial-intelligence language in the SEC filings of
S\&P~500 firms, linked to Yahoo Finance returns and standard factor data, we have shown that the
pricing of AI disclosure is a \emph{conditional, regime-dependent} phenomenon. For most of the
2010s, how intensively a firm wrote about adopting, building, or being threatened by AI bore only a
faint relation to its subsequent returns, and a risk-adjusted long--short strategy earned no
reliable alpha. After the launch of ChatGPT in late 2022, the same disclosure became strongly
priced: the high-minus-low AI-disclosure return spread jumped from roughly $+2$ to roughly $+15$
percentage points per year. The break is robust to signal definition, horizon, dimension reduction,
and the exclusion of mega-cap winners; it is specific to AI-related language (ESG and climate
disclosure, tested identically, were \emph{penalized} in the same window); it is corroborated by a
market-model event study in which high-AI firms earned $+10$ points of additional abnormal return
in the year after ChatGPT; and it operates as a cross-sectional re-rating of persistent firm- and
sector-level exposure rather than a reward for year-to-year wording, fading under firm fixed
effects.

The economic message is that markets price the \emph{language} of AI---but only once AI commands
their attention. The methodological message is that embedding-based measurement of disclosure
content opens a productive avenue for asset pricing, capturing economically meaningful, multi-faceted
exposures at scale and revealing pricing dynamics that keyword counts and firm-differenced designs
would miss. As generative AI continues to reshape the corporate landscape, the semantic footprint
it leaves in mandatory disclosure offers a rich, and now demonstrably value-relevant, signal.

\clearpage
\begin{thebibliography}{99}\small
\setlength{\itemsep}{2pt}
\bibitem[Acemoglu and Restrepo(2020)]{acemoglu2020wrong} Acemoglu, D., and P.\ Restrepo (2020). Robots and Jobs: Evidence from US Labor Markets. \textit{Journal of Political Economy} 128(6), 2188--2244.
\bibitem[Babina et al.(2024)]{babina2024aiinvestment} Babina, T., A.\ Fedyk, A.\ He, and J.\ Hodson (2024). Artificial Intelligence, Firm Growth, and Product Innovation. \textit{Journal of Financial Economics} 151, 103745.
\bibitem[Baker and Wurgler(2006)]{bakerwurgler2006} Baker, M., and J.\ Wurgler (2006). Investor Sentiment and the Cross-Section of Stock Returns. \textit{Journal of Finance} 61(4), 1645--1680.
\bibitem[Barber and Odean(2008)]{barberodean2008} Barber, B.\ M., and T.\ Odean (2008). All That Glitters: The Effect of Attention and News on the Buying Behavior of Individual and Institutional Investors. \textit{Review of Financial Studies} 21(2), 785--818.
\bibitem[Bolton and Kacperczyk(2021)]{bolton2021risk} Bolton, P., and M.\ Kacperczyk (2021). Do Investors Care about Carbon Risk? \textit{Journal of Financial Economics} 142(2), 517--549.
\bibitem[Bybee et al.(2024)]{bybee2024narrative} Bybee, L., B.\ Kelly, A.\ Manela, and D.\ Xiu (2024). Business News and Business Cycles. \textit{Journal of Finance} 79(5), 3105--3147.
\bibitem[Cameron et al.(2011)]{cameron2011} Cameron, A.\ C., J.\ B.\ Gelbach, and D.\ L.\ Miller (2011). Robust Inference with Multiway Clustering. \textit{Journal of Business \& Economic Statistics} 29(2), 238--249.
\bibitem[Cohen et al.(2020)]{cohen2020lazy} Cohen, L., C.\ Malloy, and Q.\ Nguyen (2020). Lazy Prices. \textit{Journal of Finance} 75(3), 1371--1415.
\bibitem[Cong et al.(2022)]{cong2022textfactor} Cong, L.\ W., T.\ Liang, and X.\ Zhang (2022). Textual Factors: A Scalable, Interpretable, and Data-Driven Approach to Analyzing Unstructured Information. Working paper.
\bibitem[Da et al.(2011)]{da2011} Da, Z., J.\ Engelberg, and P.\ Gao (2011). In Search of Attention. \textit{Journal of Finance} 66(5), 1461--1499.
\bibitem[DellaVigna and Pollet(2009)]{dellavigna2009} DellaVigna, S., and J.\ M.\ Pollet (2009). Investor Inattention and Friday Earnings Announcements. \textit{Journal of Finance} 64(2), 709--749.
\bibitem[Eisfeldt et al.(2023)]{eisfeldt2023generative} Eisfeldt, A.\ L., G.\ Schubert, and M.\ B.\ Zhang (2023). Generative AI and Firm Values. NBER Working Paper 31222.
\bibitem[Fama and French(2015)]{famafrench2015} Fama, E.\ F., and K.\ R.\ French (2015). A Five-Factor Asset Pricing Model. \textit{Journal of Financial Economics} 116(1), 1--22.
\bibitem[Fama and MacBeth(1973)]{famamacbeth1973} Fama, E.\ F., and J.\ D.\ MacBeth (1973). Risk, Return, and Equilibrium: Empirical Tests. \textit{Journal of Political Economy} 81(3), 607--636.
\bibitem[Gentzkow et al.(2019)]{gentzkow2019} Gentzkow, M., B.\ Kelly, and M.\ Taddy (2019). Text as Data. \textit{Journal of Economic Literature} 57(3), 535--574.
\bibitem[Hirshleifer et al.(2009)]{hirshleifer2009driven} Hirshleifer, D., S.\ S.\ Lim, and S.\ H.\ Teoh (2009). Driven to Distraction: Extraneous Events and Underreaction to Earnings News. \textit{Journal of Finance} 64(5), 2289--2325.
\bibitem[Hirshleifer et al.(2013)]{hirshleifer2013innovative} Hirshleifer, D., P.-H.\ Hsu, and D.\ Li (2013). Innovative Efficiency and Stock Returns. \textit{Journal of Financial Economics} 107(3), 632--654.
\bibitem[Hoberg and Phillips(2010)]{hoberg2010product} Hoberg, G., and G.\ Phillips (2010). Product Market Synergies and Competition in Mergers and Acquisitions: A Text-Based Analysis. \textit{Review of Financial Studies} 23(10), 3773--3811.
\bibitem[Hoberg and Phillips(2016)]{hobergphillips2016} Hoberg, G., and G.\ Phillips (2016). Text-Based Network Industries and Endogenous Product Differentiation. \textit{Journal of Political Economy} 124(5), 1423--1465.
\bibitem[Kelly et al.(2019)]{kelly2019textregression} Kelly, B., A.\ Manela, and A.\ Moreira (2019). Text Selection. Working paper / \textit{Journal of Business \& Economic Statistics} (2021).
\bibitem[Kogan and Papanikolaou(2014)]{koganpapanikolaou2014} Kogan, L., and D.\ Papanikolaou (2014). Growth Opportunities, Technology Shocks, and Asset Prices. \textit{Journal of Finance} 69(2), 675--718.
\bibitem[Kogan et al.(2017)]{kogan2017technological} Kogan, L., D.\ Papanikolaou, A.\ Seru, and N.\ Stoffman (2017). Technological Innovation, Resource Allocation, and Growth. \textit{Quarterly Journal of Economics} 132(2), 665--712.
\bibitem[Lopez-Lira(2023)]{lopezlira2023risk} Lopez-Lira, A. (2023). Risk Factors That Matter: Textual Analysis of Risk Disclosures for the Cross-Section of Returns. Working paper.
\bibitem[Loughran and McDonald(2011)]{lm2011} Loughran, T., and B.\ McDonald (2011). When Is a Liability Not a Liability? Textual Analysis, Dictionaries, and 10-Ks. \textit{Journal of Finance} 66(1), 35--65.
\bibitem[Loughran and McDonald(2016)]{lm2016} Loughran, T., and B.\ McDonald (2016). Textual Analysis in Accounting and Finance: A Survey. \textit{Journal of Accounting Research} 54(4), 1187--1230.
\bibitem[MacKinlay(1997)]{mackinlay1997} MacKinlay, A.\ C. (1997). Event Studies in Economics and Finance. \textit{Journal of Economic Literature} 35(1), 13--39.
\bibitem[Newey and West(1987)]{neweywest1987} Newey, W.\ K., and K.\ D.\ West (1987). A Simple, Positive Semi-Definite, Heteroskedasticity and Autocorrelation Consistent Covariance Matrix. \textit{Econometrica} 55(3), 703--708.
\bibitem[P\'astor et al.(2021)]{pastor2021sustainable} P\'astor, \v{L}., R.\ F.\ Stambaugh, and L.\ A.\ Taylor (2021). Sustainable Investing in Equilibrium. \textit{Journal of Financial Economics} 142(2), 550--571.
\bibitem[P\'astor et al.(2022)]{pastor2022dissecting} P\'astor, \v{L}., R.\ F.\ Stambaugh, and L.\ A.\ Taylor (2022). Dissecting Green Returns. \textit{Journal of Financial Economics} 146(2), 403--424.
\bibitem[Petersen(2009)]{petersen2009} Petersen, M.\ A. (2009). Estimating Standard Errors in Finance Panel Data Sets: Comparing Approaches. \textit{Review of Financial Studies} 22(1), 435--480.
\bibitem[Reimers and Gurevych(2019)]{reimers2019sbert} Reimers, N., and I.\ Gurevych (2019). Sentence-BERT: Sentence Embeddings Using Siamese BERT-Networks. \textit{EMNLP}.
\bibitem[Shiller(2017)]{shiller2017narrative} Shiller, R.\ J. (2017). Narrative Economics. \textit{American Economic Review} 107(4), 967--1004.
\bibitem[Tetlock(2007)]{tetlock2007} Tetlock, P.\ C. (2007). Giving Content to Investor Sentiment: The Role of Media in the Stock Market. \textit{Journal of Finance} 62(3), 1139--1168.
\bibitem[Tetlock et al.(2008)]{tetlock2008} Tetlock, P.\ C., M.\ Saar-Tsechansky, and S.\ Macskassy (2008). More Than Words: Quantifying Language to Measure Firms' Fundamentals. \textit{Journal of Finance} 63(3), 1437--1467.
\bibitem[Webb(2020)]{webb2020} Webb, M. (2020). The Impact of Artificial Intelligence on the Labor Market. Working paper, Stanford University.
\end{thebibliography}

\clearpage
\appendix
\section*{Appendix A. Variable Definitions}
\addcontentsline{toc}{section}{Appendix A. Variable Definitions}
\begin{description}[leftmargin=1.5em,style=nextline,itemsep=2pt]
\item[AI\,Adopt (\texttt{ai\_adoption\_wmean}).] Section-weighted mean cosine similarity of a firm's 10-K chunks to the AI-internal-adoption anchor in fiscal year $Y$.
\item[AI\,Infra (\texttt{ai\_infrastructure\_wmean}).] Section-weighted mean similarity to the AI-compute-infrastructure anchor.
\item[AI\,Risk (\texttt{ai\_disruption\_risk\_wmean}).] Section-weighted mean similarity to the AI-product-disruption-risk anchor.
\item[AI\,Compos.] $\tfrac12[z(\text{AI\,Adopt})+z(\text{AI\,Infra})]$, with $z$ the within-year cross-sectional standardization.
\item[Forward return.] Buy-and-hold simple return from July of $Y{+}1$ to June of $Y{+}2$ (adjusted close, dividends reinvested).
\item[Market-adjusted return.] Forward return minus the contemporaneous SPY return.
\item[Size.] Natural log of the market-cap proxy (latest shares outstanding $\times$ price) at formation.
\item[Momentum.] Own return over the 12 months ending at formation, skipping the last month.
\item[Volatility.] Annualized standard deviation of monthly returns over the trailing 12 months.
\item[Post.] Indicator equal to one for formation (fiscal) years $\ge 2022$.
\item[CAR.] Cumulative market-model abnormal return; market model estimated Jan--Oct 2022, event window Nov 2022--Jun 2024.
\end{description}

\section*{Appendix B. Data and Reproducibility}
\addcontentsline{toc}{section}{Appendix B. Data and Reproducibility}
\noindent Semantic factors are from the EDGAR Research Database (UQO). Equity prices (monthly and
daily) are from Yahoo Finance; the Fama--French five factors, momentum, and the risk-free rate are
from the Kenneth R.\ French data library. All return tests use look-ahead-free July-formation
forward returns. The analysis pipeline (\texttt{analysis/01--09}) downloads prices, builds the
merged panel, runs the sorts/regressions/event study, and emits all figures and tables; results are
serialized to \texttt{results.json}, \texttt{results2.json}, and \texttt{results\_event.json}.

\end{document}