TeX 69.3%
Python 30.6%
1\documentclass[11pt]{article}2\usepackage[margin=1in]{geometry}3\usepackage{graphicx,booktabs,amsmath,amssymb,siunitx,caption,float,multirow}4\usepackage[round,authoryear]{natbib}5\usepackage[hidelinks]{hyperref}6\usepackage{xcolor,sectsty,setspace,enumitem}7\usepackage{titling}8\definecolor{navy}{HTML}{1b3a6b}9\sectionfont{\color{navy}\large}\subsectionfont{\color{navy}\normalsize}10\captionsetup{font=small,labelfont={bf,color=navy}}11\graphicspath{{../figures/}}12\setlength{\parskip}{0.4em}13\sisetup{group-minimum-digits=4,group-separator={,}}14\newcommand{\tstat}[1]{{\footnotesize(\textit{t}=#1)}}15\doublespacing16\hypersetup{linkcolor=navy,citecolor=navy}1718\title{\color{navy}\bfseries Does the Language of Artificial Intelligence Move Markets?\\[3mm]19\large A Semantic Analysis of S\&P~500 SEC Filings and the Cross-Section of Stock Returns, 2010--2026}20\author{%21\begin{tabular}{c}22\textbf{Manel Kammoun}\\23\texttt{manel.kammoun@uqo.ca}\\[0.5em]24\textbf{Djerrycharli Tandjambianda}\\25\texttt{djerrycharli.tandjambianda@uqo.ca}\\[0.5em]26\textbf{Simon-Pierre Boucher}\\27\texttt{simon-pierre.boucher@uqo.ca}\\[0.8em]28{\large\textbf{Universit\'e du Qu\'ebec en Outaouais (UQO)}}\\29\end{tabular}%30}31\date{June 2026}3233\begin{document}34\maketitle35\thispagestyle{empty}3637\begin{abstract}38\noindent\normalsize39We construct a \emph{semantic} measure of artificial-intelligence (AI) engagement for every40S\&P~500 firm by embedding the text of its annual SEC filings (Form~10-K) with a sentence41transformer and scoring each passage by cosine similarity to professionally curated AI42``anchor'' concepts---internal AI adoption, AI infrastructure exposure, and AI product-disruption43risk. Linking these scores to monthly equity prices from Yahoo Finance and to the Fama--French44factor data, we assemble a panel of \num{7869} firm-years (500 firms, fiscal years 2010--2026)45and ask whether the intensity of AI language predicts the cross-section of subsequent returns.46Our central result is \textbf{conditional and economically large}. Over the full sample the47relation between AI-disclosure intensity and future returns is positive but modest and only48marginally significant; a risk-adjusted long--short portfolio earns an insignificant49$\sim$2\%/year. After the public launch of ChatGPT (30 November 2022), however, the relation50\emph{more than sextuples}: the spread between the top and bottom quintile of AI disclosure rises51from $+2.4$ percentage points per year before 2022 to $+14.6$ points per year afterward. The52break survives section-weighting choices, alternative signal definitions and horizons, dimension53reduction, and---crucially---the exclusion of the ``Magnificent~7'' mega-caps. A battery of54identification tests sharpens the interpretation: (i) a \emph{placebo} across five non-AI55semantic factors shows the post-2022 jump is specific to AI-related language (ESG and climate56disclosure, by contrast, earn \emph{negative} post-2022 spreads); (ii) a market-model57\emph{event study} finds that firms with higher pre-event AI disclosure earned $+10.1\%$ higher58cumulative abnormal returns over the year after ChatGPT \tstat{1.97}; and (iii) the effect is59absorbed by firm fixed effects and roughly halved by sector fixed effects, indicating a60\emph{cross-sectional, regime-dependent re-rating}---the market repriced \emph{which firms and61sectors} are AI-exposed once AI became salient, rather than rewarding year-to-year changes in any62single firm's wording. The findings provide direct, text-based evidence for an attention- and63narrative-driven repricing of AI exposure, and illustrate the value of embedding-based disclosure64measurement for asset pricing.6566\vspace{2mm}67\noindent\textbf{Keywords:} textual analysis, sentence embeddings, artificial intelligence,68disclosure, cross-section of returns, investor attention, event study, generative AI.\\69\noindent\textbf{JEL:} G12, G14, G17, M41, O33.70\end{abstract}7172\clearpage73\setcounter{page}{1}74\tableofcontents75\clearpage7677\section{Introduction}78The arrival of generative artificial intelligence has been the defining corporate-narrative shift79of the 2020s. Between the public launch of ChatGPT in late 2022 and 2025, the market80capitalization of AI-linked firms swelled by trillions of dollars, ``AI'' became the most81frequently invoked theme on earnings calls, and the language of corporate disclosure filled with82references to large language models, generative tools, intelligent automation, and the compute83infrastructure that powers them. This paper asks a deceptively simple question: did the84\emph{language} of AI in firms' mandatory disclosures carry information about their subsequent85stock returns---and if so, when and through what channel?8687Answering this question well requires three ingredients that have only recently become jointly88feasible: a large, machine-readable corpus of corporate disclosure; a measurement technology that89captures the \emph{meaning} of text rather than the mere frequency of keywords; and a clean90mapping from disclosure to tradeable returns. We bring all three to bear. Our laboratory is the91universe of S\&P~500 constituents over fiscal years 2010--2026. Our raw material is the full text92of their annual reports (Form~10-K), parsed section by section. Our measurement engine is the93EDGAR Research Database, which embeds every text chunk with the \texttt{all-MiniLM-L6-v2} sentence94transformer and scores it by cosine similarity to a set of carefully written ``anchor'' passages95that define eight economic concepts. Three of these anchors isolate distinct facets of AI96engagement: \emph{internal AI adoption} (using AI to run the business), \emph{AI infrastructure97exposure} (capital tied to AI compute), and \emph{AI product-disruption risk} (the danger that AI98commoditizes the firm's offerings). The result is a continuous, comparable, firm-year measure of99how intensively---and in what sense---each firm engages with AI in its own words.100101We pair these semantic scores with monthly adjusted prices for all 500 tickers from Yahoo Finance102and with the Fama--French five factors plus momentum. To avoid look-ahead bias we adopt the103standard portfolio-formation convention: because a 10-K covering fiscal year $Y$ is public by104early in year $Y{+}1$, we form signals at the end of June of $Y{+}1$ and measure the forward105twelve-month return from July of $Y{+}1$ to June of $Y{+}2$. The signal is therefore always known106before the holding period begins. This yields \num{6852} firm-years with a strictly tradeable107forward return, spanning formation years 2010 through 2024.108109Our headline finding is that the pricing of AI language is \emph{not} a stable factor premium but110a \textbf{regime-dependent re-rating that switched on with the generative-AI boom}. For most of111the 2010s, sorting S\&P~500 firms on the intensity of their AI disclosure produced only a faint112return pattern: the equal-weighted spread between the highest and lowest quintiles averaged about113$+2.4$ percentage points per year, and a risk-adjusted long--short portfolio earned an114economically small and statistically insignificant alpha. After the launch of ChatGPT in November1152022, the same sort produced a spread of $+14.6$ percentage points per year. The cross-sectional116re-rating is visible in every post-2022 formation cohort, in a cumulative long--short wealth117curve that inflects sharply upward at the end of 2022, and in a year-by-year Fama--MacBeth slope118that jumps from near zero to large and positive.119120A natural worry is that this pattern simply reflects the well-known rally of a handful of121technology mega-caps. We address this concern head-on and reject it. The post-2022 spread is122$+14.1$ percentage points (\textit{t}=1.9) when the ``Magnificent~7'' are excluded and $+14.5$123points when the ten largest firms by market capitalization are dropped---the effect is124broad-based across the index, not an artifact of Nvidia, Microsoft, or their peers. It is robust125to using the simple mean, the section-weighted mean, or the top-5\% tail of similarity scores; to126six-, twelve-, and twenty-four-month horizons; and to replacing the three AI signals with their127first principal component (which alone explains 87\% of their common variation).128129We then deploy three complementary tests to pin down what the result is---and is not. First, a130\emph{placebo} exercise re-runs the entire design on the five non-AI semantic factors in the131database (ESG, climate, cybersecurity, automation/labor, and macro risk). If the post-2022 jump132merely reflected a generic ``more disclosure earns more'' phenomenon, or a mechanical artifact of133our pipeline, every factor should display it. They do not. The post-2022 spread is large and134positive for the AI factors, essentially zero for macro-risk language, and \emph{negative} for135ESG and climate disclosure---high-ESG firms \emph{underperformed} after 2022. The post-ChatGPT136re-rating is specific to AI-related content. Second, we run a market-model \emph{event study}137around the 30 November 2022 ChatGPT launch. Firms in the top tercile of pre-event AI disclosure138(measured from their FY2021 10-K, the last report filed before the event) earned cumulative139abnormal returns $+10.1$ percentage points higher than the bottom tercile over the following year140\tstat{1.97} and $+14.6$ points higher over eighteen months. Because abnormal returns are141purged of market-beta exposure, this is a clean cross-sectional response to an unanticipated142narrative shock. Third, we use two-way fixed effects and difference-in-differences to localize the143mechanism. The premium is absorbed by firm fixed effects and roughly halved by sector fixed144effects, telling us that the durable signal is the persistent \emph{level} of a firm's (and its145sector's) AI exposure relative to peers, not the year-to-year wiggle in its wording.146147Taken together, the evidence supports an \textbf{attention- and narrative-driven repricing}148interpretation. The textual evidence of firms' AI engagement accumulated quietly for a decade and149was barely priced. A salient catalyst---the consumer breakthrough of generative AI---then drew150investor attention to that pre-existing information, and capital flowed toward the firms and151sectors whose disclosures revealed the greatest AI exposure. This is the asset-pricing signature152of limited attention and narrative economics, observed not through a sentiment proxy but directly153in the semantic content of mandatory disclosure.154155\paragraph{Contribution.} We make four contributions. \emph{First}, methodologically, we156demonstrate the value of embedding-based semantic measurement of disclosure for asset pricing: a157cosine-similarity-to-anchor design captures economically meaningful, multi-faceted exposure158(adoption vs.\ infrastructure vs.\ risk) that keyword counts cannot, and it does so at scale across159the entire S\&P~500 and a decade and a half of filings. \emph{Second}, substantively, we provide160the first text-based, look-ahead-free evidence that the cross-sectional pricing of AI exposure is161a post-2022 regime phenomenon rather than a long-standing premium. \emph{Third}, we discipline the162finding with a placebo across non-AI factors and an event study, distinguishing a genuine163AI-specific repricing from a mechanical or market-wide artifact. \emph{Fourth}, we document the164\emph{anatomy} of the effect---its concentration in smaller firms, its operation substantially165through a sector channel, and its disappearance under firm fixed effects---which sharpens the166economic interpretation and warns against naive trading strategies.167168The remainder of the paper proceeds as follows. Section~\ref{sec:lit} situates the study in five169literatures. Section~\ref{sec:hyp} states our hypotheses. Section~\ref{sec:data} describes the170data and the construction of the semantic measures. Section~\ref{sec:method} lays out the171empirical methodology. Section~\ref{sec:results} presents the main results, and172Section~\ref{sec:robust} the robustness and identification tests. Section~\ref{sec:disc} discusses173mechanisms, Section~\ref{sec:limits} the limitations, and Section~\ref{sec:concl} concludes.174175\section{Related Literature}\label{sec:lit}176Our study sits at the intersection of five strands of research: textual analysis in finance, the177pricing of corporate disclosure, technology and innovation in asset pricing, investor attention178and narrative economics, and the nascent literature on the economics of artificial intelligence.179180\subsection{Textual analysis and natural language processing in finance}181A now-large literature extracts economic signals from financial text. The foundational work of182\citet{tetlock2007} showed that the tone of a Wall Street Journal column forecasts aggregate183returns, and \citet{tetlock2008} extended this to firm-level news. \citet{lm2011} demonstrated184that finance-specific sentiment dictionaries materially outperform general-purpose word lists in18510-K analysis, launching a wave of dictionary-based studies and the widely used Loughran--McDonald186word lists. \citet{lm2016} survey this field. A second generation moved from counting words to187modeling them: \citet{gentzkow2019} review text-as-data methods in economics, while188\citet{kelly2019textregression} and \citet{bybee2024narrative} use machine-learning189representations of news to construct interpretable factors and a ``state of the economy'' from190text. Our paper belongs to the most recent, embedding-based generation. Rather than counting AI191keywords---which conflate ``we use no AI'' with ``we are an AI leader'' and miss paraphrase---we192measure the \emph{semantic similarity} of disclosure to anchor concepts using a sentence193transformer \citep{reimers2019sbert}. This follows a growing practice of using contextual194embeddings to quantify disclosure content \citep{cong2022textfactor}, and is closest in spirit to195similarity-to-concept designs that score documents against reference passages.196197\subsection{The pricing of corporate disclosure}198A parallel literature asks whether the \emph{content} and \emph{changes} of mandatory disclosure199predict returns. \citet{cohen2020lazy} show that firms that change the language of their 10-K200relative to the prior year (``lazy prices'' when they do not) experience predictable future201returns and operating outcomes. \citet{hoberg2010product} and \citet{hobergphillips2016} use 10-K202product descriptions to build text-based industry classifications and measures of product-market203fluidity that carry asset-pricing content. \citet{lopezlira2023risk} extract risk exposures from20410-K risk-factor sections. The general lesson is that the qualitative portions of filings contain205value-relevant information that markets process incompletely and with delay. We contribute a206focused, concept-level measure---AI engagement---and show that its pricing is strongly207state-dependent, a dimension the disclosure-pricing literature has largely left implicit.208209\subsection{Technology, innovation, and the cross-section of returns}210Asset pricing has long sought to connect technological change to risk and return.211\citet{hirshleifer2013innovative} show that innovative efficiency predicts returns;212\citet{kogan2017technological} build market-based measures of innovation from patents and link213them to growth and reallocation. \citet{koganpapanikolaou2014} show that investment-specific214technology shocks shape the cross-section. A more recent literature studies AI exposure directly:215\citet{babina2024aiinvestment} measure firm-level AI investment from job-posting and resume data216and document faster growth among AI-adopting firms; \citet{eisfeldt2023generative} construct a217firm-level measure of \emph{generative}-AI exposure from the task content of occupations and show218that a generative-AI factor was priced precisely in the weeks around ChatGPT's release.219\citet{acemoglu2020wrong} and \citet{webb2020} study the labor-market exposure of firms to AI and220automation. Our contribution relative to this work is to measure AI exposure from the firm's221\emph{own disclosure}---a self-reported, narrative signal available for the entire index at annual222frequency---and to characterize the full time profile of its pricing across a decade and a half,223not only the event window. Where \citet{eisfeldt2023generative} use occupation-task exposure, we224use the language firms choose, providing a complementary and highly granular measure.225226\subsection{Investor attention, salience, and narrative economics}227Why might value-relevant disclosure be priced only after a catalyst? The behavioral literature228offers a clear mechanism: limited investor attention. \citet{barberodean2008} show that retail229investors are drawn to attention-grabbing stocks; \citet{da2011} construct a direct attention230proxy from Google search and link it to prices; \citet{dellavigna2009} document under-reaction to231information released when attention is low; and \citet{hirshleifer2009driven} show that earnings232news is underweighted when investors are distracted. \citet{shiller2017narrative} argues that233contagious narratives drive major asset-price movements, and \citet{bakerwurgler2006} show that234sentiment differentially affects hard-to-value, speculative stocks. Our results provide a sharp235illustration: the information about firms' AI engagement existed in filings for years but was236repriced abruptly when a salient consumer event---ChatGPT---focused investor attention on AI. The237asymmetry we document (AI rewarded, ESG penalized after 2022) is itself consistent with a shift in238the prevailing market narrative.239240\subsection{ESG and climate disclosure as a comparison case}241Finally, our placebo design connects to the literature on the pricing of ESG and climate242disclosure. \citet{pastor2021sustainable} model how sustainability preferences depress the243expected returns of green assets in equilibrium, and \citet{pastor2022dissecting} show that green244stocks' strong \emph{realized} returns through 2021 reflected unexpected increases in climate245concern rather than high expected returns. \citet{bolton2021risk} document a carbon premium.246Consistent with the post-2021 reversal in this literature, we find that high-ESG and high-climate247disclosure earned \emph{negative} spreads after 2022, exactly when AI disclosure began to earn248large positive ones---underscoring that the post-2022 re-rating reflects a thematic rotation249toward AI rather than a generic disclosure effect.250251\section{Hypotheses}\label{sec:hyp}252Our analysis is organized around four hypotheses.253254\begin{description}[leftmargin=2.4em,style=nextline]255\item[H1 (Unconditional pricing).] If AI exposure carries a stable risk premium or persistent256mispricing, then sorting on AI-disclosure intensity yields a positive and significant257risk-adjusted return spread over the full 2010--2026 sample.258\item[H2 (Regime dependence).] If instead AI exposure is repriced only once AI becomes salient,259the return spread is small before and large after the late-2022 generative-AI catalyst, with a260detectable structural break.261\item[H3 (AI-specificity).] If the post-2022 spread reflects a genuine AI repricing rather than a262mechanical or market-wide artifact, it appears for AI-related semantic factors but not for263unrelated factors (e.g.\ ESG, climate, macro risk).264\item[H4 (Cross-sectional, not within-firm).] If the market repriced \emph{which} firms are265AI-exposed rather than rewarding incremental wording, the effect is concentrated in the266cross-section and attenuates under firm and sector fixed effects.267\end{description}268269\noindent The data, as we show, decisively favor H2--H4 over H1.270271\section{Data and Variable Construction}\label{sec:data}272\subsection{The semantic factor panel}273The backbone of our study is the EDGAR Research Database, an infrastructure that downloads,274parses, and semantically scores SEC filings for S\&P~500 firms. For each filing the pipeline275(i)~extracts the canonical sections (Risk Factors, MD\&A, legal proceedings, financial statements,276and so on); (ii)~splits each section into overlapping text chunks; (iii)~embeds every chunk into a277384-dimensional vector with the \texttt{all-MiniLM-L6-v2} sentence transformer; and (iv)~scores278each chunk by cosine similarity to a set of \emph{anchor} passages defining eight economic279concepts. The anchors are short, deliberately written paragraphs that articulate the concept in280natural language; for example, the AI-adoption anchor describes a firm ``internally using281artificial intelligence, machine learning, generative AI, large language models, automation,282predictive analytics \dots to improve productivity, operations, efficiency, decision-making.''283Cosine similarity to such an anchor measures how closely a passage's \emph{meaning} matches the284concept, regardless of the exact words used---capturing paraphrase and context in a way keyword285counts cannot.286287Chunk-level scores are aggregated to the filing level using section weights that emphasize the288parts of the document where the concept is most likely to be discussed substantively (e.g.\ MD\&A289and Risk Factors), and then to the firm-fiscal-year level. For each concept the database reports290the mean, maximum, standard deviation, top-5\% tail, and section-weighted mean (\texttt{wmean}) of291chunk similarities, along with first and second differences, percentage changes, and three-year292rolling statistics. The eight concepts comprise three AI facets---\emph{AI adoption}, \emph{AI293infrastructure}, \emph{AI disruption risk}---and five others used as placebos: \emph{ESG294commitment}, \emph{climate risk}, \emph{cybersecurity}, \emph{automation/labor}, and \emph{macro295risk}.296297The resulting panel covers all 500 current S\&P~500 members across fiscal years 2010--2026, a298total of \num{7869} firm-years. We take the section-weighted mean (\texttt{wmean}) as our headline299signal because it up-weights the forward-looking, strategy-rich sections; we verify robustness to300the simple mean and the top-5\% tail in Section~\ref{sec:robust}. Our three primary signals are301therefore \texttt{ai\_adoption\_wmean} (\textbf{AI\,Adopt}), \texttt{ai\_infrastructure\_wmean}302(\textbf{AI\,Infra}), and \texttt{ai\_disruption\_risk\_wmean} (\textbf{AI\,Risk}); we also form a303standardized opportunity composite, $\text{AI\,Compos}=\tfrac12[z(\text{AI\,Adopt})+z(\text{AI\,Infra})]$.304305\subsection{Price and factor data}306We download monthly adjusted-close prices for all 500 tickers from Yahoo Finance over 2009--2026307(handling share-class tickers such as BRK.B and BF.B), the SPDR S\&P~500 ETF (SPY) as the market308benchmark, and the Fama--French five factors, the momentum factor, and the risk-free rate from the309Kenneth R.\ French data library. Monthly simple returns are computed from adjusted prices, which310incorporate dividends and splits. A market-capitalization proxy is formed as shares outstanding311(from Yahoo Finance) times the historical price; we use it for value-weighting and as a size312control, acknowledging the shares-held-constant approximation.313314\subsection{Timing and the construction of tradeable returns}315The single most important design choice in a disclosure--return study is timing. A 10-K covering316fiscal year $Y$ becomes public only when filed, typically in February or March of year $Y{+}1$. To317guarantee that our signal is in the public information set before any return is measured, we follow318the Fama--French portfolio convention: signals from fiscal year $Y$ are used to form portfolios at319the \textbf{end of June of $Y{+}1$}, and we measure the \textbf{forward twelve-month return from320July of $Y{+}1$ through June of $Y{+}2$}. The roughly three-month gap between the typical filing321date and June~30 ensures no look-ahead bias even for late filers. After requiring a valid forward322return with adequate monthly coverage, we obtain \num{6852} firm-years spanning formation years3232010--2024, with return realizations extending through mid-2026.324325\subsection{Controls and standardization}326For each formation year we compute cross-sectional $z$-scores of every signal, so that all327coefficients are expressed per cross-sectional standard deviation and the powerful secular rise in328AI language (documented below) is differenced out---in any given year we compare firms to their329contemporaneous peers. We control for the three workhorse return determinants: \emph{size}330(log market-cap proxy), \emph{momentum} (the firm's own return over the twelve months ending at331formation, skipping the last month), and \emph{idiosyncratic volatility} (the annualized standard332deviation of monthly returns over the trailing year). In regressions, returns and continuous333controls are winsorized at the 1st and 99th percentiles to limit the influence of outliers; in334portfolio sorts we use raw returns.335336\subsection{Summary statistics}337Table~\ref{tab:summ} reports summary statistics for the estimation sample. The average forward338twelve-month return is 18.6\% (reflecting the strong 2011--2026 bull market for U.S.\ large caps),339with a wide cross-sectional dispersion (standard deviation 41.5\%). The AI signals are small in340level---cosine similarities of disclosure to narrow anchors---but exhibit meaningful341cross-sectional variation. Table~\ref{tab:corr} reports correlations. Two features stand out.342First, the three AI signals are highly collinear (pairwise $\rho$ between 0.75 and 0.86): firms343that write more about adopting AI also write more about AI infrastructure and AI risk. They are344best understood as three lenses on a single latent construct---\emph{AI salience}---a point we345formalize with principal-components analysis in Section~\ref{sec:robust}. Second, the raw346(unconditional) correlation between each AI signal and the forward return is positive but small347($\approx 0.04$), foreshadowing that the unconditional relation is weak and that the action lies in348its time variation.349350\input{tab_summary}351\input{tab_corr}352353\subsection{The secular rise of AI language}354Before turning to returns, Figure~\ref{fig:trend} documents the raw phenomenon that motivates our355within-year design. The cross-sectional mean similarity to all three AI concepts trended steadily356upward through the 2010s and inflected sharply after 2022; by 2025--2026 AI language saturates357S\&P~500 disclosure. This is precisely why our tests rely on within-year cross-sectional ranking:358the interesting question is not whether the index as a whole talks more about AI over time (it359emphatically does), but whether, \emph{in a given year}, firms that talk about AI more than their360peers subsequently earn different returns. Figure~\ref{fig:sectorbox} shows that AI disclosure is361also unevenly distributed across sectors, concentrated in Information Technology, Communication362Services, and Industrials---a fact that will matter for interpreting the sector channel.363364\begin{figure}[H]\centering365\includegraphics[width=0.9\textwidth]{fig4_ai_language_trend.png}366\caption{Cross-sectional mean of section-weighted cosine similarity to each AI anchor, by fiscal367year. All three measures trend upward and inflect after the 2022 launch of ChatGPT (dashed368line).}\label{fig:trend}369\end{figure}370371\begin{figure}[H]\centering372\includegraphics[width=0.9\textwidth]{fig12_sector_box.png}373\caption{Distribution of the AI-adoption score by GICS sector (box plots, outliers suppressed; red374line = median). AI disclosure is concentrated in technology-intensive sectors.}\label{fig:sectorbox}375\end{figure}376377\subsection{Which firms disclose the most AI? Descriptive evidence}378Table~\ref{tab:topfirms} lists the ten most AI-intensive S\&P~500 firms by average AI-adoption379score over the full sample. The list is instructive in two ways. First, it has strong face380validity: it features enterprise-technology and software firms (Hewlett Packard Enterprise, Intel,381IBM, Tyler Technologies, Autodesk), an algorithm-centric platform (Uber), and data-intensive382financials (BNY Mellon, Berkshire Hathaway), exactly the kinds of firms one would expect to write383substantively about deploying AI. Second, the list is \emph{not} confined to the technology sector:384industrials (Cummins) and energy (EOG Resources) appear, confirming that intensive AI disclosure385cuts across the GICS taxonomy and that our continuous, text-based measure captures within- and386cross-sector variation that a sector dummy would miss.387388A further descriptive fact is central to our interpretation: the AI signal is highly389\emph{persistent}. The year-over-year cross-sectional rank autocorrelation of the AI composite is3900.88, and even at a five-year horizon it remains 0.73. Firms that are AI-intensive disclosers stay391AI-intensive; the measure captures a slow-moving structural characteristic of the firm rather than392transient wording. This persistence has two implications we exploit later: it explains why firm393fixed effects absorb most of the signal (there is little within-firm variation to identify), and it394underpins the difference-in-differences design, in which we fix each firm's exposure at its395pre-event level precisely because that level is stable.396397\input{tab_topfirms}398399\section{Methodology}\label{sec:method}400We bring six complementary research designs to bear, ranging from simple portfolio sorts to a401market-model event study. Using multiple designs with different identifying assumptions guards402against any single method's weaknesses and lets us triangulate the economic interpretation.403404\subsection{Portfolio sorts}405Each formation year we sort firms into quintiles on a given (within-year standardized) AI signal406and compute the equal-weighted (EW) and value-weighted (VW) forward return of each quintile, plus407the high-minus-low (Q5$-$Q1) spread. We report the time-series average of the annual spreads and408test its significance with Newey--West \citep{neweywest1987} standard errors (three lags) to409account for any serial correlation in the overlapping return windows. Sorts are model-free and410transparent, and the EW--VW comparison reveals whether an effect is concentrated in larger or411smaller firms.412413\subsection{Factor-model spanning of a long--short portfolio}414To ask whether the AI spread survives risk adjustment, we build an annually rebalanced,415equal-weighted, self-financing long--short portfolio (top minus bottom quintile), record its416monthly return, and regress it on (i)~the CAPM market factor and (ii)~the Fama--French five417factors plus momentum (FF6). The intercept (alpha) measures the average return unexplained by418known risk factors; we use heteroskedasticity- and autocorrelation-consistent (HAC) standard419errors. A significant alpha would indicate pricing beyond standard exposures.420421\subsection{Two-way fixed-effects panel regressions}422We estimate423\begin{equation}424r_{i,t\to t+12} = \beta\, \text{AI}_{i,t} + \gamma' X_{i,t} + \alpha_i + \delta_t + \varepsilon_{i,t},425\end{equation}426where $r$ is the winsorized forward return, $\text{AI}_{i,t}$ the standardized signal, $X$ the427controls, and $\alpha_i,\delta_t$ firm and year fixed effects. Standard errors are double-clustered428by firm and year \citep{petersen2009,cameron2011}. Firm fixed effects absorb all429time-invariant cross-firm differences, so $\beta$ identifies whether \emph{within-firm} changes in430AI language predict returns; year fixed effects absorb common shocks. Contrasting this estimate431with the cross-sectional ones isolates whether the signal lives across or within firms (H4).432433\subsection{Fama--MacBeth regressions}434We also run annual cross-sectional regressions of the forward return on the AI signal and controls435and average the slopes \`a la \citet{famamacbeth1973}, with Newey--West errors on the time series of436coefficients. This estimator weights each year equally and yields an interpretable437per-standard-deviation premium; plotting the annual slopes reveals the time profile of the premium438directly.439440\subsection{Structural-break and difference-in-differences designs}441To test regime dependence (H2) we split the sample at the 30~November 2022 ChatGPT launch and442compare sub-period spreads, and we estimate pooled interaction models of the form443$r = \beta_1 \text{AI} + \beta_2 (\text{AI}\times \text{Post}) + \gamma' X + \varepsilon$, where444$\text{Post}=\mathbb{1}[\text{formation year}\ge 2022]$ and $\beta_2$ captures the change in the445premium. We complement this with a difference-in-differences design in which treatment is the446firm's \emph{pre-determined} AI exposure---fixed at its 2019--2021 average so it cannot respond to447the event---interacted with Post, within firm and year fixed effects. An event-time version448interacts pre-determined exposure with each year dummy (relative to 2021) to trace the dynamic449response.450451\subsection{Market-model event study}452Finally, we run a classic event study \citep{mackinlay1997} around the ChatGPT launch. For each453stock we estimate a market model ($r_{i,d}=a_i+b_i r^{m}_{d}+e_{i,d}$) over the estimation window454January--October 2022 and compute daily abnormal returns $\text{AR}_{i,d}=r_{i,d}-(\hat a_i+\hat455b_i r^m_d)$ over the event window November 2022--June 2024. Cumulative abnormal returns (CAR) are456compared across terciles of \emph{pre-event} AI disclosure (from the FY2021 10-K), and we regress457firm-level CAR on pre-event exposure with size, beta, and sector controls. Because abnormal458returns net out market-beta exposure, the event study isolates the cross-sectional response to the459AI-narrative shock from the contemporaneous (and initially negative) market drift of 2022--2023.460461\section{Results}\label{sec:results}462\subsection{Portfolio sorts: a weak unconditional relation}463Table~\ref{tab:sorts} reports the quintile sorts over the full sample. The relation between AI464disclosure and future returns is positive and broadly monotone but economically modest. The465equal-weighted Q5$-$Q1 spread is $+4.6$ percentage points per year for AI adoption \tstat{1.83},466$+3.9$ for AI infrastructure \tstat{1.58}, $+5.3$ for AI disruption risk \tstat{2.27}, and $+4.9$467for the composite \tstat{1.88}. Value-weighting attenuates the adoption and composite spreads,468indicating the effect is stronger among smaller index members. The left panel of469Figure~\ref{fig:quint} displays the full-sample pattern: returns rise gently from the lowest to470the highest AI quintile. On its own, this is suggestive but far from decisive evidence---and, as471we now show, it averages over two very different regimes.472473\input{tab_sorts}474475\subsection{Does AI disclosure earn risk-adjusted alpha unconditionally? (H1)}476Table~\ref{tab:alpha} reports CAPM and FF6 alphas of the long--short portfolios. The477AI-opportunity composite earns a raw spread of about $+2.3\%$ per year, of which the FF6 alpha is478$+1.9\%$---but statistically insignificant \tstat{1.24}. The AI-adoption and AI-infrastructure479long--shorts are similarly insignificant after risk adjustment. The one exception is the480AI-disruption-risk spread, whose FF6 alpha of $+2.8\%$ per year is marginally significant481\tstat{2.10}, hinting at a small risk premium for firms that flag AI as a competitive threat. The482factor loadings (not the focus here) show the AI long--short is roughly market-neutral with modest483negative profitability (RMW) and investment (CMA) tilts, consistent with a growth-oriented long484leg. The verdict on H1 is clear: \textbf{unconditionally, AI disclosure is at best weakly485priced.} This is not the end of the story but the beginning of it.486487\input{tab_alpha}488489\subsection{Panel and Fama--MacBeth estimates}490Table~\ref{tab:panel} juxtaposes the two-way fixed-effects panel with the Fama--MacBeth estimator.491The contrast is instructive. In the Fama--MacBeth (cross-sectional) regressions, the AI-adoption492and AI-infrastructure premia are positive---$+1.3$ and $+1.6$ percentage points per493standard-deviation---with \textit{t}-statistics near 1.6--1.7, indicating a weak but positive494cross-sectional premium on average. In the two-way fixed-effects panel, however, every AI495coefficient collapses to essentially zero. Because firm fixed effects remove all permanent496cross-firm differences, this tells us the signal lives \emph{across} firms, not \emph{within}497them: a given firm writing relatively more about AI from one year to the next does not predict its498own subsequent return (anticipating H4). Throughout, size carries the familiar negative premium499($t\approx-8$) and volatility a positive one, consistent with the broader cross-sectional500literature in this large-cap universe.501502\input{tab_panel}503504\subsection{The structural break: AI pricing switched on after ChatGPT (H2)}505The full-sample averages mask a sharp regime change that is the paper's central result. Splitting506the sample at the November 2022 ChatGPT launch, the center and right panels of507Figure~\ref{fig:quint} reveal a striking transformation: the quintile pattern is nearly flat508before 2022 and steeply increasing afterward. The high-minus-low composite spread is $+2.4$509percentage points per year pre-ChatGPT but $+14.6$ points per year post-ChatGPT---a more than510sixfold increase.511512\begin{figure}[H]\centering513\includegraphics[width=\textwidth]{fig1_quintile_returns.png}514\caption{Mean forward twelve-month return by AI-disclosure composite quintile, full sample and by515sub-period. The high-minus-low spread is $+2.4$ percentage points before ChatGPT and $+14.6$ points516after.}\label{fig:quint}517\end{figure}518519Table~\ref{tab:break} quantifies the break two ways. In sub-period sorts, the AI-adoption Q5$-$Q1520spread rises from $+2.2\%$ per year \tstat{2.04} pre-2022 to $+14.3\%$ \tstat{2.15} post-2022; the521composite rises from $+2.5\%$ to $+14.6\%$; AI infrastructure from $+1.4\%$ to $+13.6\%$. In the522pooled interaction regressions, the post-2022 incremental slope is positive for all three signals523and statistically significant for AI infrastructure \tstat{2.63}---the facet most directly tied to524the chip, cloud, data-center, and power firms that led the 2023--2025 rally. Figure~\ref{fig:spread}525shows the annual spreads: small and mixed-sign through 2021, uniformly large and positive for every526post-2022 cohort.527528\input{tab_break}529530\begin{figure}[H]\centering531\includegraphics[width=0.9\textwidth]{fig3_annual_spread.png}532\caption{Annual high-minus-low (Q5$-$Q1) forward-return spread on the AI-disclosure composite, by53310-K fiscal year. Post-ChatGPT cohorts (red) are uniformly and substantially positive.}\label{fig:spread}534\end{figure}535536The same regime change is visible in two further lenses. Figure~\ref{fig:cum} plots the cumulative537value of one dollar invested in the equal-weighted, annually rebalanced, self-financing538high-minus-low portfolio: it drifts sideways-to-up for a decade and then inflects upward after the539dashed ChatGPT line. Figure~\ref{fig:fmyear} plots the year-by-year Fama--MacBeth slope on the AI540composite (with controls): it hovers near zero for the entire 2010s and then turns sharply541positive, peaking in the 2024 cohort. The latter figure also delivers an important nuance: the542size-controlled premium is largest in the most recent cohort, consistent with the AI543\emph{infrastructure} super-cycle of 2024--2025 (the build-out of data centers and accelerated544compute). We return to this concentration when discussing limitations.545546\begin{figure}[H]\centering547\includegraphics[width=0.86\textwidth]{fig2_cumulative_ls.png}548\caption{Cumulative value of \$1 invested in the equal-weighted, annually rebalanced, self-financing549high-minus-low AI-disclosure portfolio (AI opportunity composite and AI disruption-risk).}\label{fig:cum}550\end{figure}551552\begin{figure}[H]\centering553\includegraphics[width=0.9\textwidth]{fig6_fm_byyear.png}554\caption{Year-by-year Fama--MacBeth slope on the AI-disclosure composite (forward return per555cross-sectional standard deviation), controlling for size, momentum, and volatility. Bars are556$\pm1$ standard error.}\label{fig:fmyear}557\end{figure}558559Finally, Figure~\ref{fig:bin} confirms the pattern is not an artifact of a few influential560observations: the binned relation between the AI composite and forward returns is essentially flat561in the pre-ChatGPT sub-period and distinctly upward-sloping afterward, across the entire562distribution of AI disclosure rather than only at the extreme quintiles.563564\begin{figure}[H]\centering565\includegraphics[width=\textwidth]{fig5_binscatter.png}566\caption{Binned scatter of forward twelve-month return against the AI-disclosure composite, by567sub-period. Each point is a portfolio of firms in a bin of the AI signal; the line is an OLS fit.568The conditional slope steepens markedly after ChatGPT.}\label{fig:bin}569\end{figure}570571\subsection{The anatomy of the premium: size and sector}572Where in the cross-section does the post-2022 premium concentrate? Table~\ref{tab:doublesort} and573Figure~\ref{fig:double} report an AI~$\times$~size double sort. Before 2022 the AI gradient is flat574within both size halves. After 2022 it is steep, and dramatically so among smaller firms: small,575high-AI firms earned a mean forward return of 34.6\%, versus 19.5\% for small, low-AI firms---a57615-point gap---while the corresponding gap among large firms is 7 points. The premium is thus577\emph{not} merely a mega-cap phenomenon (a point we reinforce in Section~\ref{sec:robust}); if578anything it is strongest among the smaller, less-followed names, consistent with an579attention-based mechanism in which information diffuses more slowly into smaller stocks.580581\input{tab_doublesort}582583\begin{figure}[H]\centering584\includegraphics[width=\textwidth]{fig9_double_sort.png}585\caption{AI-disclosure $\times$ firm-size double sort: mean forward twelve-month return by AI586tercile within small and large firms, pre- and post-ChatGPT.}\label{fig:double}587\end{figure}588589\subsection{Economic significance}590The statistical break is matched by an economic one. The Sharpe ratio of the equal-weighted591AI-composite long--short portfolio---a direct gauge of risk-adjusted attractiveness---rises from5920.15 in the pre-ChatGPT decade to 1.33 afterward; for the AI-disruption-risk spread it rises from5930.33 to 1.15. A Sharpe ratio above one is exceptional for a single-characteristic, self-financing594equity strategy, and its emergence precisely at the ChatGPT boundary is hard to reconcile with a595constant risk premium. The point estimates in the quintile sorts are equally vivid: a 14-point596annual spread between the top and bottom AI quintiles is roughly three times the unconditional597equity risk premium and an order of magnitude larger than the pre-2022 spread. Even acknowledging598that a portion of this reflects the higher realized volatility of AI-exposed firms (documented in599Section~\ref{sec:robust}) and the broad technology rally, the magnitude of the post-2022 re-rating600is among the largest single-theme cross-sectional effects in the recent asset-pricing record. The601contrast between an essentially flat pre-period and a near-vertical post-period is, to us, the most602compelling single piece of evidence that a genuine regime change occurred.603604\section{Robustness and Identification}\label{sec:robust}605The structural break is economically large; this section asks how much we should believe it. We606subject it to a placebo across non-AI factors, an event study, a difference-in-differences design,607and a battery of construction and universe robustness checks.608609\subsection{Placebo: is the effect specific to AI? (H3)}610If the post-2022 spread were a mechanical artifact of our pipeline---or a generic ``high-disclosure611firms did well'' phenomenon---it should appear for any semantic factor. Table~\ref{tab:placebo} and612Figure~\ref{fig:placebo} re-run the entire sub-period and interaction design on all eight factors.613The result strongly supports AI-specificity. The three AI factors show large positive post-2022614spreads ($+13.6$ to $+14.3$ points). Macro-risk language shows essentially no spread ($-0.5$,615$t=-0.1$). And, strikingly, ESG-commitment and climate-risk disclosure earn \emph{negative}616post-2022 spreads ($-8.8$ and $-5.0$ points, both significant)---high-ESG firms \emph{underperformed}617after 2022, in line with the post-2021 reversal in the sustainable-investing literature618\citep{pastor2022dissecting}. The one non-AI factor that does show a large positive post-2022619spread is cybersecurity ($+13.5$); we read this not as a violation but as confirmation, since620cybersecurity disclosure is tightly bundled with technology exposure and is itself increasingly621AI-related. The cleanest contrast---AI rewarded, ESG and climate penalized---makes a powerful case622that the post-ChatGPT re-rating reflects a thematic rotation \emph{toward AI}, not a generic623disclosure effect.624625\input{tab_placebo}626627\begin{figure}[H]\centering628\includegraphics[width=0.92\textwidth]{fig7_placebo.png}629\caption{Placebo test: post-ChatGPT high-minus-low return spread for each of the eight semantic630factors. AI-related factors (red) cluster at the top; ESG and climate disclosure earn negative631spreads.}\label{fig:placebo}632\end{figure}633634\subsection{Event study around ChatGPT}635The portfolio results use annual formation and forward returns; the event study provides a636higher-frequency, cleanly identified complement. Table~\ref{tab:event} and Figure~\ref{fig:event}637report market-model cumulative abnormal returns by pre-event AI tercile. The ordering is monotone638at every horizon and widens steadily over time. The High-minus-Low CAR difference is $+2.9$ points639at 60 trading days \tstat{1.77}, $+10.1$ points at one year \tstat{1.97}, and $+14.6$ points at640eighteen months \tstat{1.89}. Because these are abnormal returns---purged of each firm's market641beta---the spread cannot be attributed to high-AI firms simply being higher-beta in a recovering642market. A cross-sectional regression of one-year CAR on pre-event AI exposure yields $+3.1$ points643per standard deviation \tstat{1.39}; adding sector fixed effects halves it to $+2.1$ points644\tstat{0.99}, again locating much---but not all---of the effect in a sector channel.645Figure~\ref{fig:eventbin} shows the relation is smooth across the exposure distribution, not driven646by a few outliers.647648\input{tab_event}649650\begin{figure}[H]\centering651\includegraphics[width=\textwidth]{fig8_event_car.png}652\caption{ChatGPT event study. Left: mean cumulative abnormal return (market model) by pre-event AI653tercile. Right: the High$-$Low CAR spread.}\label{fig:event}654\end{figure}655656\begin{figure}[H]\centering657\includegraphics[width=0.66\textwidth]{fig13_event_binscatter.png}658\caption{One-year cumulative abnormal return after ChatGPT versus pre-event AI exposure (binned659scatter with fitted line).}\label{fig:eventbin}660\end{figure}661662\subsection{Difference-in-differences and the cross-sectional nature of the effect (H4)}663Table~\ref{tab:did} reports difference-in-differences estimates in which treatment is the firm's664pre-determined (2019--2021) AI exposure interacted with the post-2022 indicator, within firm and665year fixed effects. The continuous-exposure DiD coefficient is $+2.7$ percentage points per666standard deviation \tstat{1.51}; the binary version is $+1.7$ points \tstat{0.78}. These are667positive and economically meaningful but weaker than the cross-sectional estimates---exactly the668pattern predicted by H4. Once firm fixed effects absorb the permanent component of AI exposure,669only the residual within-firm variation identifies the effect, and that variation is small. The670event-time profile in Figure~\ref{fig:eventtime} reinforces this reading: the return premium on a671firm's \emph{fixed} pre-event AI exposure is flat and near zero through 2021 and jumps after 2022,672peaking in 2024. The market did not reward firms for adding AI sentences to their filings; it673repriced the firms whose pre-existing, persistent exposure made them AI winners once the narrative674turned.675676\input{tab_did}677678\begin{figure}[H]\centering679\includegraphics[width=0.82\textwidth]{fig11_eventtime.png}680\caption{Event-study coefficients: the return premium on pre-determined (2019--2021) AI exposure by681year (95\% confidence intervals), relative to 2021. The premium activates after ChatGPT.}\label{fig:eventtime}682\end{figure}683684\subsection{Construction, horizon, and universe robustness}685Table~\ref{tab:robust} collects the remaining checks. The post-2022 spread is essentially unchanged686whether the signal is the section-weighted mean (base, $+14.6$), the simple mean ($+13.4$), or the687top-5\% tail ($+14.9$). It holds at six-, twelve-, and twenty-four-month horizons (growing with the688horizon, as a persistent premium should). It survives replacing the three AI signals with their689first principal component---which alone explains 86.8\% of their common variance, confirming the690``single latent AI construct'' interpretation---yielding a post-2022 PC1 spread of $+13.4$ points691\tstat{1.97}. Most importantly for the mega-cap concern, the spread is $+14.1$ points \tstat{1.92}692when the Magnificent~7 are excluded and $+14.5$ points \tstat{1.93} when the ten largest firms are693dropped (Figure~\ref{fig:mag7}). The AI re-rating is a broad-based feature of the index, not the694mechanical footprint of a handful of trillion-dollar winners. Finally, the sector-neutral695composite---which compares firms only to others \emph{within} their GICS sector---produces a much696smaller and insignificant post-2022 spread ($+3.8$ points, $t=0.84$). This is not a failure of the697result but a sharpening of it: a large share of the AI premium operates through a698\emph{between-sector} channel, as the market re-rated AI-exposed sectors (technology above all)699relative to others, with a smaller within-sector residual.700701\input{tab_robust}702703\begin{figure}[H]\centering704\includegraphics[width=0.78\textwidth]{fig10_mag7.png}705\caption{Robustness of the AI return spread to excluding mega-cap winners. The post-2022 composite706spread is stable when the Magnificent~7 or the ten largest firms are removed.}\label{fig:mag7}707\end{figure}708709\subsection{An alternative performance outcome: risk}710Performance is not only about mean returns. We also relate AI disclosure to subsequent realized711volatility. High-AI firms exhibit modestly higher forward volatility, and the gap widens after7122022 (a Q5$-$Q1 difference of about 2.9 annualized volatility points, $t=3.1$). The post-2022 AI713premium therefore came bundled with somewhat higher risk---part, though by no means all, of the714elevated returns is compensation for the greater volatility of AI-exposed firms during a period of715rapid technological repricing.716717\section{Discussion: Mechanism and Interpretation}\label{sec:disc}718What economic mechanism best fits this constellation of facts? We find an \emph{attention- and719narrative-driven repricing} the most parsimonious explanation, and we read the evidence as follows.720721\paragraph{The information pre-existed; the attention did not.} Firms disclosed their AI engagement722for years before it was priced. The semantic content of those disclosures was, in cross-section,723only faintly related to returns through 2021. What changed in late 2022 was not the information but724its \emph{salience}: a consumer breakthrough (ChatGPT) made AI the dominant market narrative725\citep{shiller2017narrative}, drawing investor attention \citep{da2011,barberodean2008} to a body726of disclosure that had been hiding in plain sight. The abrupt, near-simultaneous repricing across727the cross-section---visible in the event study within months of the launch---is the signature of an728attention shock rather than the gradual accrual of a risk premium.729730\paragraph{It is a re-rating of levels, not of changes.} The disappearance of the effect under firm731fixed effects, the muted difference-in-differences estimates, and the event-time profile all point732the same way: the market repriced the persistent \emph{level} of a firm's AI exposure, not the733year-to-year change in its wording. This has a clear practical corollary---firms cannot734mechanically earn the premium by adding AI boilerplate to their filings (and our top-5\% and735simple-mean robustness suggests the measure is not easily gamed)---and a clear methodological one:736studies that difference out firm effects will understate the cross-sectional importance of737disclosure content.738739\paragraph{It operates substantially, but not entirely, through sectors.} The sector-neutral and740sector-fixed-effects results show that much of the AI premium reflects the re-rating of AI-exposed741\emph{sectors}---technology, communication services, and the industrial firms supplying the AI742build-out---relative to the rest of the index. This is economically sensible: the generative-AI743boom was, to first order, a sectoral reallocation of expected cash flows and valuations. Yet a744positive within-sector residual remains in the event study, indicating that the firm's own745disclosure carried incremental information even among industry peers.746747\paragraph{The thematic asymmetry is telling.} Perhaps the most revealing single fact is the748placebo: in the very same post-2022 window, high-AI firms were rewarded while high-ESG and749high-climate firms were penalized. Markets did not simply reward verbose or virtuous-sounding750disclosure; they rotated toward one theme (AI) and away from another (sustainability). This is751difficult to reconcile with a purely mechanical or risk-based account and natural under a752narrative-rotation view of investor attention.753754\paragraph{Relation to the AI-asset-pricing literature.} Our findings complement755\citet{eisfeldt2023generative}, who show that an occupation-task measure of generative-AI exposure756was priced around ChatGPT. We reach a convergent conclusion from an entirely different757measurement---firms' own disclosure language---and extend it along the time dimension, showing that758the same exposure was \emph{un}priced for a decade beforehand. The agreement of two independent759measures on the timing and sign of the effect strengthens confidence that a real AI repricing760occurred.761762\subsection{Implications}763Our findings carry distinct implications for three audiences. \emph{For investors and quantitative764strategists}, the results are a caution as much as an opportunity. The AI-disclosure signal765delivered a striking post-2022 Sharpe ratio, but the same signal was nearly inert for the preceding766decade, and the effect is absorbed by firm fixed effects---so a strategy that traded on767year-to-year changes in AI verbosity would have captured little. The tradeable content lies in the768persistent cross-sectional \emph{level} of AI exposure, and even that is partly a sector bet whose769durability is unproven. The honest reading is that disclosure-based AI signals are best viewed as a770lens on a regime-dependent theme rather than as a stable alpha source. \emph{For corporate managers771and disclosure policy}, the asymmetry between AI (rewarded) and ESG/climate (penalized) after 2022772illustrates how powerfully the prevailing market narrative conditions the valuation consequences of773what firms choose to emphasize. Because the premium attaches to persistent, hard-to-fake exposure774rather than to boilerplate, the result does not invite cheap disclosure manipulation; but it does775suggest that the market reads filings for genuine strategic content and reprices it when attention776turns. \emph{For researchers}, the study demonstrates that embedding-based semantic measurement777uncovers pricing dynamics invisible to both keyword counts (which cannot distinguish meaning) and778firm-differenced panels (which discard the persistent cross-sectional variation that turns out to779matter). The broader agenda---measuring multi-faceted, concept-level exposure from disclosure at780scale and studying its state-dependent pricing---appears fertile.781782\subsection{Summary of the evidence across designs}783It is worth stepping back to see how the designs corroborate one another, since each has different784weaknesses. The portfolio sorts establish the raw magnitude and its post-2022 timing but are785model-free and could in principle reflect risk; the factor-model spanning regressions show the786unconditional spread is not explained by standard factors yet is statistically weak over the full787sample; the sub-period and interaction tests localize the effect to the post-ChatGPT regime; the788placebo rules out a mechanical or generic-disclosure explanation by showing the opposite sign for789ESG and climate; the event study, using daily abnormal returns purged of beta, provides790cleanly-timed, cross-sectional confirmation within months of the catalyst; and the fixed-effects791and difference-in-differences estimates pin down that the priced object is persistent firm- and792sector-level exposure, not incremental wording. No single design is decisive, but their793convergence---on the same sign, the same timing, and a consistent economic story---is what gives us794confidence in the conclusion. A spurious result would have to survive all six designs795simultaneously while also producing the precise ESG/climate sign reversal, which strains credulity.796797\section{Limitations}\label{sec:limits}798Several caveats temper our conclusions. \emph{First}, the post-2022 window contains only three799formation cohorts, and the size-controlled premium is concentrated in the most recent (2024)800cohort, which captures the 2024--2025 AI-infrastructure capital-expenditure boom. The regime result801is economically large and consistent across designs (sorts, event study, DiD), but its post-period802rests on a short sample and could in principle reflect a single extraordinary episode rather than a803durable new regime; only future data can settle this. \emph{Second}, our universe is current804S\&P~500 membership, which introduces survivorship and look-ahead-at-the-index-level bias: firms805that exited the index are absent, and current members were selected partly on past success. This806biases the level of returns upward but is less likely to drive the \emph{cross-sectional}807AI--return relation that is our focus. \emph{Third}, the market-capitalization proxy holds shares808outstanding at their latest value, approximating historical size with error. \emph{Fourth}, the809semantic anchors, though carefully written, embed the authors' definitional choices; different810anchor wording would yield somewhat different scores, although the robustness to mean/tail811aggregations and to PCA mitigates this concern. \emph{Fifth}, disclosure intensity is correlated812with sector, and while our within-year, size-, momentum-, and volatility-controlled tests and the813sector-FE event study address this, we cannot fully separate ``AI exposure'' from ``technology814sector'' as the priced characteristic---indeed, our sector-channel results suggest the two are815deeply intertwined. \emph{Finally}, our designs establish a robust association and an816event-study-based timing argument, but not strict causality; the attention-repricing interpretation817is the most consistent with the evidence rather than the only logically possible one.818819\section{Conclusion}\label{sec:concl}820Using a semantic, embedding-based measure of artificial-intelligence language in the SEC filings of821S\&P~500 firms, linked to Yahoo Finance returns and standard factor data, we have shown that the822pricing of AI disclosure is a \emph{conditional, regime-dependent} phenomenon. For most of the8232010s, how intensively a firm wrote about adopting, building, or being threatened by AI bore only a824faint relation to its subsequent returns, and a risk-adjusted long--short strategy earned no825reliable alpha. After the launch of ChatGPT in late 2022, the same disclosure became strongly826priced: the high-minus-low AI-disclosure return spread jumped from roughly $+2$ to roughly $+15$827percentage points per year. The break is robust to signal definition, horizon, dimension reduction,828and the exclusion of mega-cap winners; it is specific to AI-related language (ESG and climate829disclosure, tested identically, were \emph{penalized} in the same window); it is corroborated by a830market-model event study in which high-AI firms earned $+10$ points of additional abnormal return831in the year after ChatGPT; and it operates as a cross-sectional re-rating of persistent firm- and832sector-level exposure rather than a reward for year-to-year wording, fading under firm fixed833effects.834835The economic message is that markets price the \emph{language} of AI---but only once AI commands836their attention. The methodological message is that embedding-based measurement of disclosure837content opens a productive avenue for asset pricing, capturing economically meaningful, multi-faceted838exposures at scale and revealing pricing dynamics that keyword counts and firm-differenced designs839would miss. As generative AI continues to reshape the corporate landscape, the semantic footprint840it leaves in mandatory disclosure offers a rich, and now demonstrably value-relevant, signal.841842\clearpage843\begin{thebibliography}{99}\small844\setlength{\itemsep}{2pt}845\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--2244.846\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, 103745.847\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--1680.848\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--818.849\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--549.850\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--3147.851\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--249.852\bibitem[Cohen et al.(2020)]{cohen2020lazy} Cohen, L., C.\ Malloy, and Q.\ Nguyen (2020). Lazy Prices. \textit{Journal of Finance} 75(3), 1371--1415.853\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.854\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--1499.855\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--749.856\bibitem[Eisfeldt et al.(2023)]{eisfeldt2023generative} Eisfeldt, A.\ L., G.\ Schubert, and M.\ B.\ Zhang (2023). Generative AI and Firm Values. NBER Working Paper 31222.857\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--22.858\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--636.859\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--574.860\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--2325.861\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--654.862\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--3811.863\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--1465.864\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).865\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--718.866\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--712.867\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.868\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--65.869\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--1230.870\bibitem[MacKinlay(1997)]{mackinlay1997} MacKinlay, A.\ C. (1997). Event Studies in Economics and Finance. \textit{Journal of Economic Literature} 35(1), 13--39.871\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--708.872\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--571.873\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--424.874\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--480.875\bibitem[Reimers and Gurevych(2019)]{reimers2019sbert} Reimers, N., and I.\ Gurevych (2019). Sentence-BERT: Sentence Embeddings Using Siamese BERT-Networks. \textit{EMNLP}.876\bibitem[Shiller(2017)]{shiller2017narrative} Shiller, R.\ J. (2017). Narrative Economics. \textit{American Economic Review} 107(4), 967--1004.877\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--1168.878\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--1467.879\bibitem[Webb(2020)]{webb2020} Webb, M. (2020). The Impact of Artificial Intelligence on the Labor Market. Working paper, Stanford University.880\end{thebibliography}881882\clearpage883\appendix884\section*{Appendix A. Variable Definitions}885\addcontentsline{toc}{section}{Appendix A. Variable Definitions}886\begin{description}[leftmargin=1.5em,style=nextline,itemsep=2pt]887\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 $Y$.888\item[AI\,Infra (\texttt{ai\_infrastructure\_wmean}).] Section-weighted mean similarity to the AI-compute-infrastructure anchor.889\item[AI\,Risk (\texttt{ai\_disruption\_risk\_wmean}).] Section-weighted mean similarity to the AI-product-disruption-risk anchor.890\item[AI\,Compos.] $\tfrac12[z(\text{AI\,Adopt})+z(\text{AI\,Infra})]$, with $z$ the within-year cross-sectional standardization.891\item[Forward return.] Buy-and-hold simple return from July of $Y{+}1$ to June of $Y{+}2$ (adjusted close, dividends reinvested).892\item[Market-adjusted return.] Forward return minus the contemporaneous SPY return.893\item[Size.] Natural log of the market-cap proxy (latest shares outstanding $\times$ price) at formation.894\item[Momentum.] Own return over the 12 months ending at formation, skipping the last month.895\item[Volatility.] Annualized standard deviation of monthly returns over the trailing 12 months.896\item[Post.] Indicator equal to one for formation (fiscal) years $\ge 2022$.897\item[CAR.] Cumulative market-model abnormal return; market model estimated Jan--Oct 2022, event window Nov 2022--Jun 2024.898\end{description}899900\section*{Appendix B. Data and Reproducibility}901\addcontentsline{toc}{section}{Appendix B. Data and Reproducibility}902\noindent Semantic factors are from the EDGAR Research Database (UQO). Equity prices (monthly and903daily) are from Yahoo Finance; the Fama--French five factors, momentum, and the risk-free rate are904from the Kenneth R.\ French data library. All return tests use look-ahead-free July-formation905forward returns. The analysis pipeline (\texttt{analysis/01--09}) downloads prices, builds the906merged panel, runs the sorts/regressions/event study, and emits all figures and tables; results are907serialized to \texttt{results.json}, \texttt{results2.json}, and \texttt{results\_event.json}.908909\end{document}910