% ============================================================================
\section{Introduction}
\label{sec:introduction}
% ============================================================================

The arrival of generative artificial intelligence has been the defining corporate-narrative shift
of the 2020s. In the span of barely three years following the public launch of ChatGPT in late
2022, the market capitalization of AI-linked firms swelled by trillions of dollars, ``AI'' vaulted
past every other theme in frequency 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. A technology that had been the preserve of research labs
and back-office analytics suddenly became the organizing story of public equity markets. This paper
asks a question that is simple to state but surprisingly demanding to answer well: did the
\emph{language} of AI in firms' mandatory disclosures carry information about their subsequent stock
returns---and if so, when did that information begin to be priced, and through what economic channel?

Answering the question convincingly requires three ingredients that have only recently become jointly
feasible. The first is a large, machine-readable corpus of corporate disclosure. The second is a
measurement technology that captures the \emph{meaning} of text rather than the mere frequency of
keywords---one that recognizes that a firm describing how it ``deploys machine-learning models to
optimize its supply chain'' is talking about AI even when the three-letter acronym never appears. The
third is a clean, look-ahead-free 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, each corresponding to a different way the technology bears
on firm value: \emph{internal AI adoption} (using AI to run the business more productively),
\emph{AI infrastructure exposure} (capital and revenue tied to the compute build-out), and \emph{AI
product-disruption risk} (the danger that AI commoditizes or displaces the firm's own 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, available for the entire index and spanning a decade and a
half.

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 filed and 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 roughly three-month cushion between the typical
filing date and portfolio formation guarantees that the signal is in the public information set
before the holding period begins, even for late filers. This yields 6{,}852 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
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---economically modest and statistically fragile---and a risk-adjusted long--short
portfolio earned an alpha that was both small and statistically insignificant. The information was
there in the filings, but the market was not paying for it. After the launch of ChatGPT in November
2022, the very same sort produced a spread of $+14.6$ percentage points per year, a more than sixfold
jump. This is not a subtle tilt buried in a long-run average; it is a sharp, datable change in how
the cross-section prices a fixed characteristic. The re-rating is visible from three independent
angles: it appears in every post-2022 formation cohort taken on its own, in a cumulative long--short
wealth curve that drifts sideways for a decade and then inflects sharply upward at the end of 2022,
and in a year-by-year Fama--MacBeth slope that hovers near zero through 2021 before turning large and
positive.

A natural worry is that this pattern simply reflects the well-known rally of a handful of technology
mega-caps---that we have rediscovered the Magnificent~7 dressed in textual clothing. We confront this
concern head-on and reject it. The post-2022 spread is $+14.1$ percentage points \tstat{1.9} when the
``Magnificent~7'' are excluded and $+14.5$ points when the ten largest firms by market capitalization
are dropped: removing the very names that dominate the headline rally barely moves the estimate, so
the effect is broad-based across the index rather than an artifact of Nvidia, Microsoft, or their
peers. It is equally robust to how the signal is built and measured---to using the simple mean, the
section-weighted mean, or the top-5\% tail of similarity scores; to six-, twelve-, and twenty-four-month
holding horizons; and to replacing the three correlated AI signals with their first principal
component, a single index that alone explains 87\% of their common variation.

Establishing that the spread is large and well-timed is only half the task; we then deploy three
complementary tests to pin down what the result is---and, equally important, what it is not. The
first is a \emph{placebo} exercise that re-runs the entire research design on the five non-AI semantic
factors in the database (ESG, climate, cybersecurity, automation/labor, and macro risk). The logic is
simple: if the post-2022 jump merely reflected a generic ``more disclosure earns more'' phenomenon, or
a mechanical artifact of our embedding-and-sorting pipeline, every factor should display the same break.
They do not. The post-2022 spread is large and positive for the AI factors, essentially zero for
macro-risk language, and---most tellingly---\emph{negative} for ESG and climate disclosure, meaning
high-ESG firms \emph{under}performed after 2022. The post-ChatGPT re-rating is specific to AI-related
content, not a property of disclosure intensity in general. The second test is 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 anyone could have
anticipated 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 each firm's market-beta exposure, this is a clean
cross-sectional response to an unanticipated narrative shock rather than a reflection of high-AI firms
simply being higher-beta in a recovering market. The third test uses two-way fixed effects and
difference-in-differences to localize the mechanism within the cross-section. The premium is fully
absorbed by firm fixed effects and roughly halved by sector fixed effects, which tells us that the
durable, priced 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 four pieces of evidence---the datable break, the AI-specific placebo, the
event-study timing, and the fixed-effects anatomy---converge on a single interpretation: an
attention- and narrative-driven repricing. The textual evidence of firms' AI engagement accumulated
quietly in their filings 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 had already revealed the greatest AI
exposure. What looks like new information was in large part old information newly attended to. This is
the asset-pricing signature of limited attention and narrative economics, observed here not through an
indirect 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
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 is organized as follows. Section~\ref{sec:literature} situates the study
in five literatures and states our hypotheses. Section~\ref{sec:data} describes the data and the
construction of the semantic measures. Section~\ref{sec:methodology} lays out the empirical
methodology. Section~\ref{sec:results} presents the main results, and Section~\ref{sec:robustness}
the robustness and identification tests. Section~\ref{sec:discussion} discusses mechanisms and
limitations, and Section~\ref{sec:conclusion} concludes.