% ============================================================================
\section{Data and Variable Construction}
\label{sec:data}
% ============================================================================

\subsection{The semantic factor panel}
\label{subsec:semantic}
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 proceeds in four
steps. It (i)~extracts the canonical sections of the 10-K---Risk Factors, Management's Discussion and
Analysis (MD\&A), legal proceedings, the financial statements, and so on---so that the location as
well as the content of each passage is known; (ii)~splits each section into overlapping text chunks of
a few sentences, the overlap ensuring that a thought spanning a chunk boundary is not lost;
(iii)~embeds every chunk into a 384-dimensional vector with the \texttt{all-MiniLM-L6-v2} sentence
transformer, a compact and widely validated model that maps semantically similar passages to nearby
points in the embedding space; and (iv)~scores each chunk by its cosine similarity to a set of
\emph{anchor} passages that define eight economic concepts. The anchors are short, deliberately
written paragraphs that articulate each concept in natural language. The AI-adoption anchor, for
instance, 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, on a scale from $-1$ to $1$, regardless of the exact
words used---so that a sentence on ``deploying neural networks to forecast demand'' scores highly on
AI adoption even though it shares no keyword with the anchor. This is precisely the paraphrase- and
context-sensitivity that keyword counts lack.

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 7{,}869 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:robustness}. 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}
\label{subsec:price}
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 \citep{famafrench2015}, 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}
\label{subsec:timing}
The single most important design choice in a disclosure--return study is timing, because a signal
that inadvertently peeks at information unavailable in real time can manufacture predictability that
no investor could have traded on. 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 end of June of $Y{+}1$, and we measure
the 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 builds in a margin that rules out look-ahead bias even
for late filers, while the July-to-June holding window is the standard one in the cross-sectional
literature and so keeps our estimates comparable to it. After requiring a valid forward return with
adequate monthly coverage, we obtain 6{,}852 firm-years spanning formation years 2010--2024, with
return realizations extending through mid-2026.

\subsection{Controls and standardization}
\label{subsec:controls}
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. Appendix~\ref{app:variables} gives precise definitions.

\subsection{Summary statistics}
\label{subsec:summary}
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:robustness}. 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{tables/tab_summary}
\input{tables/tab_corr}

\subsection{The secular rise of AI language}
\label{subsec:trend}
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}[htbp]\centering
\includegraphics[width=0.9\textwidth]{fig4_ai_language_trend.png}
\caption{\textbf{The secular rise of AI language in S\&P~500 disclosure.} The figure plots the
cross-sectional mean of the section-weighted cosine similarity of 10-K text to each of the three AI
anchors (adoption, infrastructure, and disruption risk), by fiscal year over 2010--2026. All three
measures trend steadily upward through the 2010s and inflect sharply after the November 2022 launch of
ChatGPT (dashed vertical line), by which point AI language saturates index disclosure. This secular
trend motivates the within-year cross-sectional design used throughout the paper.}
\label{fig:trend}
\end{figure}

\begin{figure}[htbp]\centering
\includegraphics[width=0.9\textwidth]{fig12_sector_box.png}
\caption{\textbf{AI disclosure across GICS sectors.} The figure shows box plots of the firm-level
AI-adoption score within each GICS sector (boxes span the interquartile range, the red line marks the
sector median, and outliers are suppressed for legibility). AI disclosure is concentrated in
technology-intensive sectors---Information Technology, Communication Services, and Industrials---a
concentration that motivates the sector-channel analysis and the sector-neutral robustness checks
reported later.}
\label{fig:sectorbox}
\end{figure}

\subsection{Which firms disclose the most AI? Descriptive evidence}
\label{subsec:topfirms}
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 variation 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{tables/tab_topfirms}