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

\subsection{The scored filing corpus}
\label{subsec:corpus_data}
The backbone of the study is the EDGAR Research Database, an infrastructure built by the authors
that downloads, parses, and semantically scores SEC filings for the 500 current S\&P~500
constituents. The acquisition layer retrieved every Form~10-K, 10-Q, and 8-K filed by these firms
from January 2010 to June 2026 (\num{134124} filings). The parsing layer splits each document into
its canonical sections where a section template exists (the quarterly report is parsed into its
financial statements, management's discussion and analysis (MD\&A), quantitative disclosures,
controls and procedures, legal proceedings, risk factors, and unregistered sales; the current
report is kept as a single body) and then into overlapping chunks of about 500 words, the overlap
ensuring that a thought spanning a chunk boundary is not lost. The embedding layer maps every
chunk to a 384-dimensional vector with the \texttt{all-MiniLM-L6-v2} sentence transformer
\citep{reimers2019sbert}, a compact and widely validated model that places semantically similar
passages close together.

Table~\ref{tab:corpus} describes the corpus that had actually been parsed and scored when the
panel used in this paper was built. All \num{103926} Form~8-K filings of the 500 firms were
processed, as were \num{3710} of the \num{22542} Form~10-Q filings, belonging to 84 firms, almost all
of them with tickers between SMCI and ZTS (the parser was interrupted during an alphabetical pass;
Appendix~\ref{app:corpus} lists them); none of
the \num{7655} Form~10-K filings had been processed. The first version of this paper described the
corpus as the full text of annual reports, which was incorrect. The scored text is therefore
dominated by current reports: a typical 8-K contributes two chunks and a typical firm-year contains
about thirteen filings and 28 chunks (interquartile range 17--57). Current reports are short and
event-driven---earnings releases, material agreements, executive changes, results of shareholder
votes---so the AI measure captures the extent to which a firm's \emph{announcements} of the year
are couched in AI-related language, rather than the extent of its annual strategic narrative. The
10-Q filings that were processed carry roughly thirty times more chunks each and, because they
contain MD\&A and risk-factor sections, score about 0.07 higher on the AI-adoption anchor on
average than an 8-K. Section~\ref{subsec:corpus} shows what this implies for the firm-year scores and
re-estimates the main results on 8-K filings alone.

\input{tables/tab_corpus}

\subsection{Anchor concepts and semantic scores}
\label{subsec:semantic}
Each of eight economic concepts is defined by three short paragraphs written in the register of a
filing; the full texts are reproduced in Appendix~\ref{app:anchors}. The AI-adoption anchor, for
instance, describes a firm that ``internally uses artificial intelligence, machine learning,
generative AI, large language models, automation, predictive analytics \dots\ to improve
productivity, operations, efficiency, decision-making.'' The three paragraphs of a concept are
embedded, averaged, and normalized, so that each concept $a$ is represented by a unit vector
$\bar{\mathbf a}_a$. The score of chunk $c$ on concept $a$ is the cosine similarity
\begin{equation}
  s_{c,a} \;=\; \frac{\mathbf e_c^{\prime}\,\bar{\mathbf a}_a}{\lVert \mathbf e_c\rVert\,\lVert \bar{\mathbf a}_a\rVert}\in[-1,1],
  \label{eq:cos}
\end{equation}
where $\mathbf e_c$ is the embedding of the chunk. Because similarity is computed in the embedding
space, a passage on ``deploying neural networks to forecast demand'' scores highly on AI adoption
even though it shares no keyword with the anchor; this paraphrase sensitivity is what keyword
counts lack. The eight concepts comprise three AI facets---\emph{AI adoption}, \emph{AI
infrastructure}, and \emph{AI disruption risk}---and five others used as placebos: \emph{ESG
commitment}, \emph{climate risk}, \emph{cybersecurity}, \emph{automation/labour}, and \emph{macro
risk}.

Chunk scores are aggregated to the filing level in three ways. The simple mean averages all chunks
of filing $f$; the top-5\% tail averages the chunks at or above the 95th percentile of the filing's
scores; and the section-weighted mean, our headline measure, weights chunks by the section in which
they appear,
\begin{equation}
  \bar s^{\,w}_{f,a} \;=\; \frac{\sum_{c\in f} w_{\sigma(c),a}\, s_{c,a}}{\sum_{c\in f} w_{\sigma(c),a}},
  \label{eq:wmean}
\end{equation}
where $\sigma(c)$ is the section of chunk $c$ and the weights $w$ emphasize MD\&A and risk-factor
text for the relevant concept (for AI adoption and infrastructure, MD\&A receives weight 2.0 and
risk factors 1.2; for AI disruption risk, risk factors 2.0 and MD\&A 1.5; financial statements
receive 0.5; the undivided body of an 8-K receives 1.0). Since 8-K filings have a single section,
the weighted and simple means coincide for them, and the weights only affect the 10-Q filings of
the 84 firms. Filing-level scores are then averaged, without weights, over all scored filings dated
in calendar year $Y$:
\begin{equation}
  \text{AI}_{i,Y,a} \;=\; \frac{1}{|\mathcal F_{i,Y}|}\sum_{f\in\mathcal F_{i,Y}} \bar s^{\,w}_{f,a},
  \label{eq:annual}
\end{equation}
where $\mathcal F_{i,Y}$ is the set of scored filings of firm $i$ in year $Y$. An important
consequence of equation~\eqref{eq:annual} is that a 10-Q with sixty chunks and an 8-K with two
chunks carry the same weight, so that the presence of a scored 10-Q in a firm-year raises the
annual score mechanically. The panel covers the 500 current constituents over filing years
2010--2026, a total of \num{7869} firm-years. Our three primary signals are
\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, $\text{AI\,Compos}=\tfrac12[z(\text{AI\,Adopt})+z(\text{AI\,Infra})]$,
where $z(\cdot)$ denotes the within-year cross-sectional standardization defined below.

\subsection{Price and factor data}
\label{subsec:price}
We download monthly adjusted-close prices for all 500 tickers from Yahoo Finance over June 2009 to
June 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
\citep{carhart1997}, and the risk-free rate from the Kenneth R.\ French data library (available
through April 2026). Monthly simple returns are computed from adjusted prices, which incorporate
dividends and splits. A market-capitalization proxy is formed as the latest shares outstanding
(from Yahoo Finance, June 2026) times the historical price; we use it for value-weighting and as a
size control, acknowledging that it holds shares constant. Daily prices from September 2021 to
July 2024 are used for the event study.

\subsection{Universe and survivorship}
\label{subsec:universe}
The universe is the S\&P~500 membership as of June 2026, not the historical membership. Firms that
left the index before 2026 are absent, and firms that joined it after 2010---83 of the 500 have
Yahoo Finance prices beginning after July 2009, among them Uber, Palantir, and Airbnb---enter the
panel with their pre-inclusion history. This is a survivorship and look-ahead bias at the level of
the universe \citep{browngoetzmannross1995}: it raises the average level of returns (entrants earn
10.6 percentage points more per year than seasoned firms in our panel) and could bias the
cross-sectional AI--return relation if recent entrants were concentrated in high-AI quintiles. We
show in Section~\ref{subsec:entrants} that entrants are spread evenly across the AI quintiles and
that the main results hold among firms listed since 2009, but the reader should bear in mind that
the panel is not a point-in-time index.

\subsection{Timing and the construction of tradeable returns}
\label{subsec:timing}
The most important design choice in a disclosure--return study is timing, because a signal that
inadvertently uses information unavailable in real time can manufacture predictability that no
investor could have traded on. Every filing contributing to the signal of year $Y$ is dated between
1~January and 31~December of $Y$, and it is public on its filing date. We follow the
portfolio-formation convention of \citet{famafrench2015}: signals of year $Y$ are used to form
portfolios at the end of June of $Y{+}1$, and the forward return is the buy-and-hold return from
July of $Y{+}1$ through June of $Y{+}2$,
\begin{equation}
  r_{i,Y+1} \;=\; \prod_{m=\text{Jul}\,Y+1}^{\text{Jun}\,Y+2}\bigl(1+r_{i,m}\bigr)-1 ,
  \label{eq:fwd}
\end{equation}
computed when at least six monthly returns are available in the window. The gap between the last
filing of year $Y$ and portfolio formation is therefore at least six months and on average about
twelve, which rules out look-ahead bias but also means that the signal is stale relative to the
quarterly cadence of current reports; this works against finding predictability. Two consequences
of the calendar-year convention deserve notice. The signal of the 2022 cohort includes 8-K filings
dated in December 2022, after the ChatGPT launch, so a small part of that signal is itself a
reaction to the event; the difference-in-differences design of Section~\ref{sec:robustness}, which
fixes exposure at its 2019--2021 level, is immune to this. And the holding period of the 2021
cohort (July 2022--June 2023) straddles the launch; we treat this cohort as pre-launch in the
baseline and show in Section~\ref{sec:robustness} that moving it does not change the conclusions.
After requiring a valid forward return we obtain \num{6852} firm-years (497 firms) spanning
formation years 2010--2024, with return realizations extending to June 2026; the forward return of
the 2024 cohort ends with the partial June 2026 observation, prices having been downloaded on
17~June 2026. Filing years 2025 and 2026 contribute signals but no forward return.

\subsection{Controls and standardization}
\label{subsec:controls}
For each formation year we compute cross-sectional $z$-scores of every signal,
\begin{equation}
  z(x_{i,Y}) \;=\; \frac{x_{i,Y}-\bar x_{Y}}{\mathrm{sd}_Y(x)} ,
  \label{eq:z}
\end{equation}
so that all coefficients are expressed per cross-sectional standard deviation and the secular rise
in AI language documented below is removed: in any given year we compare firms to their
contemporaneous peers. We control for three return determinants measured at formation:
\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{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; in portfolio sorts we use raw returns and
report winsorized spreads as a robustness check. 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 market for U.S.\ large caps and the
survivorship of the universe, with a wide cross-sectional dispersion (standard deviation 41.5\% and
a maximum of 969\%). The AI signals are small in level---cosine similarities of short passages to
narrow anchors---but exhibit meaningful cross-sectional variation. Table~\ref{tab:corr} reports
correlations. The three AI signals are highly collinear (pairwise $\rho$ between 0.74 and 0.86):
firms whose filings resemble the adoption anchor also resemble the infrastructure and risk anchors.
They are best understood as three lenses on a single latent construct, AI salience, a point we
formalize with principal components in Section~\ref{sec:robustness}. The raw correlation between
each AI signal and the forward return is positive but small ($\approx 0.04$).

\input{tables/tab_summary}
\input{tables/tab_corr}

\subsection{The secular rise of AI language}
\label{subsec:trend}
Figure~\ref{fig:trend} documents the phenomenon that motivates the within-year design. The
cross-sectional mean similarity to all three AI concepts trended upward through the 2010s and
inflected sharply after 2022; by 2025--2026 AI language is pervasive in the filings of the index.
The interesting question is therefore not whether the index as a whole talks more about AI over
time, but whether, in a given year, firms that talk about AI more than their peers subsequently earn
different returns. Figure~\ref{fig:sectorbox} shows that AI language is unevenly distributed across
sectors, concentrated in Information Technology, Communication Services, and Industrials, a fact
that matters for the sector analysis.

\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 filings.} The figure plots the
cross-sectional mean of the section-weighted cosine similarity of the scored filings (Form~8-K and,
for 84 firms, Form~10-Q) to each of the three AI anchors, by filing year over 2010--2026. All three
measures trend upward through the 2010s and inflect after the November 2022 launch of ChatGPT
(dashed vertical line). The trend motivates the within-year cross-sectional design.}
\label{fig:trend}
\end{figure}

\begin{figure}[htbp]\centering
\includegraphics[width=0.9\textwidth]{fig12_sector_box.png}
\caption{\textbf{AI language across GICS sectors.} Box plots of the firm-year AI-adoption score
within each GICS sector (boxes span the interquartile range, the red line marks the sector median,
outliers are suppressed). AI language is concentrated in Information Technology, Communication
Services, and Industrials, which motivates the sector-neutral and sector-factor analyses of
Section~\ref{sec:robustness}.}
\label{fig:sectorbox}
\end{figure}

\subsection{Which firms score highest?}
\label{subsec:topfirms}
Table~\ref{tab:topfirms} lists the ten firms with the highest average AI-adoption score. The list
contains enterprise-technology and software firms (Hewlett Packard Enterprise, Intel, IBM, Tyler
Technologies, Autodesk), a platform company (Uber), data-intensive financials (BNY Mellon,
Berkshire Hathaway), an industrial (Cummins), and an energy producer (EOG Resources). Four of the
ten belong to the group of 84 firms whose 10-Q filings were scored, which mechanically raises their
scores; the ranking should therefore be read as descriptive. A further fact is central to the
interpretation: the signal is persistent. The average year-over-year cross-sectional rank
correlation of the composite is 0.88, and at a five-year lag it is still 0.73. Firms whose filings
score high on AI stay high; the measure captures a slow-moving characteristic rather than transient
wording. This explains why firm fixed effects absorb most of the signal, and it underpins the
difference-in-differences design that fixes each firm's exposure at its pre-launch level.

\input{tables/tab_topfirms}