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

The arrival of generative artificial intelligence has been the defining corporate-narrative shift
of the 2020s. In the three years following the public launch of ChatGPT in late 2022, the market
capitalization of AI-linked firms rose by trillions of dollars, ``AI'' became the most frequently
invoked theme on earnings calls, and the language of corporate disclosure filled with references to
large language models, generative tools, intelligent automation, and the compute infrastructure
that powers them. This paper asks a question that is simple to state but 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 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 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 acronym never appears. The third is
a look-ahead-free mapping from disclosure to tradeable returns. Our laboratory is the set of 500
current S\&P~500 constituents over 2010--2026. Our raw material is the text of their SEC filings as
parsed by the EDGAR Research Database, which embeds every text chunk with the
\texttt{all-MiniLM-L6-v2} sentence transformer \citep{reimers2019sbert} and scores it by cosine
similarity to a set of written ``anchor'' passages that define eight economic concepts. Three of
these anchors isolate distinct facets of AI engagement: \emph{internal AI adoption} (using AI to run
the business), \emph{AI infrastructure exposure} (capital and revenue tied to the compute
build-out), and \emph{AI product-disruption risk} (the danger that AI commoditizes 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.

Two features of the corpus must be stated at the outset, because they were misdescribed in the
first version of this paper and they shape the interpretation of everything that follows. First,
the filings that had been parsed and scored when the panel was built are the Form~8-K current
reports of all 500 firms (\num{103926} filings) and the Form~10-Q quarterly reports of 84 firms
(\num{3710} filings); no Form~10-K annual report had been processed. The AI measure is therefore a
measure of the language of current reports---material agreements, earnings releases, executive
changes, shareholder votes---supplemented, for a sixth of the firm-years, by quarterly reports.
Second, the ``year'' of the panel is the calendar year in which the filings were dated, not the
fiscal year they cover. We treat both features as measurement facts to be documented and tested
rather than as reasons to discard the data: Section~\ref{sec:data} describes the corpus in detail,
and Section~\ref{sec:robustness} shows that the main results are unchanged when the signal is
rebuilt from 8-K filings alone.

We pair the semantic scores with monthly adjusted prices for the 500 tickers from Yahoo Finance and
with the Fama--French five factors plus momentum. To avoid look-ahead bias we follow the
portfolio-formation convention of the cross-sectional literature: the signal of filing year $Y$ is
used to form portfolios at the end of June of $Y{+}1$ and we measure the forward twelve-month
return from July of $Y{+}1$ to June of $Y{+}2$. Because every filing of year $Y$ is dated on or
before 31~December of $Y$, the signal is in the public information set at least six months before
the holding period begins. This yields \num{6852} firm-years with a tradeable forward return,
spanning formation years 2010 through 2024.

Our first finding is that AI language is at best weakly priced unconditionally. Sorting firms into
quintiles of AI-disclosure intensity each year produces an equal-weighted top-minus-bottom spread
of about 5 percentage points per year with a $t$-statistic below two, and the six-factor alpha of
the corresponding long--short portfolio is 1.9\% per year and insignificant. Our second finding is
that the relation strengthens after the launch of ChatGPT. Over the three post-launch formation
cohorts (2022--2024) the composite spread averages $+14.6$ points per year against $+2.5$ over the
twelve earlier cohorts; the monthly long--short portfolio earns 7.5\% per year with a Sharpe ratio
of 1.32, against 1.1\% and 0.22 before; a market-model event study finds that firms in the top
tercile of pre-launch AI disclosure earned cumulative abnormal returns $+10.1$ points higher than
the bottom tercile over the year after 30~November 2022; and a placebo that runs five non-AI
semantic factors through the same pipeline finds negative post-2022 spreads for ESG and climate
language. The post-2022 spread survives the exclusion of the ``Magnificent~7'' and of the ten
largest firms, alternative signal definitions, and the restriction to firms listed since 2009.

Our third finding, which the first version of this paper did not report, is that the post-2022
evidence is considerably thinner than these point estimates suggest, and we regard documenting its
fragility as the paper's main contribution. The post-period consists of three annual cohorts. The
2022 cohort earned a spread of $+9.3$ points, the 2023 cohort $-4.0$, and the 2024 cohort $+38.4$;
two pre-launch cohorts (2012 and 2015) earned more than the 2022 cohort. Dropping the 2024 cohort
returns the post-period spread to $+2.6$ points, indistinguishable from the pre-period, and
winsorizing returns at the first and ninety-ninth percentiles halves the pooled post-period spread
from $+14.7$ to $+7.2$ points, because the 2024 cohort contains several memory and storage
manufacturers with forward returns above 600\%. The long--short return has a loading of 0.47 on an
Information Technology sector factor built from the panel itself, and once that factor is added
to the six-factor model the post-period alpha falls from 5.2\% to 1.9\% per year and loses
significance; sector-by-year fixed effects halve the post-period cross-sectional slope and leave it
insignificant. In a difference-in-differences design that fixes each firm's AI exposure at its
2019--2021 level, the year-by-year premium is significantly positive not only in 2022 and 2024 but
also in 2015 and 2016, so the pre-trend is not flat. Finally, the firm-level inference of the
event study overstates precision because every firm shares the same event date; a portfolio test
that allows for cross-sectional dependence gives $t=2.3$ at the one-year horizon but only $t=1.4$
at sixty trading days.

Taken together, the evidence is consistent with a sector-level re-rating of persistent AI exposure
after 2022, concentrated in the 2024 cohort and in the technology sector, rather than with a stable
premium on AI language or with a sharp, datable repricing of firm-specific disclosure. The
textual content of filings existed for a decade and was barely priced; after 2022 it became
correlated with returns, but largely through the sectors that the AI build-out favoured. Whether
this reflects a durable change in the pricing of AI exposure or a single extraordinary episode
cannot be settled with three post-launch cohorts, and we say so.

\paragraph{Contribution.} The paper makes three contributions. First, methodologically, it
demonstrates an embedding-based, anchor-similarity design for measuring a multi-faceted economic
theme in disclosure at the scale of an index over fifteen years, and documents its measurement
properties, including the sensitivity of annual scores to the composition of the scored corpus.
Second, substantively, it provides look-ahead-free evidence on the time profile of the pricing of
AI language, showing that the cross-sectional relation was negligible before 2022 and large but
concentrated afterward. Third, it provides a template for disciplining regime claims built on short
post-event samples: cohort-by-cohort and leave-one-out estimates, pooled firm-clustered tests,
sector-factor spanning, winsorization, and event-study inference robust to cross-sectional
dependence. The remainder of the paper is organized as follows. Section~\ref{sec:literature}
situates the study and states the hypotheses. Section~\ref{sec:data} describes the corpus, the
semantic measures, and the timing convention. Section~\ref{sec:methodology} lays out the empirical
designs and their inferential limits. Section~\ref{sec:results} presents the main results,
Section~\ref{sec:robustness} the robustness and identification tests, Section~\ref{sec:discussion}
the interpretation and limitations, and Section~\ref{sec:conclusion} concludes.