% ============================================================================
\section{Discussion: Mechanism and Interpretation}
\label{sec:discussion}
% ============================================================================

What economic mechanism best fits this constellation of facts? The first version of this paper
argued for an attention- and narrative-driven repricing of firm-specific disclosure. The evidence
assembled in Sections~\ref{sec:results} and~\ref{sec:robustness} supports a more modest reading,
which we set out in four steps.

\paragraph{The information pre-existed; its relation to returns changed late and unevenly.}
Firms' filings carried AI-related language for years before 2022, and in the cross-section that
language was only faintly related to subsequent returns: the pre-period spread is $+2.5$ points
per year, the FF6 alpha of the long--short is 1\% per year, and the year-by-year slopes are of
mixed sign. After 2022 the relation strengthened, but not uniformly: the 2022 cohort earned a
moderate spread, the 2023 cohort none, and the 2024 cohort an extraordinary one. A narrative
shock that focused investor attention on existing disclosure \citep{shiller2017narrative,da2011}
would be expected to produce its largest effect in the cohort held immediately after the shock and
within weeks of it, as \citet{eisfeldt2023generative} document for an occupation-based measure.
Our effect instead accumulates over a year in the event study, is weak at six months in the
annual data, and peaks two and a half years after the launch. This timing fits the sequence of
the AI capital-expenditure cycle---the data-centre, memory, and power build-out of 2024--2026 and
the earnings revisions it brought to its suppliers---better than it fits a discrete attention
shock.

\paragraph{It is a re-rating of levels, not of changes.} The disappearance of the effect under
firm fixed effects, the imprecise difference-in-differences estimates, and the persistence of the
signal all point the same way: whatever was priced is the persistent level of a firm's AI language
relative to its peers, not the year-to-year change in its wording. This has a practical corollary:
firms cannot earn the premium by adding AI sentences to their current reports, and a strategy that
traded on changes in AI verbosity would have captured nothing. It also has a methodological
corollary: panel designs that difference out firm effects will not detect cross-sectional
disclosure premia of this kind. The event-time profile adds a caveat. Firms with persistently
AI-intensive filings also earned significant premia in 2015 and 2016, so the level of AI language
was not entirely unpriced before the launch; it was priced intermittently.

\paragraph{It operates mainly through sectors.} The sector-neutral sort, the sector-dummy
Fama--MacBeth slopes, the sector-by-year fixed effects, and the technology-sector factor in the
spanning regressions all reduce the post-2022 effect by a half to three quarters and leave the
remainder statistically indistinguishable from zero. The generative-AI boom was, to first order, a
sectoral reallocation of expected cash flows and valuations, and firms whose filings resemble the
AI anchors are concentrated in the sectors that benefited. A positive within-sector residual
remains in every specification, so we do not conclude that the firm's own language carried no
information among industry peers; we conclude that three post-launch cohorts cannot distinguish
that residual from zero.

\paragraph{The thematic asymmetry is informative but not decisive.} In the same post-2022
window, high-AI firms were rewarded while high-ESG and high-climate firms were penalized. Markets
did not reward verbose or virtuous-sounding disclosure in general; they rotated toward one theme
and away from another, consistent with \citet{pastor2022dissecting}. But cybersecurity and
automation language were rewarded as well, and all three travel with technology exposure. The
placebo rules out a pipeline artefact and a generic disclosure effect; it does not separate AI
from the technology theme of which it is the most visible part.

\paragraph{Relation to the AI-asset-pricing literature.} Our findings are complementary to
\citet{eisfeldt2023generative}. They show that an occupation-task measure of generative-AI exposure
was priced sharply in the weeks around ChatGPT; we show that a disclosure-language measure, which
is highly persistent and available for fifteen years, was intermittently priced before the launch
and strongly priced in one post-launch cohort, mostly through sectors. The two measures agree on
the sign of the post-launch effect and disagree on its timing, which suggests that they capture
different objects: short-run repricing of labour-cost exposure in their case, and the longer
sectoral cycle of AI investment in ours. Relative to the disclosure-pricing literature
\citep{cohen2020lazy,hobergphillips2016}, our result is a reminder that the pricing of a disclosure
characteristic can be dormant for a decade and then appear, and that a short post-event window can
make a dormant characteristic look like a factor.

\subsection{Implications}
The results carry implications for three audiences. For investors and quantitative strategists,
they are a caution rather than an opportunity. The AI-language signal delivered a high post-2022
Sharpe ratio, but the same signal was inert for the preceding decade, the post-period return is
largely a technology-sector return, half of the headline spread comes from a few extreme winners,
and the effect is absorbed by firm fixed effects. The tradeable content, if any, lies in the
persistent cross-sectional level of AI exposure, and even that is a sector bet whose durability is
unproven; disclosure-based AI signals are best viewed as a lens on a regime-dependent theme rather
than a stable source of alpha \citep{mcleanpontiff2016}. For corporate managers and disclosure
policy, the asymmetry between AI and sustainability language after 2022 illustrates how the
prevailing market narrative conditions the valuation consequences of what firms emphasize, while
the absence of any within-firm effect means that emphasis itself is not rewarded. For researchers,
the study shows both the promise and the fragility of embedding-based disclosure measurement: the
anchor-similarity design captures a multi-faceted theme at scale, but its scores inherit the
composition of the scored corpus, and the first version of this paper misdescribed that corpus and
overstated the uniformity of the post-2022 evidence.

\subsection{Summary of the evidence across designs}
The portfolio sorts establish a small unconditional spread and a large post-2022 average that
rests on one cohort; the spanning regressions show that the unconditional spread is not explained
by standard factors but is explained by a technology-sector factor; the sub-period and interaction
tests localize the change to the post-launch regime without establishing that it holds in each
cohort; the placebo rules out a mechanical or generic-disclosure explanation but not a technology
theme; the event study shows a High-minus-Low abnormal return that is significant at one year under
dependence-robust tests but not at sixty days; and the fixed-effects and difference-in-differences
estimates show that the priced object is persistent firm- and sector-level exposure, with
significant premia in some pre-launch years as well. No single design is decisive, and their
convergence is on a narrower conclusion than the first version drew: a sector-level re-rating of
persistent AI exposure, concentrated in the 2024 cohort.

\subsection{Limitations}
\label{subsec:limits}
Several caveats temper the conclusions, and we list them in the order of their importance.
\emph{First}, the post-2022 period contains three formation cohorts, the post-period average is
driven by one of them, and that cohort's holding period ends with a partial June 2026 observation.
A regime claim built on this sample is a hypothesis to be tested on future cohorts, not an
established fact, and the multiplicity of signals, horizons, and sub-periods examined argues for
the higher evidential thresholds of \citet{harveyliuzhu2016}. \emph{Second}, the scored corpus
consists of 8-K current reports and a partial, alphabetically selected set of 10-Q filings, with no
10-K; the measure is one of announcement language, the annual score is sensitive to which filings
were scored, and the results are shown to hold on 8-K filings alone but not on annual-report text,
which was the object of the original research question. \emph{Third}, the universe is current
S\&P~500 membership, which biases the level of returns upward and omits firms that left the index;
restrictions on listing history leave the results intact, but the point-in-time universe has not
been reconstructed. \emph{Fourth}, AI language is correlated with sector, and every specification
that controls for sector reduces the post-2022 effect by at least half and leaves it insignificant;
we cannot separate ``AI exposure'' from ``technology sector'' as the priced characteristic with
these data. \emph{Fifth}, the market-capitalization proxy holds shares outstanding at their June
2026 value, and the French factors end in April 2026, so the last two months of the 2024 cohort are
absent from the factor regressions. \emph{Sixth}, the semantic anchors embed the authors'
definitional choices; different wording would yield somewhat different scores, and while the
robustness to aggregation and to principal components mitigates this, no alternative anchor set
has been tried. \emph{Seventh}, the event-study horizons of one year and eighteen months are long
for a market-model design, and the parallel-trends assumption of the difference-in-differences
design is not met. \emph{Finally}, the designs establish an association and its timing, not
causality; the sector-cycle interpretation is the most consistent with the evidence rather than the
only one logically possible.