spb/wp17_uqo

1% ============================================================================2\section{Discussion: Mechanism and Interpretation}3\label{sec:discussion}4% ============================================================================56What economic mechanism best fits this constellation of facts? The first version of this paper7argued for an attention- and narrative-driven repricing of firm-specific disclosure. The evidence8assembled in Sections~\ref{sec:results} and~\ref{sec:robustness} supports a more modest reading,9which we set out in four steps.1011\paragraph{The information pre-existed; its relation to returns changed late and unevenly.}12Firms' filings carried AI-related language for years before 2022, and in the cross-section that13language was only faintly related to subsequent returns: the pre-period spread is $+2.5$ points14per year, the FF6 alpha of the long--short is 1\% per year, and the year-by-year slopes are of15mixed sign. After 2022 the relation strengthened, but not uniformly: the 2022 cohort earned a16moderate spread, the 2023 cohort none, and the 2024 cohort an extraordinary one. A narrative17shock that focused investor attention on existing disclosure \citep{shiller2017narrative,da2011}18would be expected to produce its largest effect in the cohort held immediately after the shock and19within weeks of it, as \citet{eisfeldt2023generative} document for an occupation-based measure.20Our effect instead accumulates over a year in the event study, is weak at six months in the21annual data, and peaks two and a half years after the launch. This timing fits the sequence of22the AI capital-expenditure cycle---the data-centre, memory, and power build-out of 2024--2026 and23the earnings revisions it brought to its suppliers---better than it fits a discrete attention24shock.2526\paragraph{It is a re-rating of levels, not of changes.} The disappearance of the effect under27firm fixed effects, the imprecise difference-in-differences estimates, and the persistence of the28signal all point the same way: whatever was priced is the persistent level of a firm's AI language29relative to its peers, not the year-to-year change in its wording. This has a practical corollary:30firms cannot earn the premium by adding AI sentences to their current reports, and a strategy that31traded on changes in AI verbosity would have captured nothing. It also has a methodological32corollary: panel designs that difference out firm effects will not detect cross-sectional33disclosure premia of this kind. The event-time profile adds a caveat. Firms with persistently34AI-intensive filings also earned significant premia in 2015 and 2016, so the level of AI language35was not entirely unpriced before the launch; it was priced intermittently.3637\paragraph{It operates mainly through sectors.} The sector-neutral sort, the sector-dummy38Fama--MacBeth slopes, the sector-by-year fixed effects, and the technology-sector factor in the39spanning regressions all reduce the post-2022 effect by a half to three quarters and leave the40remainder statistically indistinguishable from zero. The generative-AI boom was, to first order, a41sectoral reallocation of expected cash flows and valuations, and firms whose filings resemble the42AI anchors are concentrated in the sectors that benefited. A positive within-sector residual43remains in every specification, so we do not conclude that the firm's own language carried no44information among industry peers; we conclude that three post-launch cohorts cannot distinguish45that residual from zero.4647\paragraph{The thematic asymmetry is informative but not decisive.} In the same post-202248window, high-AI firms were rewarded while high-ESG and high-climate firms were penalized. Markets49did not reward verbose or virtuous-sounding disclosure in general; they rotated toward one theme50and away from another, consistent with \citet{pastor2022dissecting}. But cybersecurity and51automation language were rewarded as well, and all three travel with technology exposure. The52placebo rules out a pipeline artefact and a generic disclosure effect; it does not separate AI53from the technology theme of which it is the most visible part.5455\paragraph{Relation to the AI-asset-pricing literature.} Our findings are complementary to56\citet{eisfeldt2023generative}. They show that an occupation-task measure of generative-AI exposure57was priced sharply in the weeks around ChatGPT; we show that a disclosure-language measure, which58is highly persistent and available for fifteen years, was intermittently priced before the launch59and strongly priced in one post-launch cohort, mostly through sectors. The two measures agree on60the sign of the post-launch effect and disagree on its timing, which suggests that they capture61different objects: short-run repricing of labour-cost exposure in their case, and the longer62sectoral cycle of AI investment in ours. Relative to the disclosure-pricing literature63\citep{cohen2020lazy,hobergphillips2016}, our result is a reminder that the pricing of a disclosure64characteristic can be dormant for a decade and then appear, and that a short post-event window can65make a dormant characteristic look like a factor.6667\subsection{Implications}68The results carry implications for three audiences. For investors and quantitative strategists,69they are a caution rather than an opportunity. The AI-language signal delivered a high post-202270Sharpe ratio, but the same signal was inert for the preceding decade, the post-period return is71largely a technology-sector return, half of the headline spread comes from a few extreme winners,72and the effect is absorbed by firm fixed effects. The tradeable content, if any, lies in the73persistent cross-sectional level of AI exposure, and even that is a sector bet whose durability is74unproven; disclosure-based AI signals are best viewed as a lens on a regime-dependent theme rather75than a stable source of alpha \citep{mcleanpontiff2016}. For corporate managers and disclosure76policy, the asymmetry between AI and sustainability language after 2022 illustrates how the77prevailing market narrative conditions the valuation consequences of what firms emphasize, while78the absence of any within-firm effect means that emphasis itself is not rewarded. For researchers,79the study shows both the promise and the fragility of embedding-based disclosure measurement: the80anchor-similarity design captures a multi-faceted theme at scale, but its scores inherit the81composition of the scored corpus, and the first version of this paper misdescribed that corpus and82overstated the uniformity of the post-2022 evidence.8384\subsection{Summary of the evidence across designs}85The portfolio sorts establish a small unconditional spread and a large post-2022 average that86rests on one cohort; the spanning regressions show that the unconditional spread is not explained87by standard factors but is explained by a technology-sector factor; the sub-period and interaction88tests localize the change to the post-launch regime without establishing that it holds in each89cohort; the placebo rules out a mechanical or generic-disclosure explanation but not a technology90theme; the event study shows a High-minus-Low abnormal return that is significant at one year under91dependence-robust tests but not at sixty days; and the fixed-effects and difference-in-differences92estimates show that the priced object is persistent firm- and sector-level exposure, with93significant premia in some pre-launch years as well. No single design is decisive, and their94convergence is on a narrower conclusion than the first version drew: a sector-level re-rating of95persistent AI exposure, concentrated in the 2024 cohort.9697\subsection{Limitations}98\label{subsec:limits}99Several caveats temper the conclusions, and we list them in the order of their importance.100\emph{First}, the post-2022 period contains three formation cohorts, the post-period average is101driven by one of them, and that cohort's holding period ends with a partial June 2026 observation.102A regime claim built on this sample is a hypothesis to be tested on future cohorts, not an103established fact, and the multiplicity of signals, horizons, and sub-periods examined argues for104the higher evidential thresholds of \citet{harveyliuzhu2016}. \emph{Second}, the scored corpus105consists of 8-K current reports and a partial, alphabetically selected set of 10-Q filings, with no10610-K; the measure is one of announcement language, the annual score is sensitive to which filings107were scored, and the results are shown to hold on 8-K filings alone but not on annual-report text,108which was the object of the original research question. \emph{Third}, the universe is current109S\&P~500 membership, which biases the level of returns upward and omits firms that left the index;110restrictions on listing history leave the results intact, but the point-in-time universe has not111been reconstructed. \emph{Fourth}, AI language is correlated with sector, and every specification112that controls for sector reduces the post-2022 effect by at least half and leaves it insignificant;113we cannot separate ``AI exposure'' from ``technology sector'' as the priced characteristic with114these data. \emph{Fifth}, the market-capitalization proxy holds shares outstanding at their June1152026 value, and the French factors end in April 2026, so the last two months of the 2024 cohort are116absent from the factor regressions. \emph{Sixth}, the semantic anchors embed the authors'117definitional choices; different wording would yield somewhat different scores, and while the118robustness to aggregation and to principal components mitigates this, no alternative anchor set119has been tried. \emph{Seventh}, the event-study horizons of one year and eighteen months are long120for a market-model design, and the parallel-trends assumption of the difference-in-differences121design is not met. \emph{Finally}, the designs establish an association and its timing, not122causality; the sector-cycle interpretation is the most consistent with the evidence rather than the123only one logically possible.124