spb/wp17_uqo

1% ============================================================================2\section{Introduction}3\label{sec:introduction}4% ============================================================================56The arrival of generative artificial intelligence has been the defining corporate-narrative shift7of the 2020s. In the three years following the public launch of ChatGPT in late 2022, the market8capitalization of AI-linked firms rose by trillions of dollars, ``AI'' became the most frequently9invoked theme on earnings calls, and the language of corporate disclosure filled with references to10large language models, generative tools, intelligent automation, and the compute infrastructure11that powers them. This paper asks a question that is simple to state but demanding to answer12well: did the \emph{language} of AI in firms' mandatory disclosures carry information about their13subsequent stock returns, and if so, when did that information begin to be priced and through14what economic channel?1516Answering the question requires three ingredients that have only recently become jointly17feasible. The first is a large, machine-readable corpus of corporate disclosure. The second is a18measurement technology that captures the \emph{meaning} of text rather than the frequency of19keywords---one that recognizes that a firm describing how it ``deploys machine-learning models to20optimize its supply chain'' is talking about AI even when the acronym never appears. The third is21a look-ahead-free mapping from disclosure to tradeable returns. Our laboratory is the set of 50022current S\&P~500 constituents over 2010--2026. Our raw material is the text of their SEC filings as23parsed by the EDGAR Research Database, which embeds every text chunk with the24\texttt{all-MiniLM-L6-v2} sentence transformer \citep{reimers2019sbert} and scores it by cosine25similarity to a set of written ``anchor'' passages that define eight economic concepts. Three of26these anchors isolate distinct facets of AI engagement: \emph{internal AI adoption} (using AI to run27the business), \emph{AI infrastructure exposure} (capital and revenue tied to the compute28build-out), and \emph{AI product-disruption risk} (the danger that AI commoditizes the firm's own29offerings). The result is a continuous, comparable, firm-year measure of how intensively, and in30what sense, each firm engages with AI in its own words.3132Two features of the corpus must be stated at the outset, because they were misdescribed in the33first version of this paper and they shape the interpretation of everything that follows. First,34the filings that had been parsed and scored when the panel was built are the Form~8-K current35reports of all 500 firms (\num{103926} filings) and the Form~10-Q quarterly reports of 84 firms36(\num{3710} filings); no Form~10-K annual report had been processed. The AI measure is therefore a37measure of the language of current reports---material agreements, earnings releases, executive38changes, shareholder votes---supplemented, for a sixth of the firm-years, by quarterly reports.39Second, the ``year'' of the panel is the calendar year in which the filings were dated, not the40fiscal year they cover. We treat both features as measurement facts to be documented and tested41rather than as reasons to discard the data: Section~\ref{sec:data} describes the corpus in detail,42and Section~\ref{sec:robustness} shows that the main results are unchanged when the signal is43rebuilt from 8-K filings alone.4445We pair the semantic scores with monthly adjusted prices for the 500 tickers from Yahoo Finance and46with the Fama--French five factors plus momentum. To avoid look-ahead bias we follow the47portfolio-formation convention of the cross-sectional literature: the signal of filing year $Y$ is48used to form portfolios at the end of June of $Y{+}1$ and we measure the forward twelve-month49return from July of $Y{+}1$ to June of $Y{+}2$. Because every filing of year $Y$ is dated on or50before 31~December of $Y$, the signal is in the public information set at least six months before51the holding period begins. This yields \num{6852} firm-years with a tradeable forward return,52spanning formation years 2010 through 2024.5354Our first finding is that AI language is at best weakly priced unconditionally. Sorting firms into55quintiles of AI-disclosure intensity each year produces an equal-weighted top-minus-bottom spread56of about 5 percentage points per year with a $t$-statistic below two, and the six-factor alpha of57the corresponding long--short portfolio is 1.9\% per year and insignificant. Our second finding is58that the relation strengthens after the launch of ChatGPT. Over the three post-launch formation59cohorts (2022--2024) the composite spread averages $+14.6$ points per year against $+2.5$ over the60twelve earlier cohorts; the monthly long--short portfolio earns 7.5\% per year with a Sharpe ratio61of 1.32, against 1.1\% and 0.22 before; a market-model event study finds that firms in the top62tercile of pre-launch AI disclosure earned cumulative abnormal returns $+10.1$ points higher than63the bottom tercile over the year after 30~November 2022; and a placebo that runs five non-AI64semantic factors through the same pipeline finds negative post-2022 spreads for ESG and climate65language. The post-2022 spread survives the exclusion of the ``Magnificent~7'' and of the ten66largest firms, alternative signal definitions, and the restriction to firms listed since 2009.6768Our third finding, which the first version of this paper did not report, is that the post-202269evidence is considerably thinner than these point estimates suggest, and we regard documenting its70fragility as the paper's main contribution. The post-period consists of three annual cohorts. The712022 cohort earned a spread of $+9.3$ points, the 2023 cohort $-4.0$, and the 2024 cohort $+38.4$;72two pre-launch cohorts (2012 and 2015) earned more than the 2022 cohort. Dropping the 2024 cohort73returns the post-period spread to $+2.6$ points, indistinguishable from the pre-period, and74winsorizing returns at the first and ninety-ninth percentiles halves the pooled post-period spread75from $+14.7$ to $+7.2$ points, because the 2024 cohort contains several memory and storage76manufacturers with forward returns above 600\%. The long--short return has a loading of 0.47 on an77Information Technology sector factor built from the panel itself, and once that factor is added78to the six-factor model the post-period alpha falls from 5.2\% to 1.9\% per year and loses79significance; sector-by-year fixed effects halve the post-period cross-sectional slope and leave it80insignificant. In a difference-in-differences design that fixes each firm's AI exposure at its812019--2021 level, the year-by-year premium is significantly positive not only in 2022 and 2024 but82also in 2015 and 2016, so the pre-trend is not flat. Finally, the firm-level inference of the83event study overstates precision because every firm shares the same event date; a portfolio test84that allows for cross-sectional dependence gives $t=2.3$ at the one-year horizon but only $t=1.4$85at sixty trading days.8687Taken together, the evidence is consistent with a sector-level re-rating of persistent AI exposure88after 2022, concentrated in the 2024 cohort and in the technology sector, rather than with a stable89premium on AI language or with a sharp, datable repricing of firm-specific disclosure. The90textual content of filings existed for a decade and was barely priced; after 2022 it became91correlated with returns, but largely through the sectors that the AI build-out favoured. Whether92this reflects a durable change in the pricing of AI exposure or a single extraordinary episode93cannot be settled with three post-launch cohorts, and we say so.9495\paragraph{Contribution.} The paper makes three contributions. First, methodologically, it96demonstrates an embedding-based, anchor-similarity design for measuring a multi-faceted economic97theme in disclosure at the scale of an index over fifteen years, and documents its measurement98properties, including the sensitivity of annual scores to the composition of the scored corpus.99Second, substantively, it provides look-ahead-free evidence on the time profile of the pricing of100AI language, showing that the cross-sectional relation was negligible before 2022 and large but101concentrated afterward. Third, it provides a template for disciplining regime claims built on short102post-event samples: cohort-by-cohort and leave-one-out estimates, pooled firm-clustered tests,103sector-factor spanning, winsorization, and event-study inference robust to cross-sectional104dependence. The remainder of the paper is organized as follows. Section~\ref{sec:literature}105situates the study and states the hypotheses. Section~\ref{sec:data} describes the corpus, the106semantic measures, and the timing convention. Section~\ref{sec:methodology} lays out the empirical107designs and their inferential limits. Section~\ref{sec:results} presents the main results,108Section~\ref{sec:robustness} the robustness and identification tests, Section~\ref{sec:discussion}109the interpretation and limitations, and Section~\ref{sec:conclusion} concludes.110