The introduction of ChatGPT coincided with faster pricing of complex earnings information: firms whose annual reports were harder to parse saw a significant compression of post‑earnings drift after the tool's release, with the strongest effects for non‑US issuers and low‑institutional‑ownership stocks.

Main Finding

The release of ChatGPT (Nov 30, 2022) is associated with faster price discovery for complex earnings disclosures consistent with a reduction in investors' information-processing costs. Using a pre-release, text-based processing-cost index (and token-level transformer perplexity), the author finds that after ChatGPT became available the continuation component of post‑earnings announcement returns falls more for firms whose annual reports were harder to process before the release (coef. −0.010, t = −3.4). Effects are strongest for non‑U.S. ADRs and firms with low institutional ownership; a later mobile‑app release (May 18, 2023) produces an additional compression (continuous spec coef. −0.021, t = −2.84). Results are described as consistent with a consumption‑margin (investor processing) channel, but some pre‑period dynamics and underpowered reversal tests counsel caution about strong causal claims.

Key Points

• Conceptual framing: separates three margins through which AI can affect disclosure outcomes — production (firms), intermediation (analysts/auditors), and consumption (investors). Only the consumption margin makes direct predictions about price discovery.
• Identification strategy: uses ChatGPT’s dated, universal public release as an exogenous shock to investors’ processing technology; treatment intensity varies continuously by a pre‑release processing‑cost measure of 10‑K and 20‑F filings.
• Measurement innovations:
  • A processing‑cost index built from pre‑release textual features of annual reports (predetermined relative to the shock).
  • Token‑level perplexity from a transformer model to capture statistical predictability; this measure diverges from human‑oriented readability (correlation −0.45 across 587 filings).
• Main empirical results:
  • Post‑announcement continuation (drift) declines more for higher pre‑release processing‑cost firms after ChatGPT (coef. −0.010, t = −3.4).
  • Strongest effects among non‑U.S. ADRs (highest processing costs) and firms with low institutional ownership (coef. −0.017, t = −3.32).
  • Mobile app release on May 18, 2023 gives a further decline in a continuous specification (coef. −0.021, t = −2.84).
  • Drift‑based long‑short strategies earn less after the release in high‑complexity firms; the processing‑cost premium in extreme‑surprise firms narrows toward zero (not always statistically significant).
• Robustness & caveats:
  • Binary high/low complexity group tests are noisier; placebo analysis shows pre‑period dynamics that caution against a strong causal interpretation.
  • ChatGPT outage episodes offer a reversal check but are underpowered (few high‑complexity announcements in outage windows).
  • The analysis observes aggregate market outcomes, not individual investors’ ChatGPT use.

Data & Methods

• Data sources:
  • SEC annual filings (10‑K and 20‑F) to construct a pre‑release processing‑cost index.
  • Transformer language model outputs (token‑level perplexity) as an alternate complexity/statistical‑predictability measure.
  • Earnings announcement returns, parsed press‑release EPS, daily stock prices.
  • 13F filings to proxy institutional ownership.
• Sample:
  • US large caps, US small caps, and non‑US issuers listed as ADRs (three market segments to capture variation in processing costs and investor sophistication).
• Empirical design:
  • Difference‑in‑differences–style analysis exploiting the dated, universal ChatGPT release as the shock and continuous treatment intensity given by the pre‑release processing‑cost index.
  • Outcomes: decomposition of returns into announcement (immediate) vs continuation (drift) windows; efficiency ratio (share of total reaction at announcement); profitability of drift‑exploiting strategies.
  • Validation: language‑model perplexity used to validate and augment the processing‑cost index; results largely reproduce with the augmented index despite divergence between perplexity and readability.
  • Additional diagnostics: examination of institutional‑ownership interactions, sub‑sample analyses (non‑U.S. ADRs, small caps), mobile‑app release as a second dated shock, ChatGPT outage episodes as weak reversal tests, and a mid‑2021 placebo.
• Limitations explicitly acknowledged:
  • Cannot observe individual or institutional ChatGPT usage; inference is from equilibrium price outcomes.
  • Pre‑period dynamics and noisy binary splits weaken claims of sharp causality.
  • Reversal (outage) evidence is thin and underpowered.

Implications for AI Economics

• Empirical strategy lesson: When studying AI’s market effects, explicitly distinguish the margin affected (production, intermediation, consumption). Only consumption‑margin shocks map cleanly to immediate price‑discovery predictions.
• Measurement: Statistical predictability from LLMs (perplexity) can diverge substantially from human readability — researchers should choose complexity measures that match the cognitive object of interest (human processing vs. model prediction).
• Market microstructure & asset pricing:
  • A broad, dated reduction in investors’ processing costs can compress post‑announcement drift and reduce profitability of drift‑based strategies—consistent with the idea that some documented anomalies reflect costly human processing rather than risk or pure arbitrage limits.
  • Effects concentrate where processing is hardest and where institutional (sophisticated) capital is scarce, implying AI may be particularly impactful in less‑covered or cross‑border contexts.
• Policy and disclosure implications:
  • Regulators and standard‑setters concerned about disclosure complexity should consider that reductions in consumer processing costs (via public AI tools) alter how quickly and fully markets incorporate information.
  • Access and availability of consumer AI tools matter: dated public deployments (and accessibility changes like mobile apps or bans/outages) can shift equilibrium pricing.
• Directions for future research:
  • Link individual investor/institutional usage data (logs) to price outcomes to directly identify who benefits from AI.
  • Examine other disclosure types and higher‑frequency interactions (earnings calls, 8‑Ks) to generalize beyond annual reports and earnings announcements.
  • Study equilibrium responses in the production and intermediation margins (e.g., do firms change disclosure length/format when investors use AI? do analysts change coverage strategy?), and potential feedback loops across margins.
• Caution: The paper’s evidence is consistent with a consumption‑channel effect of generative AI on price discovery, but pre‑trends and underpowered reversal tests mean results should be interpreted as suggestive rather than definitive causal proof.