Since ChatGPT's launch, firms more dependent on AI have been repriced: they suffer larger negative moves after unexpected Fed forward-guidance shifts and show stronger bubble-like price behavior, implying AI exposure magnifies both monetary-policy sensitivity and speculative valuation risk.

Main Finding

Firms with higher pre-2023 AI exposure in the S&P 500 (as measured from SEC 10‑K disclosures) show two consistent patterns: (1) they are more negatively affected than lower-AI firms by hawkish forward‑guidance surprises at medium to long horizons (impact ~0, negative divergence emerges over time), and (2) they display stronger evidence of mildly explosive price dynamics (GSADF/PSY tests), particularly in concentrated, mega-cap AI baskets. These patterns are robust across many specifications but are not offered as causal proof of mispricing — rather as a cross-sectional regularity linking AI exposure to both discount‑rate sensitivity and explosive price behavior.

Key Points

• Motivation: Post‑ChatGPT (Nov 30, 2022) revaluation concentrated in a small set of AI‑linked mega caps (the “Magnificent Seven”), with these firms accounting for an outsized share of S&P 500 gains and index concentration.
• Two complementary empirical lenses:
  • Local projections (LP) using high‑frequency FOMC forward‑guidance surprises to estimate cross‑sectional return sensitivity by AI exposure.
  • Phillips‑Shi‑Yu (PSY) recursive right‑tailed unit‑root (GSADF) tests to detect mildly explosive price dynamics across AI exposure portfolios and concentration baskets.
• Main LP result: High‑AI firms show near‑zero impact responses to path shocks but develop a statistically significant negative differential at longer horizons relative to No‑AI firms; Moderate‑AI is weaker and often indistinguishable from zero.
• Main PSY result: Explosive dynamics are strongest for High‑AI portfolios and concentrated AI/mega‑cap baskets (especially Magnificent‑Seven ratios). Date‑stamping shows much of the broad High‑AI vs No‑AI differential predates ChatGPT, while post‑2022 explosiveness is clearer in concentrated mega‑cap level series.
• Robustness: Results persist to alternative shock constructions (including separating monetary vs information components), alternative outcome measures (cumulative abnormal returns), sample windows, lag orders, weighting schemes (value vs equal), and excluding mega‑cap leaders. Debt data are limited, but leverage does not fully account for the LP differential.
• Limitations: The thesis does not identify a causal effect of AI per se, cannot prove deviations from fundamentals, and the PSY framework identifies statistical explosiveness (not necessarily irrational bubble pricing). Identification of the LP relies on within-sector, within‑event variation; interpretation of forward‑guidance shocks is subject to information‑effect concerns.

Data & Methods

• Sample
  • S&P 500 constituents, market data (CRSP/Yahoo/FRED), analysis window roughly 2021–April 2026 (event focus around ChatGPT, Nov 2022).
  • AI exposure coded from pre‑2023 SEC 10‑K filings into No‑AI / Moderate‑AI / High‑AI groups; additional concentrated baskets (Magnificent Seven, top AI leaders) and portfolio constructions (value‑ and equal‑weighted).
  • Supplementary descriptive evidence on capex and cash‑flow (showing a large, concentrated increase in AI‑related capex and largely self‑financed investment).
• Local projection (LP) analysis — forward‑guidance sensitivity
  • Shock: high‑frequency FOMC announcement surprises decomposed to emphasize forward‑guidance (path) surprises; also robustness checks separating monetary vs information components.
  • Design: panel LPs estimating event‑horizon return responses with fixed effects (sector × event) to isolate within‑sector, within‑meeting variation; coefficients interpreted as differential horizon responses for Moderate‑ and High‑AI versus the No‑AI omitted category.
  • Inference & robustness: clustering/inference choices and specification grid; placebo pre‑event tests; alternative shocks and outcome measures; exclusions (mega caps) and checks on leverage.
• PSY (GSADF) explosive‑dynamics analysis
  • Tests: recursive right‑tailed unit‑root testing (GSADF) with bootstrap critical values to detect episodes of mild explosiveness and date‑stamp explosive periods.
  • Units: portfolios by AI exposure, concentration baskets, and broad S&P references; both value‑ and equal‑weighted series examined.
  • Robustness: lag‑order sensitivity, wild‑bootstrap multipliers, log vs level specifications, and alternative reference choices.
• Interpretation
  • LP: evidence consistent with a duration/discount‑rate channel — growth‑type (AI‑exposed) firms more sensitive to changes in expected discount‑rate paths — but no identified causal mechanism.
  • PSY: detects mildly explosive dynamics concentrated in AI‑exposed and concentrated mega‑cap series; timing nuance (pre‑ vs post‑ChatGPT) means not all explosiveness is a simple post‑2022 phenomenon.

Implications for AI Economics

• Asset‑pricing dimension: AI exposure is a meaningful firm‑level heterogeneity dimension for monetary‑policy transmission to equity returns. Researchers and practitioners should consider AI exposure (or related growth/intangible metrics) alongside traditional financial characteristics (leverage, size, book‑to‑market) when modeling cross‑sectional responses to policy news.
• Macro‑financial and policy concerns: The concentration of market capitalisation in a few AI leaders means that forward‑guidance surprises and monetary shocks can have outsized, non‑uniform effects across the equity cross‑section and on index returns. Policymakers should be aware that communications shocks may differentially affect long‑duration, tech‑centric firms, potentially amplifying redistributional or concentration effects.
• Risk and portfolio management: Investors with passive exposure to broad indices are implicitly exposed to AI concentration risk. Active managers and risk modelers may need to account for greater discount‑rate sensitivity and mild explosiveness in AI‑exposed portfolios when constructing hedges, stress tests, and allocation rules.
• Research directions: The joint pattern — that the same AI partition organizes both forward‑guidance sensitivity and explosive price behavior — motivates causal work to unpack channels (discount‑rate vs cash‑flow vs sentiment/speculation), better measurement of AI exposure (dynamic/textual measures), and integration of investment (capex) outcomes with valuation dynamics.
• Caution: Statistical explosiveness does not equal irrational mispricing. The findings flag a cross‑sectional regularity with economic importance, but further work is needed to link price dynamics to fundamentals (real returns on AI investment, competition effects from open‑source models, and realized cash‑flows).