The AI Narrative, Tested Three Ways

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

The AI boom carries a set of assumptions that are rarely checked against data directly: that AI- exposed stocks now move together more than before, that a bellwether like NVIDIA reliably rallies on its own earnings, and that a theme with this much capital and attention flowing into it should show exploitable short-horizon structure. This paper runs three independent tests of those three claims, using this program’s own standing methodology throughout — real point estimates, genuine replication — and finds no support for any of them: (1) a daily tail-co-movement test, applying the core conditional-exceedance framework to a real, dated AI-stock universe (the 52 US-listed holdings of Global X’s AIQ ETF exactly as they stood in December 2022, not selected with hindsight) in matched windows before and after ChatGPT’s launch; (2) an earnings- reaction event study, covering all 71 of NVIDIA’s quarterly reports since 2006 against the S&P 500, AMD, and the semiconductor sector; and (3) an intraday feasibility test, applying the exact two methods already used to find no short-horizon signal in crypto (naive conditional- exceedance and multifractal/structure-function machine learning) to real 1-minute NVDA, AMD, and SPY data. Test 1 finds real tail-level co-movement falling from 2,132 surviving configurations in a length- matched window before ChatGPT’s launch to 522 after — and NVIDIA, Microsoft, Meta, Amazon, and Apple each show zero surviving tail-level predictability from the rest of the basket in the post- launch window. This decline survives a direct robustness check: excluding the COVID crash from the pre-window does not shrink it, it grows to 2,678, making the decline sharper, not an artifact of one crisis episode. Test 2 finds NVIDIA beats the S&P 500, AMD, and the semiconductor sector on only 44–55% of its earnings reactions depending on the window — close to even — and its one- month return specifically relative to its own sector peers has not improved through the 2023– 2026 AI earnings supercycle, if anything running more negative than before 2023. Test 3 finds the same result already found in crypto: zero surviving configurations under the naive-exceedance method (maximum CPE achieved anywhere 0.416 for equities, 0.378 for crypto under an identical restriction, both well short of the 0.80 bar), and directional accuracy at chance (49.2–51.4%) under the machine-learning method, across three horizons and five model families. Three independently designed tests, three different timescales, three different targets — the same answer throughout: the specific, tradeable patterns the AI narrative implies do not survive contact with the data, even though the underlying price trend is real.