研究：模型发布后市场如何重估「被AI替代」的公司
作者用15次模型发布做事件研究，发现2025年后市场把AI从「更便宜的劳动力」改读成「竞争对手」，产品被替代的公司两日跌约1%。
This note asks a simple question: which businesses is artificial intelligence replacing? The usual answer counts how many of a firm's employees do tasks a model can do and call the firm exposed. That is the right measure for a question about jobs and the wrong one for a question about firms. A firm whose workers are exposed may gain, because its costs fall. A firm whose product is exposed loses, because its revenue does. The two groups overlap much less than the debate assumes. Stock prices are best for telling them apart. Every time a lab releases a major new model, investors have to decide, within hours, who wins and who loses. So I ran an event study. I took fifteen releases, from ChatGPT on 30 November 2022 to Claude Opus 4.6 on 5 February 2026, and measured the abnormal return of hundreds of securities over the two days that followed each one, against a market model estimated on the preceding year. The data come in three layers: the forty-nine Fama–French daily industry portfolios, thirty-six sector and thematic ETFs, and seventy-four individual stocks grouped by what they sell rather than whom they employ. GPT-6 Astra, released two days before I write this, is treated as a postscript. The sign of the reaction flipped. Around the releases of 2022–2024, firms whose products compete with the machine earned slightly positive abnormal returns; the market read AI as cheaper labour for them. Around the releases of 2025–2026 the same firms lost about one per cent in two days, and around the three agentic releases of early 2026 they lost three. Labour-heavy incumbents that buy information services moved the other way. The market has changed its mind about what the technology is. It used to be an input. Now it is a competitor, and the businesses it competes with are the ones whose moat was a proprietary corpus and the workflow built around it. I want to be explicit about how this note was produced, because the method is part of the finding. I did not write a single line of code. I described the hypothesis to Claude and asked for the study and a short paper. It fetched French's archives, pulled daily prices for seventy-four stocks and thirty-six ETFs from Yahoo's API, ran the regressions in JavaScript, drew three figures, and wrote five thousand words with nineteen references. The first complete draft came back in about an hour. A round of corrections took another. Two hours, from a paragraph of instructions to a twelve-page note with tables, figures and a literature review. I have run studies like this for forty years. At UCLA this was a month of a doctoral student's time. The finding of the note is that the market now prices frontier models as substitutes for firms that sell information work. The making of the note is the same finding, observed from the inside. Read the results with the caveats they deserve: fifteen events is few, some are confounded, and the t-statistics mean little. https://drive.google.com/file/d/1D-oaKKQiHXf5phkj40ivHatI9VLwKflw/view?usp=share_link