Author: Ben Thompson Publication: Stratechery (Daily Update) Date: July 7, 2025 URL: https://stratechery.com/2025/training-ai-is-not-fair-use-llms-and-scale-pushing-on-a-string/
A July 7, 2025 Stratechery Daily Update by Ben Thompson reading the Northern District of California's two June 2025 AI fair-use rulings — Bartz v. Anthropic (Judge Alsup, June 23) and Kadrey v. Meta (Judge Chhabria, June 25). Thompson argues that Chhabria's "market dilution" theory under the fourth fair-use factor is legally sound and, if widely adopted, would render most AI training on copyrighted works not fair use, even though Chhabria ruled for Meta on procedural grounds. He treats that outcome as economically damaging for open-weight model developers.
Summary of argument
Thompson contrasts the two rulings along the four statutory fair-use factors. Both Alsup and Chhabria agreed on factors 1 through 3: that the use is transformative; that it is commercial but allowed; and that the total-use-but-no-individual-work-essential character cuts in the defendants' favor.
The judges diverged on factor 4, the effect on the market. Alsup found no harm to the market for the original works. Chhabria found that LLMs uniquely flood the market with secondary works — "market dilution" — and that this will likely cause plaintiffs to decisively win the fourth factor in future cases. Chhabria nonetheless ruled for Meta because the plaintiffs failed to develop the market-dilution evidentiary record, and the opinion explicitly disclaims that it represents a holding that AI training is fair use.
Restating Chhabria's reasoning, Thompson observes that the canonical fair-use cases — Google Books and Perfect 10 — involved tools that accessed original works, whereas AI training is qualitatively different because it generates millions of competing secondary works that compete for the same market the original works occupy. Thompson agrees that "market dilution" is the right factor-4 framing for that effect, even if the doctrine has not previously needed to operate at this scale.
Key claims
Thompson accepts the legal reasoning as sound and draws out three consequences:
- The plaintiffs' evidentiary bar going forward is lower than commonly assumed. Once a plaintiff develops the market-dilution record, which Thompson regards as relatively easy with statistical evidence, Chhabria's analysis suggests fair-use defenses will fail.
- The factor-3 defense that the scale of training data renders any single work immaterial does not save the market-dilution argument, because the factor-4 question concerns output market effect, not input necessity.
- If LLM training is not fair use, the litigation cost falls hardest on entities without large-scale licensing relationships, that is, open-weight model developers. Thompson identifies this as the same "regulation favors incumbents" pattern he advances in Attenuating Innovation (AI), reappearing in copyright form.
Provenance
This is the primary Stratechery anchor for the AI fair-use debate, alongside related material:
- Fair Learning — Lemley & Casey (2021) — defendant-side academic position.
- The Hypocrisy at the Heart of the AI Industry — Reisner (2026) — empirical evidence (Schmidt video, Amodei memo) that AI labs internally treat training-data acquisition as litigation risk to manage, not as legitimate fair use.
- Bartz v. Anthropic, Kadrey v. Meta, and NYT v. Microsoft, OpenAI et al. — case-tracking pages.
Relationships
- depends-on: Ben Thompson
- related: Bartz v. Anthropic
- related: Kadrey v. Meta
- related: NYT v. Microsoft, OpenAI et al.
- related: Fair Learning — Mark A. Lemley & Bryan Casey (Texas Law Review, 2021)
- related: The Hypocrisy at the Heart of the AI Industry — Alex Reisner (The Atlantic, March 2026)
- related: AI Copyright Litigation — Analysis
- related: Ab 2013 — California training-data documentation law
- supports: Open-Source AI / Open-Weight Models — Thompson's downstream concern is that copyright maximalism most damages open-weight developers