The U.S. tells courts to treat LLM training on copyrighted text as fair use—shifting AI’s biggest IP risk from “content owners” to “model builders”

The core event is a filing by the U.S. Department of Justice in the “OpenAI, Inc. Copyright Infringement Litigation” pending in the Southern District of New York. In a statement of interest dated Sept. 1, 2026 and filed Sept. 1, 2026, the government asked the court to reject an attempt to narrow fair-use doctrine so that it would effectively prohibit training LLMs on copyrighted text.

In plain terms, the government’s position treats “training copies” as qualitatively different from “reusing the work,” and it repeatedly frames the training stage as transformative learning rather than content substitution.

Copyright fair use in the U.S. is a case-by-case test across statutory factors—so a policy filing doesn’t “change the law” overnight. But the government can influence what courts weigh first.

Here, the statement of interest attempts to shift the analysis away from a broad “content gets paid” logic and toward a use-by-use, function-by-function view of what the training copy is doing. The government emphasizes: (1) the training purpose is to build an LLM’s internal statistical representations; (2) commercial operation doesn’t automatically defeat fair use when the use is highly transformative; and (3) training copies generally do not act as public substitutes for the original work. The filing also warns that an overly restrictive approach would tend to raise barriers that only large players can afford.

The practical chain is not “publishers → LLMs” directly. It runs through: (a) publishing text as training inputs; (b) ingestion and storage; (c) model training where copies are used to learn patterns; (d) deployment where the model generates outputs; and (e) user-facing applications. The government’s statement is primarily about step (c)—and it explicitly argues that questions about outputs should be assessed case-by-case rather than used to automatically condemn the training stage.

For listed companies tied to LLM infrastructure, the biggest near-term “model” change is not revenue—it’s expected risk and uncertainty. When a court accepts a narrower view of fair use, the market often prices the AI layer as if it must buy rights at scale or treat ingestion as inherently settlement-driven.

If courts align with the government’s statement, the expected cost of building and retraining models can shift. That doesn’t erase IP risk (outputs, derivative claims, and other claims remain), but it can reduce the probability-weighted impact of “training-stage must license” outcomes. In other words, the filing is designed to make the highest unmodeled liability less about training ingestion and more about specific downstream uses.