A Fair-Use Stress Test for an AI System

A useful copyright judgment about a generative AI system begins with a fact record, not with the word “training.” The record connects the work, its source, the actor, the stage of development, the purpose, the copies made, the deployed controls, and the behavior users can reach. Putting those facts together makes it possible to compare more and less defensible patterns without pretending that a preliminary assessment decides liability.

The value of this stress test is practical. It forces the same questions to be asked of a dataset provider, a model developer, a retrieval product, and a user-facing service. It also keeps uncertainty visible: a missing training record, an untested guardrail, or an unknown access path is not evidence that the favorable version of events is true. It is a fact that still needs to be resolved.

Build the fact record before judging the use

Start with the work. Is it a novel, a news article, a photograph, a sound recording, a software program, a database, or another kind of material? Is the relevant expression protected, or is the dispute really about facts, ideas, or a general style? The kind of work will matter later, but even at the start it prevents the vague phrase “copyrighted data” from replacing the object of the inquiry.

Next identify the actor and stage. A collector may download and curate the work. A trainer may copy a dataset into high-performance storage and present examples in batches. A distributor may release model weights. A deployer may add a retrieval database, search, filters, or other tools. A user may request an output. These actors can be related by contract or business strategy, but they may not have made the same copy or pursued the same purpose.

Then describe the specific act. Was the work downloaded, transferred, reformatted, filtered, compiled, shown to the model, retained in weights, placed in a retrieval index, retrieved for a particular question, or reproduced in an output? “The model used it” is not specific enough. The act determines which technical evidence and which exclusive right may be relevant.

The source and access path belong in the same record. A work may be owned, licensed, in the public domain, available through an authorized subscription, visible on the open web, or obtained from an unauthorized source. The file that enters a dataset may look identical in each case. The Copyright Office’s analysis nevertheless treats lawful access as a material fact and says knowing use of pirated or illegally accessed material should weigh against fair use without automatically deciding the defense.

Purpose comes next, and it must describe the deployed function rather than only the immediate engineering step. A developer may say that a dataset was assembled for research, while a commercial product later uses the resulting model to generate expressive material. Conversely, a company may conduct a noncommercial research project, publish its findings, or release weights without deploying the model as a monetized substitute for source works. The point is not to distrust every description of purpose.