The DOJ Backed AI Training Fair Use. The Copyright Office Didn't. Here's What That Split Means.

The US Department of Justice filed a brief in the New York Times vs OpenAI and Microsoft case supporting AI training as fair use. The Copyright Office, which spent two years developing a position on exactly this question, reached the opposite conclusion. When two arms of the federal government contradict each other on the same legal question, the implications are not theoretical.

What Happened

The DOJ brief argues that AI model training — the process of exposing a model to large volumes of text or images to develop its capabilities — constitutes transformative fair use under copyright law. The core argument: training copies are never made publicly available; the model learns patterns from data rather than reproducing the content; and the resulting output is sufficiently different from the input to qualify as transformative.

This directly contradicts the position the Copyright Office developed and issued earlier in the year, which rejected a blanket fair use exception for AI training and proposed that some training uses constitute infringement depending on the nature of the material and the resulting output.

The timing is notable: the former US Copyright Register was reportedly fired earlier this year amid political pressure related to AI copyright positions. Whether that sequence has a direct connection to the DOJ’s current brief is not documented, but it adds context to the regulatory environment in which the brief was filed.

Why It Matters

The NYT vs OpenAI/Microsoft case is the highest-profile AI copyright litigation in the United States, and the DOJ’s intervention as an amicus curiae signals that the federal government’s position on AI training legality has commercial implications the administration is actively trying to shape.

The split between DOJ and the Copyright Office creates a real legal uncertainty problem for organizations building AI systems on training data. If the courts follow the DOJ’s fair use framing, the current training data practices used by the major model developers are retroactively validated. If the courts follow the Copyright Office’s more restrictive framework, the exposure is significant.

For enterprises deploying AI systems — particularly organizations considering training custom models on proprietary or licensed data — this split is not resolved by waiting. The case will take years to litigate fully.

What This Means for Builders

For organizations building on top of foundation models (rather than training their own), the litigation outcome matters primarily for what it signals about the model providers’ legal exposure and whether training data practices will require remediation that affects model capabilities or licensing terms.

For organizations exploring fine-tuning or custom training on data they own or license, the DOJ brief offers cover — but not legal certainty. The Copyright Office’s contrary position is the more cautious baseline for enterprise legal teams assessing risk.

Source: The Decoder

Content created with AI assistance and reviewed for accuracy.