The Times case is about training, not your prompt

On 4 September, OpenAI, Microsoft, the New York Times, and a group of authors filed competing summary-judgment motions in Manhattan. They are asking Judge Sidney Stein to decide major parts of the consolidated copyright fight without a trial. The central dispute is whether copying works to train the models was fair use. This is not a ruling. I am not a lawyer; this is the operator read.

Training is one copying event. A user pasting an article into a product is another. An output that substitutes for the original can create a third question. One eventual ruling may influence the others, but a win for a model company would not become a blanket license for customers to ingest and republish whatever they want.

Four honest ways to read it

• The model-company read: training turns works into statistical capabilities and does not ordinarily supply readers with a market substitute for a specific article or book.
• The publisher-and-author read: the systems copied protected works without permission, can reproduce or closely imitate them, and compete in markets the rightsholders built.
• The government read: fair use should remain flexible and fact-specific; the Justice Department has supported applying existing doctrine broadly rather than creating a categorical rule.
• The operator read: the motions do not change Tuesday’s workflow. Source rights, product terms, retrieval configuration, output review, and publication risk still belong on separate lines of the checklist.

What is real

The case has reached a consequential phase: competing motions now put core fair-use arguments in front of the court. The parties disagree not only about transformation but also about market harm, memorization, outputs, licensing, and the significance of acquiring works through unauthorized channels. Those factual questions are why a headline declaring either side the winner is premature.

The skeptical read

Both camps have incentives to describe a clean principle where the record may produce a narrower answer. “Training is transformative” skips how the data was obtained and what outputs can do. “The model contains our works” can skip the difference between learned parameters and a searchable archive. Summary-judgment briefs are advocacy, not neutral explanations of the evidence.

What other people are saying

• CAIDP litigation tracker and case documents — A useful public index of the consolidated docket, the competing theories, and the September motion schedule.
• Associated Press on the Justice Department filing — The government urged a broad, fact-specific application of fair use without endorsing every practice alleged in the case.
• Axios on the summary-judgment phase — A concise account of what each side is asking the judge to decide and why the result matters beyond these parties.

What to watch next

Watch which questions the judge treats as legal issues and which require disputed facts, then watch any distinction among acquisition, training, memorization, retrieval, and outputs. Amicus briefs are due in October. The useful signal will be the court’s boundaries, not which side writes the more triumphant press release.

What to do on Tuesday

• Do not paste paywalled news into a consumer tab and republish the summary.
• If the job is “what did the Times say,” cite and link the Times. If the job is “draft our memo,” use material you own or are licensed to use.
• An enterprise workspace can improve contractual and data-handling terms. It does not erase copyright or your publication obligations.

Skip this brief if

Skip it if you do not put other people’s sentences into a model and then into a customer-facing page. Skip it if counsel already has a license and a review queue. Buy nothing because of a motion. Change the paste habit because of the motion.