IP & Technology · March 12, 2025 · 7 min read
When the research team walks: protecting AI trade secrets in 2025
Model weights, training pipelines and evaluation data now sit at the center of valuations — and they leave the building with people. What courts in New York are granting, and what companies should have in place before the resignation letters arrive.
For most AI companies, the most valuable assets are not patents. They are the trained weights of a model, the data pipeline that produced them and the evaluation sets that tell the team whether the next run is better. None of that is registered anywhere. All of it can be copied in an afternoon — and the people who know where it lives are the people most likely to be recruited by a competitor.
Why AI cases look different
Traditional trade secret disputes turn on customer lists, formulas or manufacturing know-how. AI disputes turn on artifacts that are large, technical and easy to describe badly. A court asked to enjoin the use of “the training methodology” will want to know exactly which methodology, how it differs from what is published, and why the competitor could not have arrived at it independently.
That is why the first week of these cases is won or lost on specificity. Under the and New York common law, the claimant must identify the secret with enough particularity for the defendant to know what it may not use — and for the court to write an order it can enforce.
What courts are granting
Temporary restraining orders
Where there is forensic evidence of exfiltration — a bulk download of a model repository the week before a resignation, a personal cloud account synced to a work laptop — federal courts in the Southern District have been willing to grant temporary restraining orders within days, typically requiring return of materials, a forensic examination of devices and a prohibition on use pending a preliminary injunction hearing.
Limits on the new employer
Courts are more cautious about orders that stop a competitor from shipping a product. Relief against the new employer usually depends on evidence that the company knew, or should have known, what its hires brought with them. That makes the onboarding records of the recruiting company one of the most important early discovery targets.
What to have in place before anyone resigns
Know what is secret
Maintain a living register of the artifacts that matter — model checkpoints, data sources, labeling guidelines, evaluation suites — with owners and access lists. When a dispute starts, this register becomes the backbone of the complaint instead of something assembled under deadline.
Log access, not just storage
Encryption and access controls show reasonable measures; access logs prove what happened. Retain repository, storage-bucket and endpoint logs long enough to cover a typical notice period and garden leave.
Run departures as a process
Exit interviews, device returns, written certifications and a prompt review of the last thirty days of access activity turn a suspicion into evidence — or reassure the company that nothing is wrong, which is equally valuable.
The takeaway
Trade secret protection in AI is less about the lawsuit than about the preparation that makes a lawsuit unnecessary or short. Companies that can identify their secrets precisely, prove how they protect them and document departures consistently are the ones that obtain relief in days rather than months — and, more often, never need to ask for it.