Distillation is a copying method with a respectable name. You send a strong model thousands upon thousands of questions, keep the answers, and train a cheaper model to produce the same answers. Every query is a legitimate API call that gets billed like any other.

That is the problem OpenAI's analysts face. Nothing in a single request gives the copier away. The tell is in the pattern: many accounts asking the same shaped questions, in volume, in coordination. OpenAI's post, titled "Disrupting a coordinated model distillation campaign," says it found one and shut it down. The post is on OpenAI's own site. I have no second source for it, and the figures a reader would want to size the campaign (accounts, requests, who was behind it) are the figures to look for in it.

Think about who sits next to that traffic. A procurement lead at a bank or a retailer pays for the same model by the token. Their contract probably says what the model may do for them. It likely says little about what the vendor does when its monitoring sees a burst of structured, repetitive requests, which is how a legitimate bulk job, such as a nightly sweep of thousands of documents, also looks. A copier and a diligent customer can produce the same graph.

The vendor's incentive is plain. Copied answers train a competitor that sells the same capability at a lower price, so the vendor will tune its detection toward catching too much. The customer who gets caught by that tuning learns about it when an account is throttled or suspended.

The concern is usually argued at national scale. OpenAI's post puts it at the level of one account manager and one monitoring rule.

So the question for the next renewal is concrete. Ask the vendor whether it runs extraction monitoring on your account, what pattern trips it, who gets notified before service is cut, and how a legitimate bulk workload gets cleared in advance. If the answer is in no document you can sign, you are relying on a rule you cannot see.