AI labs make a clear promise to corporate customers: your data will not be used to train the model. The policies of OpenAI and Anthropic say as much.
But a recent episode showed the promise alone is not enough. The problem arose not in training but in how long logs are kept.
What happened
Anthropic said it would store usage logs from its flagship model Fable for 30 days. That did not mean using them for training; it only meant keeping the logs for a period.
Even so, Palantir, Nvidia and Booz Allen Hamilton pulled back from using the model for sensitive work. The policy had not changed; the retention period had, and that alone was enough.
Why the reaction was so sharp
What those three companies share is the nature of their work. In defence, intelligence and critical infrastructure, a prompt sitting anywhere for 30 days creates a contractual problem.
| Question | What the policy answers | What the customer asks |
|---|---|---|
| Is data used for training | No | — |
| Is data retained | 30 days | Where, and who can access it |
| Is deletion guaranteed | At the end of the period | Can it be audited |
The right-hand column shows the real issue. The question an enterprise customer asks is not "do you train on it" but "where does the data sit and who will verify that for me".
The speed of the reaction is notable too. The pullback came within days of the announcement rather than weeks, meaning these companies' internal review processes were already primed on this question.
Policy and trust are not the same
The privacy policies of AI companies have tightened markedly over the past two years. The trust problem remains, because a policy is a declaration while trust is a matter of verification.
With a cloud provider that distinction became institutional over the years: independent audit reports, region selection, customer-held encryption keys. The equivalents on the model provider side have not settled yet.
The case for keeping logs
There is a reasonable rationale on the provider side too. Detecting abuse, debugging and investigating security incidents all require keeping logs for a period.
So there is no easy answer: zero retention makes abuse detection impossible. The data behind the threat reports published this month is precisely those logs.
That tension will only become more visible. Providers want to keep more logs in order to detect abuse while enterprise customers demand fewer, and the two meet in the same contract.
What it means for a regulated buyer
The lesson here concerns how contracts are read. The sentence "your data is not used for training" says nothing about retention period, storage location or access rights.
For an organisation assessing this under data protection rules, three questions are clear: in which country are the logs held, for how many days, and who can access them in that window. If the answers are not written in the contract, the policy itself does not count as a guarantee.