Anyone who uses AI assistants in development, support or documentation sends content: code snippets, draft contracts, error logs, sometimes customer data.Two current reports put the question of what happens to these inputs after they are submitted at the centre – and how reliable the providers' assurances are.
Retention periods as an exclusion criterion
OpenAI and Anthropic assure business customers that their data will not be used to train the models.However, this assurance covers only part of the processing chain.When Anthropic announced that it would store usage logs of its flagship model Fable for 30 days, Palantir, Nvidia and Booz Allen Hamilton restricted its use for sensitive tasks.
What is remarkable about this is not the period itself, but the reaction.No training with customer data does not automatically mean: no storage.Logs are created for abuse detection, for support and for quality assurance – and they exist for a defined period, meaning they are in principle viewable, analysable and, if in doubt, also subject to official requests.For organisations with security clearance requirements or heavily regulated data sets, that is enough to rule out a tool for certain classes of tasks.
Humans read along
The second report concerns the evaluation of model responses.According to a report by 404 Media, OpenAI has hundreds of contract workers read real ChatGPT conversations and rate them on a scale of one to seven.The aim is, among other things, to reduce sycophancy and overly human-like behaviour in the model – that is, characteristics that can only be assessed on real dialogues.
The prompts are anonymised for this purpose.However, anonymisation removes only identifying metadata, not the content: in individual cases the texts can still contain sensitive information – for example when someone copies a complete document or a database query into the input field.Anyone who wants to rule this out must actively deactivate the Improve the model for everyone setting.It is therefore active by default.
The common thread in both cases
Both reports describe different providers and different mechanisms – log storage on the one hand, human quality evaluation on the other.The common core, however, is the same: between the public assurance we do not train on your data and the actual processing lies an area that is governed by default settings and secondary conditions and that many teams never check.
There is also a difference that is often overlooked in practice: private accounts, team plans and enterprise or API access are not subject to the same rules at the major providers.When employees work with their personal access, the conditions assured for business customers do not necessarily apply.This is precisely where shadow IT arises – not out of malice, but because the private account is available faster than the approved access.
What companies should look out for when choosing
From the incidents described, some sobering check questions can be derived that should be answered before introducing an assistant:
- Retention period: How long are inputs and logs retained, and can the period be shortened contractually or set to zero?
- Purpose of storage: Is it used exclusively for abuse detection and operation, or also for quality evaluation and product improvement?
- Third-party access: Can external service providers or contract workers view content, and under what conditions?
- Default settings: Which data-relevant options are active by default, and who in the company changes them centrally?
- Dependence on plan: Which assurances apply only to enterprise or API use and not to private accounts?
- Verifiability: Can data handling be documented so that it stands up in a data processing agreement or an audit?
Configuration is part of the architecture
The OpenAI case shows how much depends on a single default setting.Anyone introducing assistants in a company should not leave such switches to individual decisions, but set and document them centrally via administrative functions.This also includes a clear statement of which data categories belong in an input field at all – personal data, access credentials and contract documents generally do not, regardless of the provider's assurance.
Much can be defused technically before it becomes a contractual question.A gateway that bundles requests from internal systems can filter content before sending, insert placeholders for sensitive fields and log which systems call which models.Tasks involving particularly sensitive data can be shifted to models operated locally or in the company's own infrastructure, while general tasks continue to run via powerful external services.This division by sensitivity is usually more economical than trying to operate everything in-house – or to outsource everything.
Trust needs an alternative
The reaction of Palantir, Nvidia and Booz Allen Hamilton is also instructive because it was not a complete abandonment: what was restricted was the use for sensitive tasks, not the use altogether.This differentiation is the practicable way forward.However, it presupposes that the company knows which tasks are considered sensitive – and that a named alternative exists for these cases.
Interchangeability is equally important.
If a provider changes its storage practices, this should be able to result in a configuration decision and not require new development.Abstraction layers between application and model cost little to implement and keep this option open.
For web and software projects, this means: how an AI provider handles data is an architectural criterion, not just a matter for the legal department.Anyone who decides early on which content may leave their own network, who manages presets centrally and connects models behind an interchangeable interface remains able to act when conditions change.That is currently the more realistic assumption than expecting stable commitments.