What do the main providers say about training?
OpenAI states that by default it does not train on data from ChatGPT Business, ChatGPT Enterprise, ChatGPT Edu or the API Platform, unless you explicitly opt in to share data. It also says it uses data from versions of ChatGPT and other services for individuals, which puts a staff member's personal account in a different category from a company workspace. Microsoft states that prompts, completions, embeddings and training data sent to models sold by Azure, including Azure OpenAI models, are not available to OpenAI and are not used to train foundation models without your permission. AWS states that Bedrock content is not used to improve the base models and is not shared with model providers.
Which settings can change the default?
The headline statement describes the default position. Features that staff or developers switch on later can change what is kept and who sees it, and none of them requires a new contract to be signed. Review the product settings as well as the terms, and repeat the review whenever a new feature is enabled in your workspace or application. Ask the person who administers each AI product which of the features below are switched on today.
- Feedback buttons: OpenAI says data you explicitly opt in to share, for example through its feedback mechanisms, may be used to improve its services.
- Personal accounts: staff using an individual consumer account fall outside the business-tier commitments.
- Stored conversation features: on Azure, the Responses API, Assistants threads and stored completions keep content in your resource until you delete it.
- Fine-tuning: training files you upload are stored by the provider to build your own model.
How long do providers keep prompts for abuse monitoring?
Abuse monitoring is the main reason a provider keeps prompts it does not train on. OpenAI may retain API inputs and outputs for up to 30 days to provide the services and identify abuse, and customers with a qualifying use case can request zero data retention for eligible endpoints. Microsoft runs automated abuse checks on Azure and may store a sample of flagged prompts and completions for review by authorised Microsoft employees, in the geography of your resource. Approved customers can apply for modified abuse monitoring, which turns that storage off, and Microsoft shows it as a ContentLogging value of false. On Bedrock, retention is set by a data retention mode, and each model declares which modes it allows, so two models in one account can behave differently.
Who at the provider can read your data?
Training and human access are separate questions. OpenAI says access to stored business data is limited to authorised employees who need it for engineering support, abuse investigation and legal compliance, plus specialised third-party contractors bound by confidentiality who review for abuse and misuse. Microsoft says its human reviewers see Azure prompts only when the abuse system has flagged them, using secure workstations and just-in-time approval. AWS describes a retention mode called aws_review for models whose provider requires human review, carried out by AWS without sending content to the model provider. For a law firm, health provider or accounting practice, record these access paths in your risk assessment alongside the training position.
What should you check in the contract and settings?
Read the product terms and data processing addendum for the exact product and tier you will use, not a marketing summary. The OAIC suggests businesses find out whether service terms give the developer access to data your organisation inputs or generates, and carry out due diligence that a product suits its intended uses. Ask the provider for written answers to the questions below, and keep a dated copy of every page you relied on, because provider documentation changes often.
- Is customer content used to train or improve any model, and what opt-in mechanisms exist?
- How long are prompts, outputs and uploaded files kept, and in which country?
- Who can access stored content, under what approval, and are contractors involved?
- Is zero data retention or modified abuse monitoring available for this product, and what does approval require?
- How will you be told when these terms change?
Does private hosting remove the question?
Running an open-weight model on infrastructure you control takes the model provider out of the data path, so there is no provider training or abuse-monitoring retention to assess. The questions move to your own setup: who can read the application's logs, how long conversation history is kept, which administrators can see prompts, and whether a monitoring or support tool sends data elsewhere. A private deployment with loose logging can expose more than a well-configured enterprise API. Hypothetical example: a firm self-hosts a model but sends full prompt logs to a shared analytics tool that its whole IT contractor team can query. The provider question has gone, and an internal access question has replaced it.