A confidentiality checklist differs by AI product tier

The writer's problem

An editor deciding whether to paste an unpublished manuscript into a writing tool cannot rely on one answer for whether that company trains on the input, because the correct answer depends on which specific product and tier is in use, a distinction each vendor's own privacy documentation states differently. OpenAI's business data page, retrieved 16 September 2026, and Anthropic's privacy center article on Claude and model training, retrieved the same day, each draw a line between consumer and business products.

What the documents show

OpenAI's page states we don't train our models on your organization's data by default for ChatGPT Enterprise, ChatGPT Business, ChatGPT Edu, ChatGPT for Healthcare, ChatGPT for Teachers, or our API platform. That named list does not include the standard consumer ChatGPT Free or Plus product, which OpenAI documents on a separate page. Anthropic's article opens by noting it covers consumer products such as Claude Free, Pro, Max, and points readers to a different article for Claude for Work and the Anthropic API. For the consumer product, it states chats are used to improve models only if the user opts in, if a conversation is flagged for safety review, or if the user has joined a program such as the Trusted Tester Program. ProWritingAid's help center article on data safety, also retrieved 16 September 2026, states the company does not use a customer's data or writing to train its models and that submitted text is removed from its servers once analysis completes, apart from documents saved in its own Web Editor.

The editorial choice

The editorial task is matching the product to the claim: an OpenAI Business or Enterprise seat and a personal ChatGPT login are documented differently on the same company's own site, and the same split exists at Anthropic between consumer Claude and Claude for Work. A newsroom or publisher building a confidentiality checklist should record which specific product tier a contributor is using, not the vendor's name alone, before deciding whether a manuscript excerpt can be pasted in.

What stays with the author

None of these pages resolve who is liable if a policy changes after a manuscript has already been shared, and none substitute for a publisher's own confidentiality agreement with its writers. Checking the specific document, dated by its retrieval, remains the author's or editor's own responsibility each time a new tool is proposed.

• Which specific product and tier does the vendor's training-default statement actually name?
• Does the policy differ for consumer accounts versus business or API accounts at the same company?
• Has the manuscript owner's own confidentiality agreement been checked against the tool's retention terms?

A checklist built from these three pages is only as current as their retrieval date, since the default setting is the part of a living privacy document most likely to change.