Two AI vendors tell users to verify output before use

The writer's problem

A writer using a chatbot to draft a paragraph faces a specific risk that fluent prose can hide: a confident sentence is not the same as an accurate one. Editorial workflow guidance from AI providers converges on a specific response, checking a generated passage against a source before treating it as usable prose, rather than trusting how natural it reads. OpenAI's help center article on whether ChatGPT tells the truth and Notion's AI FAQ, both retrieved 16 September 2026, state this from two different vendors.

What the documents show

OpenAI's article names the failure mode directly as a hallucination, defined as the model producing responses that are not factually accurate, including fabricated quotes, studies, citations or references to non-existent sources. Its stated practical guidance is to use ChatGPT as a first draft, not a final source and to always verify quotes, data, technical information or references to external documents. Notion's FAQ, describing its own AI feature built on large language models, instructs users to double-check all answers for accuracy. Neither company frames this as an optional courtesy; both state it as the condition under which their own tool's output should be used. Neither document specifies a shared verification method, such as a required citation check or a source ledger; each states only that verification is the user's responsibility before the text is treated as reliable.

The editorial choice

Reading the two statements together supports an editorial synthesis rather than a single vendor's claim: fluency is not evidence of accuracy, and a passage generated by either of these tools needs a specific verification step, not a general sense that it reads correctly, before an editor calls it a usable draft. What that step looks like, whether a citation check, a search against a primary source, or a second reader, is an editorial choice the vendors leave open.

What stays with the author

Both companies place verification on the user, not on the tool. A writer who publishes a fabricated quote or a nonexistent citation generated by a model remains accountable for it appearing in their manuscript; neither page suggests otherwise.

• Has every quote, date or citation in a generated passage been checked against a source the writer can name?
• Does the workflow treat model fluency as a substitute for a verification step, even unintentionally?
• If a fabricated reference reaches print, whose process failed to catch it?

Two vendors, writing independently, both tell their users the same thing: treat the output as a first draft requiring a check, not a finished, verified text.