Both chat assistants warn, in writing, not to trust them fully

The writer's problem

A writer using a general-purpose chat assistant for research summaries, brainstorming or a first draft needs a documented answer from the company itself to one specific question: how reliable is this for facts, and what should be checked before any of it is used. OpenAI and Anthropic each publish a help article addressing that question directly for their own product.

What the documents show

OpenAI's article, Does ChatGPT tell the truth?, states plainly that ChatGPT can be helpful but is not always right, and names hallucination as the term for confidently stated but false information, giving incorrect dates or facts and fabricated quotes or citations as examples. It states a knowledge cutoff limits what the model knows unless search or a deep-research tool is enabled, and advises using ChatGPT as a first draft, not a final source, verifying quotes and citations against outside documents. Anthropic's article on Claude providing incorrect or misleading responses gives a parallel account, stating directly that Claude can write things that might look correct but are very mistaken, attributing this partly to outdated training data and partly to fabricated content resembling correct output, and advising that users should not rely on Claude as a singular source of truth and should check underlying source material directly. Both are the companies' own current documentation of a known limitation, not an independent study of how often it occurs, and neither states a measured error rate.

The editorial choice

The two articles use different vocabulary, hallucination against very mistaken content, to describe functionally the same behavior: confident, fluent, factually wrong output. An editor advising a writer on using either tool for research support should treat this as a limitation shared across the underlying technology rather than a defect specific to one company's product, and should apply the same verification step, checking the original sources directly, to output from either assistant rather than trusting one brand's synthesis more than the other's by default.

What stays with the author

Both companies place the verification obligation explicitly on the person using the tool. OpenAI's article calls for treating output as a first draft, not a final source, and Anthropic's calls for careful scrutiny and reviewing cited original material directly. Neither document claims its own product has solved the limitation it describes, and neither resolves what a writer should do when an assistant's synthesis is useful but no traceable source exists to check it against, which is left as a judgment call for the person doing the writing rather than something either help page settles.

• Is there an original, checkable source behind this specific claim, or only the assistant's synthesis of one?
• Would the same fact survive being checked against the underlying document rather than the summary of it?
• Does either company's own article actually claim the accuracy this piece of writing is about to rely on?

Two companies, describing two different products, arrive at the same instruction in their own documentation: treat the draft as a draft, and check it against something that is not the assistant itself.