Some tasks the vendors' own policies say need a licensed human

Start here

Knowing when to keep using an assistant is easier than knowing when to stop, and the vendors' own usage policies name specific stopping points rather than leaving the judgment entirely to a user. Reading those named categories before a task begins, rather than after something goes wrong, is the practical use of a policy most readers never open.

What the documents say

OpenAI's usage policies prohibit using its services for the 'provision of tailored advice that requires a license, such as legal or medical advice, without appropriate involvement by a licensed professional.' Anthropic's Usage Policy names a longer list of what it calls high-risk categories — legal, healthcare, insurance, finance, employment and housing, academic testing and admissions, and professional journalistic content — and states that uses in these categories require a qualified professional's human review before output reaches anyone, plus disclosure to end users 'at a minimum at the beginning of each session.' Separately, the fact-checking study cited elsewhere in this guide found that people over-relied on a language model's explanation specifically when it was fluent and wrong, a pattern the researchers reduced but did not remove by adding opposing arguments. Read together, the two policies define categories where the vendors themselves require a licensed human in the loop, and the study gives a documented reason why: a confident wrong answer is hard to catch unassisted.

Check this

This is an editorial boundary drawn from the documents above. Before relying on an assistant's output for a decision, check whether the task falls into a category either policy names as requiring licensed or qualified review. If it does, the documents state a licensed professional belongs in the loop, not that the assistant should be avoided outright. If it does not, the check described elsewhere in this guide — asking for a source, checking a number, asking for the opposite case — still applies, since the study did not find a task category immune to over-reliance.

What holds and what fails

This boundary holds for the categories the vendors actually name, which are specific and limited: legal, medical, financial, safety-critical and similarly licensed domains. It does not extend automatically to every task that feels important, and the two policies do not list identical categories. This is an editorial reading of two usage policies and one study; it does not offer medical, legal or financial advice, and neither do the documents it is built from.

• Check whether a task falls under either vendor's named high-risk categories before relying on the output.
• Route legal, medical or financial questions to a licensed professional, as both policies state.
• Apply the five-minute check to everything else, since no category was shown immune to over-reliance.

A usage policy is not only a rule for a vendor to enforce; it is also a stated boundary a reader can use before being asked to.