Can you trust that AI answer? A 4-question self-check

This tool does not judge the answer for you. There is no AI behind it. It only helps you decide whether the output is a draft, a thing to test, or a claim that needs outside verification.

Why these four questions

The four questions are not random; each one isolates a different lever that decides how far you can trust an answer. Question one asks about task type, because a language model is strongest at wording and framing and weakest at pinning down specific facts. Question two asks whether the answer actually contains concrete facts, since a paragraph with a real date, name, or citation carries risk that a tone rewrite does not. Question three asks whether you have checked the key source, because a fact you verified outside the chat stops being the model's claim and becomes yours. Question four asks what happens if it is wrong, since the same shaky fact is harmless in a birthday message and serious in a contract. Put together, the four answers separate "safe to use" from "test it first" from "do not rely on it alone." The model can sound equally confident in all three cases, so you need the structure rather than the tone.

Three worked examples

The logic is easier to see with real cases. Here are three requests people make every day, and what the check concludes for each.

Rewording an email. You paste a blunt message and ask the model to make it warmer. The task is language work, there are no new facts introduced, and being wrong costs almost nothing. The check lands on "safe to use as language work." This is the model's home turf, and second-guessing it here mostly wastes your time.

Asking for a historical date. You ask when a treaty was signed, and the answer states a specific year. Now there is a concrete fact, and the model can state a wrong year with the same confidence it states a right one. If the date is going into a school report the cost is moderate, so the check tells you to verify at least that one fact against a reliable source before you treat it as final. Confidence is not evidence.

Asking about medical dosing. You ask how much of a medication to take. This is a high-stakes health decision, so the check refuses to bless it no matter what the answer says. The verdict is "do not rely on it alone": an AI answer can be background reading, but the decision belongs to a qualified professional or an authoritative source. The downside of being wrong here is exactly why the tool treats this category separately.

Where this check does not help

Be clear about what this tool is not. It does not read your AI answer, and it cannot tell you whether a specific claim is true or false. It does not detect hallucinations for you, does not grade the answer's correctness, and has no model of any kind running behind it. There is no AI here at all; nothing you select is sent anywhere, and nothing is stored. What it does is narrow: it takes the judgment you would make anyway and gives it a structure, so you are less likely to be swept along by a confident tone and more likely to stop and check the one fact that matters. The verification itself is still your job.