AI models produce text that reads as though someone checked it. Often nobody did. This page is the plain version of what that means when you use this site.
Models invent things, fluently
A language model predicts text that fits the question. When it does not know something, it does not stop — it produces text that looks like an answer. That is where invented court cases, invented citations, invented product features and invented dates come from. The give-away is not the tone, because the tone is always calm; it is specificity without a source. A model that names a study, a page number or an exact figure is the model you should check first, not last.
Agreement is a signal, not proof
Seeing five families say the same thing is genuinely useful. It usually means the claim is common, well established and repeated consistently in the material they learned from. It does not mean the claim is true. Families can share the same popular misconception, the same outdated figure, or the same confident phrasing of something nobody has verified. Treat agreement as “probably fine, and cheap to accept” and disagreement as “stop here”.
Disagreement is the most useful thing on the page
When two families give you different numbers, different steps or different advice, you have learned something no single chat box would have told you: this is contested, or ambiguous, or one of them is wrong. Pin the two answers side by side, find the sentence where they part company, and check that one sentence. That is far less work than checking a whole answer, and it is where nearly all the value of comparing models sits. How to check if an AI answer is correct covers the habits.
When to verify before acting
Always, for anything with a real consequence. Specifically:
- Health and medicine. Dosage, interactions, symptoms, whether something needs urgent care. Ask a pharmacist or a doctor.
- Law and money. Contracts, tenancy, employment, tax, benefits, anything with a deadline. Rules differ by country and change often, and models are frequently working from an older version.
- Safety. Electrical work, gas, chemicals, food safety, load limits, anything where being wrong is dangerous.
- Anything you will publish or send. Names, quotes, statistics, links and dates in a model’s answer should be treated as unverified until you have seen the source yourself.
Our AI Disclaimer is the formal version of this section.
The agreement summary is itself an AI answer
The short “Where they agree” note at the end of a round is written by one of the model families, from the other answers. It can miss a contradiction, smooth over a difference that matters, or describe an agreement that is not really there. It is there to point you at the interesting part of the page, not to certify anything. When the summary and the answers disagree, believe the answers.
What this site cannot fix
We do not train the models and we cannot correct one. If a family gives a bad answer we can tell you which one it was, we can retry it, and we can switch it off if the problem is persistent. What we cannot do is guarantee the next answer. That is why the counter, the timings and the disagreements are all on screen: the honest version of this tool is one that shows you its workings rather than one that promises a right answer.
If something on this site produced a genuinely unsafe answer, tell us through the contact page. Describe the topic and what happened; you do not need to send the whole question.