Fact-check any AI answer against its sources.

An AI assistant writes its answer by extending text so that the whole stays consistent with the material the model was trained on. Fluency and a confident register are properties of text produced that way, and neither of them tracks whether the statement matches the world. A wrong sentence has the same syntax, the same assurance and the same length as a correct one, so the wording alone gives the reader nothing to tell them apart by. The risk is highest where an AI answer is most specific: dates, amounts, names, statute numbers, version labels and anything that changed after the training data was frozen.

A source address changes what the reader is able to do. Instead of deciding whether to trust the tool, they can open the link, read the passage and disagree with it, and a third person can repeat that inspection later, which is what makes a fact-check auditable. Citations also expose mismatches a verdict alone hides: the material may cover a different year, a different jurisdiction or a narrower case than the question. Where nothing supporting turns up, the claim is marked unclear rather than quietly passed as supported. A citation still records only what was consulted, and does not settle whether the reading of it was right. Opinions, advice and forecasts cannot be checked this way at all, because they state no facts, and the checker leaves them out.

Verifying a claim is not the same operation as generating it again. Asking once more, of the same model or a different one, draws another sample from the same process, so agreement between two runs measures stability rather than truth. The fact-checker on this page therefore does not stop at a second pass through a model: it separates the pasted text into individual claims, takes at most four of the most important ones, compares each against material found on the web, and records the outcome next to the address where it was checked. The verdict for the whole answer follows from how those claims came out — how many were supported and how many contradicted by sources — not from the model's own view of its work. The share link carries the pasted text rather than the finished result, so opening it checks the facts anew. DI Cloud applies the same split into claims to answers written inside the workspace.