How to fact-check AI answers
Models are fluent whether they are right or wrong. Here is the checking routine we use before anything AI-assisted reaches this site.
A language model generates likely text, not verified fact. Fluency is not evidence. Corewire’s own editorial rule is that every central claim needs a primary source before publish, and the same routine works for anyone.
Check the three fragile things first
- Names, numbers, dates. The most common failures. Verify each against a source you open yourself.
- Citations. Models can invent plausible papers, cases, and URLs. If it cites something, open it; if it does not exist or does not say that, discard the claim.
- Anything recent. Training data has a cutoff. For current events, check a primary source: the company’s own filing, the paper itself, the official announcement.
Ask the model to show its work
Add one line to your prompt: “For each factual claim, say whether it comes from the material I provided or from general knowledge, and flag anything you are unsure of.” This does not make the model honest, but it makes unsupported claims visible. Better still, give it the source material directly, the approach behind retrieval-augmented generation.
Escalate with the stakes
A brainstorm needs no checking. A blog post needs the fragile-three pass. Anything medical, legal, or financial needs a qualified human, full stop. We wrote about what happens when that discipline slips in our reporting on shadow AI inside companies.
Related resources
- Spotting AI mistakes: this guide as a scan-in-seconds checklist.
- How to write a prompt that works: constraints that prevent invention up front.
- Glossary: RAG, tokens, and the other terms used here.