What is fact-checking?

Fact-checking, in AI training, is verifying every factual claim in a model's answer against reliable sources, and marking anything that is wrong, outdated or unsupported. It is one of the main defences against hallucinations.

What the work looks like

You break an answer into individual claims, look each one up, and label it as supported, contradicted or unverifiable. Projects usually specify which sources count, such as official statistics, primary legislation or peer-reviewed papers, and ask you to paste the source link for each judgment. Some also ask for a corrected version of the answer.

Skills that matter

• Knowing which sources are authoritative in a field.
• Reading carefully enough to catch a wrong year, unit or name.
• Not trusting a claim because it sounds familiar.
• Keeping clear notes so a reviewer can retrace your checks.

Example

An answer about a country's minimum wage gives the right figure but says it took effect a year earlier than it did, and cites a ministry page that does not exist. The fact-checker marks the date as wrong, the citation as fabricated, and links the correct official notice.