Fact-checking AI-generated course content

Language models produce text that sounds fluent, factual and convincing. They do so even when a statement is false. That is down to how they work: they calculate which words are likely to follow each other, not what is true. Mostly they are right; sometimes they invent figures, mix up sections of laws or reproduce outdated legal positions.

In a course this is particularly delicate. Learners trust the material, learn it and apply it. A course that teaches falsehoods does more harm than no course at all. That is why fact-checking belongs in every AI-assisted authoring process.

Not every sentence in a course is a verifiable claim. Explanations, examples and recommendations cannot be proven in the narrow sense. What can be checked are statements of fact that a source can confirm or refute.

A robust procedure runs in several steps. First, the verifiable claims are extracted from the text. Then a reliable source is found for each claim and the claim is checked against it. The result is classified and documented. Where there is a contradiction, a correction is proposed, which a person accepts or rejects.

The third category alongside supported and refuted is important: unverified. If a source cannot be reached or no suitable one is found, the claim is not false, but it is not supported either. A procedure that forces a decision in this case produces either false confidence or false accusations.

A single language model checking its own text easily overlooks the same errors it made. More robust is checking by several independent models whose verdicts are combined by a fixed, traceable rule. If the verdicts differ, the claim is not quietly confirmed but flagged for a person to check.

Sources should be archived. Web pages change or disappear; evidence that consists only of a link may lead nowhere in a year. An archived copy with a retrieval date records what was there at the time of checking.

Not every question has a clear answer. In some fields there are different, well-founded positions. Fact-checking should not decide such questions but present the positions side by side with their evidence. That is more honest and more instructive than forced clarity.

Facts also age. Laws change, limit values are adjusted, standards appear in new versions. Courses should therefore be checked not just once when created but regularly, especially for legal and compliance topics.

Document every check so that it can be traced later: which claim, which source, which result, who decided. Such a record helps with learners’ queries, with audits and with the question of whether a course needs updating after a change in the law. It also shows that an organisation takes its duty of care seriously.

In Universe, verifiable claims are checked when a slide is generated, before graphics and narration are produced: two language models check independently, a rule-based judge combines their verdicts, and sources are retrieved and archived as evidence copies. Besides “supported” and “refuted” there is “unverified”; contested questions are shown with both evidenced positions. Administrators set how often the whole library is checked, and the fact-check record shows every check at the push of a button.