Ask an AI for emissions factors and it may return a convincing table with the wrong numbers. Ask for a regulation and it may invent a clause. A fabricated citation often looks exactly like a real one.

That is why I call it a brilliant editor and a terrible fact-checker. The prose may be excellent. The evidence still needs checking.

Consider a consultant preparing a carbon baseline. The assistant produces an electricity emissions factor and cites a government agency. The consultant opens the source and discovers the number is from an older reporting year. Nothing in the polished answer warned them. The mistake would have passed into a client report without a source check.

The practical rule is simple. Open the reference. Check the year, geography, unit and methodology. If the source does not support the claim, remove it. A second AI checker can help find inconsistencies, but it cannot take responsibility for the final result.

BIRI™ addresses a different part of the problem: source reliability. A source rating helps decide where to investigate, not whether a particular statistic is true. A highly rated website can still carry an outdated table.

Article 42 deals separately with what to tell clients about your AI policy. Here the point is narrower: never confuse a fluent answer with verified evidence.