In article 7, we looked at why guardrails cannot guarantee that an AI answer is correct. Someone followed up: what about the checker itself — the second AI that attacks the draft? Doesn’t that solve it?
No. It helps. It helps a lot, in specific ways. But it inherits the same fundamental flaw.
Why the checker can catch things. An adversarial checker reviews the draft the way a hostile reviewer would: find weak claims, spot missing evidence, test the logic, redo the arithmetic, flag statements that contradict the sources. This works because checking structure is easier than generating it. The checker doesn’t need to know the truth to notice that section three contradicts section one, or that a percentage doesn’t match its own table, or that a claim has no citation attached. On internal consistency, completeness, and arithmetic, adversarial review is genuinely strong. Our Foundry uses it for exactly this.
Why it can’t stop hallucination. The checker is the same kind of machine as the writer: a prediction engine built for plausibility, not truth. And they often share training DNA — similar models, similar blind spots. Two people with the same wrong map will agree on the wrong route, confidently. When a hallucinated fact sounds right, it sounds right to the checker too. Worst case: the checker “verifies” the false claim and your draft now carries a false stamp of approval.
Where checkers earn their keep. Logic, consistency, arithmetic, completeness, structure, tone. The mechanical and editorial layer. That is real value — it can catch some of what’s wrong, and it does it in seconds. The factual layer — did this study really say that, is this number current, does this regulation exist — remains human work against primary sources.
In practice. Adversarial checking is a brilliant editor and a terrible fact-checker. Use it for the first job. Never delegate the second. And never let the presence of a checker replace human review — it’s a multiplier for your judgment, not a substitute for it.
