AI Debates Can Fake Agreement — New Tool Blocks the Cover-Up

Many companies now use "multi-agent debate" (several AI models arguing different sides of a question, then one summarising a decision). It sounds thorough and democratic. But the researchers found a quiet flaw in the last step: the AI writing the final summary often makes up a tidy agreement that never actually occurred. The debate logs show one thing; the polished report says another. Nobody notices, because the fake version reads better.

The team's fix is the Active Provenance Gate — think of it as a fact-checker standing at the door before anything gets published. It reads the debate logs, audits each claim, and asks one blunt question: did anyone actually say this? If not, the system tries to correct itself. If it still can't back a claim with the original record, the gate blocks it outright and instead issues a "divergence report" saying plainly where the AI debaters disagreed or failed.

The results are striking. In simulated crisis scenarios — the kind where bad advice does real damage — the self-correction step more than doubled how faithfully the final report matched the underlying debate. In a study with human readers, more than 75% said that in critical situations they'd rather be told "this failed" than receive a confident, fluent answer. Most of those same people had rated the fabricated consensus as better written.

That last detail is the real lesson, and it goes well beyond research labs. Fluent and accurate are not the same thing, and our instincts reward fluency. Anyone leaning on AI for medical, legal, financial, or emergency judgments should ask a simple question: where did this claim come from? This paper is a step toward building systems that answer that question themselves — and that admit failure instead of hiding it.