New AI Tool Stops You From Misreading AI's Explanations

When an AI turns down your loan, rejects your job application, or flags your medical scan, it usually offers a short reason: 'your credit history,' 'insufficient experience,' 'irregular pattern.' Those explanations are simplifications — designed to be readable, not complete. Researchers call them 'leaky explanations': simple on the surface, but you can only understand them correctly if you know the details that were left out. In interviews with nine people, the team found that when an explanation didn't answer someone's real question, they didn't notice the gap. Instead, they invented a plausible-sounding reason of their own and believed it.

That matters because acting on the wrong reason wastes your time and money. If you think the AI denied you credit because of one late payment, you might spend months fixing the wrong thing while the real factor — say, how much of your available credit you're using — goes untouched. It also quietly shifts blame: the AI gets credit for being 'transparent' while you carry the risk of misreading it. And with AI now influencing hiring, insurance, lending, and medical triage, a confident wrong guess can have real consequences.

The proposed fix is a conversational helper called Explanation Navigator. Rather than dumping a longer explanation on you, it watches for a mismatch — when what you're asking about isn't something the explanation can actually answer. It then surfaces the relevant hidden details, offers extra explanations for the unmet question, and nudges you to correct your own mistaken inference instead of just handing you the answer.

The evidence comes from an online study with 316 participants, where the approach helped people recognise and fix their invented reasoning, moving them toward a more accurate understanding. Two honest caveats: this is a preprint, meaning it hasn't yet been vetted by peer review, and the interviews involved only nine people. It's a lab demonstration, not a shipping product — and having AI explain AI's explanations adds its own layer of trust to think about.