"Explainable AI" gets treated like a solved problem — add a confidence score, a highlighted feature, a plain-language summary, and the system is now transparent. But transparency and trustworthiness are not the same thing, and conflating them is exactly the risk my recent work, When Transparency Deceives, tries to unpack.
An explanation can be technically accurate and still mislead you
Here's the uncomfortable version: a system can tell you something true about how it reached a decision, and still leave you with a false sense of security about whether that decision was good. It can show you the right features, the right weights, the right visual, and still hide the thing that actually matters — whether the underlying reasoning reflects the values and context a human would consider relevant. That's the gap between an honest explainer and a merely transparent one.
Why this isn't just an academic distinction
In practice, this shows up whenever an AI system is explaining a decision that affects someone's life — a hiring recommendation, a medical triage flag, a fraud alert. If the explanation is designed to build confidence rather than to genuinely inform, people either over-trust the system (and stop questioning it) or, once they sense something's off, lose trust in explanation entirely — including the honest ones. Both outcomes are bad, and both are avoidable if we treat explanation design as an ethical safeguard rather than a UI feature.
What "good" looks like
In our work, we've been trying to systematise what separates a security- and ethics-aware explanation from a decorative one: does it expose uncertainty rather than paper over it? Does it invite scrutiny rather than discourage it? Does it hold up when the person reading it is a non-expert, not just an engineer reviewing their own model? These are the evaluation criteria we're proposing — not as a checklist, but as a starting point for taking explanation seriously as a place where ethics actually happens, not just where usability does.
I'll be presenting more on this at ICRES later this year — if you're working on XAI, robot ethics, or anything adjacent, I'd love to compare notes.