Interactive Essay
The Explanation Gap
Your system passed the audit. Could the person it just rejected understand why? Seven questions about something you already have in production.
There is a confusion I have watched repeat itself in risk committees and boardrooms for years. Someone asks: is this system auditable? Everyone nods, because the technical answer is yes. There are logs. There is traceability. There is a forty-page report a regulator could read line by line if something goes wrong.
But that is not the question that matters to the person the system has just denied a loan, a promotion or a medical coverage decision. That person does not ask is it auditable? They ask why me? And there, very often, the same system that passed the most rigorous audit has absolutely nothing to offer them.
Law 6 of the Positronic Decalogue separates the two on purpose: an AI must be auditable and capable of explaining its decisions in a manner comprehensible to appropriate audiences. The key word is audiences, plural. Auditable for a regulator and comprehensible for the affected person are not the same standard, and designing only for the first is designing to pass exams, not to treat people well.
This audit takes about three minutes. Answer for one real system that decides or recommends about people — hiring, promotion, credit, pricing, eligibility, triage. Nothing is stored and nothing is sent anywhere.