What do I do about ML models I can't fully explain to a regulator?
You built something useful that you cannot fully trace. That is not a moral failing — complex tools often work this way. But the fear you feel is a signal, not a nuisance. It is telling you to act while you still have time, not to wait until someone with authority asks first. Notice what you are actually afraid of. It is not the model's math. It is standing in a room with no answer. That is a different problem, and it is one you and your team can solve this week, not one that requires you to understand every weight in the network. A regulator does not expect you to explain the model like a mathematician. They expect you to show that a responsible process produced it, watches it, and can stop it. That process is within your power. Build it together, and stop treating your own uncertainty as a confession you must avoid making.
You will not make the model simple. Stop trying. What you control is the record around it: why you chose it, how you tested it, who approved it, what happens when it fails. Build that record now, with your team, and the unexplainable model becomes a well-governed one.
What changes unlock by starting
- A written record of why and how the model was built, ready before anyone requests it.
- One accountable owner for explainability, instead of a vague shared hope.
- A team that can answer hard questions without scrambling.
- Less private dread, because the gap is managed instead of avoided.