I applied for a credit line increase last month, and got denied in about four seconds. The reason on the screen said "insufficient data for approval." I called, and the person on the phone read me the exact same sentence back, word for word, then admitted she couldn't see anything beyond that either — the model had made the call, and nobody on her end had access to why. Not "here's the specific thing that pushed you under the threshold." Just a closed door with a sign on it that neither of us could read.
I call this the unfalsifiable no. It's not that the decision was necessarily wrong — I have no way of knowing that either. It's that there was no way to test it. A falsifiable claim is one you could, in principle, prove false if it were false. "Insufficient data" isn't that. It's a sentence shaped exactly like a reason, sitting where a reason should be, that gives you nothing to push against, appeal with, or even verify was applied consistently to the next applicant.
This isn't a hypothetical problem anymore, it's a live regulatory one. In April 2026, the Fed, the OCC, and the FDIC jointly updated the model risk framework that banks are supposed to follow for credit scoring — and explicitly excluded generative and agentic AI from it. The newest, least explainable models are, for now, the least governed ones. A Wolters Kluwer survey of banking professionals published in June found that close to three in four banks couldn't produce a documented process for catching or rolling back a lending model that started behaving badly. And this isn't abstract risk: a student loan company settled for $2.5 million last year after its underwriting model was found to weight applicants using their college's average default rate — a variable that, unnoticed by any human reviewer, systematically pushed down approval odds for Black and Hispanic borrowers. Nobody built that on purpose. Nobody caught it either, until regulators did.
The law in most places already says a lender has to give you a specific, real reason when it says no — not a generic form response. What's changed is that a growing share of the systems making that call are structurally unable to produce one, because the model itself can't fully explain its own weighting in terms a human, or a regulator, could act on. Newton's approach to underwriting starts from the opposite constraint. Instead of a trained model producing a score nobody downstream can fully unpack, the lender's criteria are written as an explicit, inspectable set of rules — the same open policy format already used for cloud infrastructure permissions — and every credit decision comes back with a cryptographic record of exactly which rule was applied, and why the applicant's verified credentials did or didn't clear it. It's not a friendlier no. It's a no you could actually go check.
I genuinely don't know if trading a trained model for an explicit rule set costs accuracy in ways that matter — a good machine-learned model might catch real risk signals a hand-written policy would miss entirely, and "explainable" isn't automatically the same thing as "fair" or "correct." A transparent rule can still be a bad rule; it's just a bad rule you can actually see and argue with, instead of one hiding behind a sentence that only sounds like an explanation.
I still don't know what "insufficient data" meant for me specifically. Next time a piece of software tells me no, I'd like it to have to show its work, not just deliver its verdict.
