Why Most AI De-Identification Fails in Production

De-identifying text is easy to demo and hard to trust.

That gap is where most of the real product work begins. In a slide deck, masking names and entities can look impressive in minutes. In a real workflow, especially one involving legal, medical, or otherwise sensitive material, the problem is much larger. You need reversibility, consistency, auditability, sensible failure modes, and an experience people can trust when the stakes are not hypothetical.

That has shaped a lot of how I think about AI products. The hard part is rarely making the model do something clever once. The hard part is building a system that behaves predictably enough for professionals to use without crossing their fingers.

In practice, that means designing beyond the model. You need structure around detection, replacement, review, recovery, and traceability. You need to think about how decisions are made, how mistakes surface, and what happens when the system encounters ambiguity. You also need to care about interface quality, because trust is influenced as much by how a system behaves as by what it claims to do.

This is one of the reasons I’ve been interested in high-trust AI workflows. SmartScrub is part of that story. So is Smartnote. Both sit inside a broader belief that AI products need more than intelligence. They need operational credibility.

The more I work in this space, the less interested I am in AI theatre. I’m much more interested in products that reduce risk, make judgment easier, and fit into real working environments without demanding blind faith from the people using them.