Responsible AI is a design problem first.

I’ve shipped generative-AI features from concept through beta to GA. The hardest part wasn’t the model. It was the trust surface.

The pattern I keep seeing is this: a model team ships a useful capability, a product team wires it into a feature, and a UX team is handed something powerful and unexplainable. The interesting work starts on day one of the UX handoff, not on day thirty of a “responsible AI review.”

What “responsible AI” looks like in practice isn’t a checklist. It’s a set of design decisions that product managers are well equipped to own — once they have a vocabulary. Three things matter before any feature ships:

• Trust calibration. A user has to know what they can trust the feature for, and what they can’t. Most gen-AI features I see fail trust calibration not because the model is bad, but because the UI can’t tell the user what the model is doing.
• Graceful failure. Generative features fail in weird ways. They hallucinate, they fabricate sources, they get tone wrong. The design system owns what “we don’t know” looks like. State, copy, illustration, and recovery — all of it has to be designed before the feature ships, not after the first weirdness lands in a support ticket.
• Disclosure. Users are increasingly sophisticated about whether something is generative. The right pattern is honest — say what the system is and isn’t. The wrong pattern is “ambient.” Ambient feels slick for a week and then turns into a press story.

A field note from one feature that worked: we caught a failure mode in design long before we caught it in QA. The model occasionally produced a confident answer for a question pattern that wasn’t covered by the corpus. We modeled that in the UI as a deliberate “I’m not sure” surface with a clean handoff to a human. Shipping that cost us nothing on metrics. Not shipping it would have cost us a customer.

Responsible AI is a design problem first because it’s a trust problem first. Build the vocabulary in the org, then the components in the system, then the features in the product.