Notes on keeping my AI experiments honest
Responsible AI doesn't have to be a sweeping framework. Sometimes it's three small habits you build into the team's daily work.
The phrase "responsible AI" usually conjures images of governance committees, policy documents, and lengthy review processes. Those things matter, but in my experience, what actually shifts how a product behaves is much smaller: a few habits, repeated.
1. Write the user back to yourself
Before shipping any new AI feature, write a short note from the user's perspective describing what they tried to do and what happened. Read it out loud. If the imagined user sounds confused, frustrated, or misled — even subtly — the feature isn't ready.
This sounds trivial. It is not. Most product teams optimize for the happy path. Reading from the user's perspective forces you to consider the unhappy ones.
2. Log the why, not just the what
It's easy to log model inputs and outputs. It's harder — and far more useful — to log why a particular response was generated. Which retrieval results were used? Which system prompt branched? Which guardrail fired?
Months later, when something looks off, the why is what lets you understand and fix it. Without it, you're guessing.
3. Designate a "concern channel"
A simple Slack channel, mailing list, or shared doc where anyone on the team can flag something that feels off — even if they can't articulate why. No bureaucracy. No required action.
Most of the time nothing happens. But the few times it does, it surfaces something that would have otherwise become a much bigger problem. Quiet unease is often the earliest signal you can get.
The point
None of these habits are heroic. They don't require new tooling or executive approval. They just require someone to start them — and the team to keep them up.
That, more than any framework, is what responsible AI looks like in practice.