Stewardship in AI
Everyone building with AI right now is making decisions whose consequences land on people who did not choose to be part of them. That’s not a reason to stop. It’s a reason to hold the decisions differently from ordinary product calls.
I build voice agents, so the version I live with is concrete: a synthetic voice on a phone call with someone who may not know, may not be technical, and may be calling about something that matters to them.
What that implies in practice
Disclosure isn’t a compliance checkbox. A person on a call should know what they’re talking to, early, in the first turn. Every argument for burying this is an argument that the interaction works better when the person is mistaken, which answers the question.
Escalation must be real. “Talk to a human” has to reach one, within the same call where possible. A system that recognises the request and routes it into a queue that ends nowhere is worse than one that never offered.
Auditability before scale. Logs, transcripts, retention limits set on purpose, and the ability to answer what happened on a specific call at a specific time. This is unglamorous and it is what lets you tell the truth after an incident rather than reconstruct a story.
Humans on the consequential branch. Not on everything — that’s theatre and it doesn’t survive volume. On the branch where the system commits to something irreversible on someone’s behalf.
The cost runs in the direction you’d expect. Telling a caller in the first turn that they’re speaking to an AI lowers completion, at least initially — some people hang up who would otherwise have carried on. That’s the trade, stated plainly. It’s also decreasingly optional: bot-disclosure rules already exist in several jurisdictions, which makes the real choice between doing it now on purpose and doing it later badly, because someone made you.
Where it breaks
Two failure modes, and the second is the one that gets less attention.
The first is obvious: ethics as positioning. A page about responsible AI, a paragraph in the deck, no mechanism behind either. Everyone can spot this from outside, eventually.
The second is safety theatre that shifts risk rather than reducing it. Consent language nobody reads. A human-in-the-loop who approves four hundred items an hour and is therefore approving nothing. Both create a record of diligence and a person to blame, which is a legal artefact, not a safety one. It’s possible to add these in good faith and end up with a system that is less safe and better defended.
And the trade is real. Disclosure lowers conversion. Escalation costs headcount. Retention limits lose you training data. Anyone claiming these are all free is selling something. The argument for them isn’t that they’re costless — it’s that the alternative is a category that gets regulated into a shape nobody chose, after an incident that was preventable and predictable.
Takeaway: ethical design is product-market fit for the future.