Who Is Accountable When the Agent Acts?
There is a quiet but enormous difference between a system that suggests and a system that acts. Autocomplete suggests; you commit. A booking agent that actually reserves the flight has crossed a line — and our frameworks for consent, authority, and liability were not built for the other side of it.
This is a first attempt at a taxonomy, grounded in the interface decisions that actually determine who ends up accountable.
The suggest → act spectrum
Accountability is not binary; it slides along how much distance sits between the model's output and a real-world effect.
- Suggest. The system proposes; a human executes. Accountability stays with the human.
- Draft-and-confirm. The system prepares the action; a human approves it. Shared, but the confirmation UI does a lot of load-bearing work.
- Act-and-notify. The system acts, then tells you. Accountability has quietly shifted, and the notification is often too late to matter.
- Act silently. The system acts within a standing authorization. Whoever granted that authorization owns the outcome — whether they understood it or not.
The interface is where authority is actually granted or withheld. Everything downstream is a consequence of that one design decision.
Consent is a UI problem before it's a legal one
We like to imagine consent as a contract. In practice it is a checkbox, a default, a phrasing. My own studies keep showing the same thing: framing changes what users believe they authorized, often dramatically, for identical underlying permissions.
If a single word in a confirmation dialog can swing what a user thinks they agreed to, then consent is a design artifact long before it is a legal one. Pretending otherwise just launders responsibility onto the least-powerful party.
A minimal accountability checklist
For any agent that can act, I want the interface to make three things impossible to miss:
- Scope. Exactly what class of actions is authorized — bounded, not vibes.
- Attribution. A durable record of who authorized what, and when.
- Reversibility. A real path to undo or halt, surfaced before the action, not buried after it.
Systems that get these right are "accountable by default." Systems that don't have simply moved the accountability somewhere the user can't see it.
Why I care about the boring version
It would be more fun to write about agency and moral status. But the decisions that will actually determine who is responsible are mundane: a default toggle, a confirmation copy, a log that does or doesn't exist.
The ethics of agentic systems will be won or lost in those small, unglamorous interface choices — which is exactly why they deserve serious attention now, before the defaults harden.