What we cannot delegate

There is something reassuring about the word delegate. It suggests that someone remains in charge. We hand over a task, receive a result, and retain the ability to correct it. Between the request and the response lies a relationship we understand, even when it goes wrong.

Artificial intelligence gives us reason to look at that relationship again. An answer can arrive before we have quite finished framing the question. It can be well written, persuasive, and consistent with what we expected. None of that tells us, on its own, whether we should act on it.

The work that remains

It helps to separate two things we often conflate: producing an answer and taking responsibility for its consequences. The first may become an increasingly simple operation. The second still involves specific people: someone on the receiving end of a decision, someone whose work changes, a community asked to accept an outcome.

Imagine an important decision made with the help of a system. The system recommends. A person agrees. Later, something goes wrong. Saying that the machine decided may describe part of the procedure. It does not answer who chose that procedure, or who could have stopped it.

Keeping hold of the question

Responsibility begins before the result. It lies in deciding what to ask of a tool, which limits to set, and what we are willing to reconsider. It also means recognizing when an answer calls for a conversation that no automated output can bring to a close.

This does not require distrust of everything. It requires judgment. A useful tool can expand our ability to act. For that to happen, we need to understand what we are trying to do and whom it might affect.

Perhaps the most important work is not competing with an intelligence that can answer faster. Perhaps it is preserving enough time to ask whether that was the question worth answering.