The researcher asked for a draft. He got four pages, an elegant conclusion, and a whole afternoon of checking. The machine had finished. He was starting.

A researcher writes while he and a robot hold up an enormous slab above documents.
An image of the burden: continuing to work and continuing to answer for the work.

One reference pointed to the right book and the wrong page. Another looked perfect until he opened the document. The third one existed. It felt like a tiny holiday.

The pile of sources on the table kept growing. The label “made with AI” had to fit hours of someone working out what it all meant. It was a poor fit.

— If the conclusion is wrong, who answers the phone?

The robot had not asked to be treated as an infallible author. It could locate documents, separate versions, place excerpts side by side. That was what the researcher asked for on the second attempt.

The researcher writes while the robot scans documents beside him.
The robot finds and organizes. The person checks and decides.

This time, each passage came alongside its source. The robot scanned; the researcher wrote. One discrepancy remained unresolved until the next day. No one had to pretend to be certain just to finish the page.

When they handed in the work, they noted the roles: assisted source organization; verification and decisions by the team. The responsibility was still there. At least now it had an address.

The draft still came out fast. The work stopped pretending it came out on its own.