GRG News — Public interest. Private misgivings.

When should an AI ask a human?

Agents can do most of the work without making human absence the design goal.

Writing in One Useful Thing, Ethan Mollick proposes the “Twilight Factory”, developed with Lilach Mollick: agents handle much of the work but deliberately call people back in.

Their four signals are approval, expertise, diverse perspectives and decisions worth keeping human. The argument is not against automation. It is for designing the hand-off before the lights go out.

GRG desk view: Automating the drudgery is useful. Automating every interesting decision is a less attractive employment proposition.

Reading brief / Background

The model is only part of the machine

Choosing an AI now means judging the brain, the product and the machinery around both.

Mollick separates an AI system into three layers: the model, the app people use and the harness that supplies tools, instructions and room to act.

The same underlying model can behave very differently when its harness can research, create files, operate a browser or test its own work. That makes a product name a poor proxy for what a system can actually finish.

This February essay is filed as background reading, not a current model ranking or buying guide. Product names, versions and prices date quickly; the model–app–harness distinction travels better.

GRG desk view: Evaluate the job a system can demonstrably finish, not just the name on its model selector.

When should an AI ask a human?

The model is only part of the machine

A note on the method.