AI can already do more than we ask, and four human edges close the gap

Ethan Mollick's argument is simple: the newest models can already do far more than almost anyone asks of them, and the bottleneck now is us, not the technology. He calls this gap the overhang. To show it, he points to a few projects that used only a fraction of a normal token budget. GPT-6 Astra turned Zork, a text adventure from 1977, into a playable 3D game and decided on its own what the monsters should look like. Fable 5.1 reconstructed the roughly 27,000-book library in Umberto Eco's Milan apartment from videos and photos, naming about 5,000 titles and marking how confident it was in each one. Another run produced a full animated book trailer, script, voiceover and edits included, work that would take a person weeks.

His point is not that the models are magic. It is that the result now depends on what the person brings. Mollick names four advantages that still belong to humans: deep knowledge, so you can judge whether an output is actually good; wide knowledge, so you can direct the model across fields; taste, so you can pick, cut and remix what it produces; and agency, the willingness to try things when nobody yet knows what works. None of these compete with the model. They steer it.

The essay reads as a correction to two common reactions, panic about being replaced and disappointment that AI has not changed daily work. Mollick's answer is that the capability is here and mostly sitting unused, waiting for people who know how to ask.

Why it matters

If you have concluded that current AI is overhyped because it has not changed your job, Mollick's case is that the limit is how you use it, not what it can do. The practical move is to find the places where your own expertise and judgment let the model do something you could not do alone.