Skip to Content

The gap between what AI can do and what we do with it

Ethan Mollick argues that today's AI models can do far more than most people use, and human strengths decide.
October 28, 2026

Ethan Mollick argues that today's AI models can do far more than most people use them for, and that human strengths make the difference.

Why this is useful to know

Ethan Mollick, a professor who has followed AI for years, argues that today's AI models can do far more than most people actually do with them. He calls this the capability overhang: the gap between what a model can do and what users do with it. In his view the newest models, which he calls GPT-6 Astra and Fable 5.1, can already do work that would take people weeks, provided you steer them well. His examples are a 3D version of the 1977 text adventure Zork and a 3D reconstruction of Umberto Eco's library. His conclusion: the difference is made not only by the model, but also by whether you bring four human strengths: deep knowledge, broad knowledge, taste and initiative. Above all, remember this: the gap is an opportunity, and people are still needed to choose, steer and judge.

When you need this

Say your team only uses AI for short texts, while much more is possible. Then you are probably leaving value on the table. Those strengths lead to a simple check. Ask yourself three things for each task. Do you know the field well enough to spot a mistake in the result? Do you know the right professional terms to steer the model precisely? And do you dare to try what is not yet possible? If the answer to any of these is no, that is your starting point. Start small: let AI do a task you know well, check the result and ask for a second version if it is not right. Decide for yourself what you do and do not leave to AI. The model executes quickly, you decide what is good enough.

How it works under the hood

Mollick does not describe a new technique but a usage pattern, and that is relevant for an application manager. The model is only part of the result. An agent that works in Blender, drives a video generator through the browser and edits the shots itself is in fact a harness: tools, a loop of planning and checking, and human decision points. Quality depends heavily on the context: the book as a source, the desired genre, a storyboard as an intermediate step. Note what he says himself: there are visible mistakes in it, and he only spotted them because he knew Zork, Eco and his own book well. For your own setup, this means you measure quality per step, keep intermediate results and record where human approval is needed. His example projects are open source, so you can check the setup yourself.

This news touches chapters 2, 25 and 26 of the course AI(l) or nothing by Newway. If you follow the course, this subject returns there at a calmer pace.

Written in response to a publication by One Useful Thing (Ethan Mollick): https://www.oneusefulthing.org/p/the-overhang