There is a large gap in AI education between technical material for engineers and content marketing dressed up as insight. Ethan Mollick occupies that gap almost alone, and does it well enough that his newsletter has become required reading for a lot of people who would never call themselves technical. What separates him from the enormous crowd of AI commentators is that he runs experiments. He teaches at Wharton, he assigns AI work to students, he collects data, he collaborates on studies with organisations, and he writes about what happened rather than what he expects would happen.
That sounds like a low bar. In this space it is not. The Boston Consulting Group study he was involved in, looking at consultant performance with and without AI assistance on different task types, produced the jagged frontier idea that has since become one of the few genuinely useful concepts in the whole discourse. Some tasks sit inside the capability boundary and AI helps enormously, some sit just outside it and AI actively hurts because the output is confident and wrong, and the boundary is invisible and irregular.
That framing has changed how I think about where to deploy these tools, and I have not seen it stated better anywhere else. The book, Co-Intelligence, sets out four principles for working with AI. Always invite it to the table, be the human in the loop, treat it like a person while remembering it is not, and assume this is the worst AI you will ever use. These are simple to the point of sounding glib, and they hold up under use, which is more than most frameworks manage.
The first one in particular has changed my own default behaviour. Trying the tool on a task before assuming it cannot help turns out to surface capability you would otherwise never find. The newsletter is the stronger product. It is free, it arrives weekly, it responds to what actually happened rather than what was true at manuscript deadline, and the writing is better because it is less constrained.
Posts typically take a new capability or a new study and work through what it means for people doing normal work. When a new model ships, Mollick has usually spent real time with it and has something specific to say rather than a summary of the announcement. Now what it is not. There is nothing technical here at all.
No architecture, no code, no evaluation methodology, no discussion of how any of this works. If you build AI systems, this is not for you, and reading it will feel like watching someone describe a car by how it feels to drive. That is a description of the audience rather than a flaw, and it is worth being clear about before you buy. The book has an unavoidable problem.
It was published in a field that moves in months, and by the time it reached shelves parts of it were already behind. The principles survive. The specific examples and capability claims date badly. Read the newsletter for current, the book for framing.
The metaphors get overworked. Centaur and cyborg as two modes of human AI collaboration is a nice image that carries about one chapter of weight and gets stretched considerably further. Treating AI like a person while knowing it is not is a useful practical heuristic dressed up as something more profound than it is. On evidence, I want to be fair but firm.
Classroom experiments with a few dozen students are suggestive rather than conclusive, and Mollick occasionally writes with more confidence than the sample supports. He is generally good about caveats, better than almost anyone else with his reach, and the caveats are sometimes several paragraphs behind the headline claim. My four point two is for the most useful writing available for the working professional trying to make sense of this, based on real experimentation, honest about uncertainty most of the time, marked down because the book is the weaker half of the offering, because nothing here helps you build anything, and because the evidence occasionally gets framed more strongly than it can bear. Start with the newsletter.
It costs nothing and it is genuinely good.