Book review: Co-Intelligence

The fourth of Ethan Mollick’s four rules is “assume this is the worst AI you will ever use”, and reading the book two years after it came out is a strange way to test that, because he was right, and so a fair amount of the book describes models nobody uses any more.

Co-Intelligence: Living and Working with AI was published in April 2024. Mollick teaches at Wharton, and he’s been writing One Useful Thing, his newsletter about using these tools, since well before ChatGPT. The book is short, around 230 pages, in two parts. The first is about what large language models are and how to think about them, ending with the four rules. The second is a set of chapters on roles the AI can play, as a person, a creative, a coworker, a tutor and a coach, and a last chapter on possible futures. I bought it when it came out, read the first chapter, and put it down. I’ve been meaning to go back to it for a while, and a colleague recommending it to our product team last month was the nudge.

What’s dated

The examples, mostly. The book was written in 2023, and its AI is a chatbot. You type, it answers, you copy the answer somewhere. Most of the advice is about how to talk to it, give it a persona, ask for options, push back. That’s still true as far as it goes, but the thing I spend my days with now is an agent that reads my repo, runs my tests and asks for permission, and there’s almost nothing in the book about a model that acts instead of answers. That isn’t a criticism of Mollick, since those barely existed when he was writing. It does mean that a reader picking this up in 2026 for practical advice will find the practical parts a bit quaint, like a 2009 book about the iPhone that’s very excited about the App Store.

I also didn’t enjoy the chapter on treating the AI as a person as much as he seems to want readers to. His argument is pragmatic, that the models respond better when you treat them as a kind of person and tell them what kind, and he’s careful to say they aren’t people. In practice, a lot of the chapter reads like encouragement to anthropomorphize, and I’ve seen what that does to how carefully people check the output. It’s how we got the support agent who’d stopped checking the details in drafts because they’d always been right.

What held up

The jagged frontier. Mollick was one of the authors of the 2023 study with Boston Consulting Group, where 758 consultants were given GPT-4 for a set of realistic tasks. On the tasks that turned out to be inside the model’s abilities, the consultants with AI finished about 12% more of them, about 25% faster, and with quality rated more than 40% higher. On a task deliberately chosen to be just outside, they were 19 percentage points less likely to get it right than the ones without it. The frontier between those is jagged, which is his word, and you can’t see where it is from the outside. Something that looks hard is easy for the model and something that looks easy is a disaster.

That’s the single most useful idea in the book, and it’s aged better than anything else because the frontier has moved without getting any smoother. Our agents will do a twenty-file refactor of a .NET service correctly and then confidently invent a configuration flag in the same afternoon. The only way I know of finding the edge is the one he recommends, which is using the tools on your real work, a lot, and keeping notes. My text file of model questions, which I’ve kept since 2021, is basically a hand-drawn map of the frontier, and it needs redrawing every time a model changes.

The centaurs and cyborgs distinction from the same study holds up too. Centaurs split the work, handing whole tasks to the AI and keeping others. Cyborgs weave it through everything they do. I’ve watched both happen on my team, and the cyborg approach is clearly where the team has gone, with the risk that comes with it, which is that nobody can say any more which parts were whose.

And his second rule, be the human in the loop, is the whole of how I work now, though I’d put it more strongly than he does. He frames it partly as keeping your skills sharp and staying in charge. For me it’s that the review is the job, and the book has the right instinct about why that matters, even if the tools it describes are older.

Should you read it?

If you’re an engineer who uses agents every day, probably not for the practical advice. If you’re trying to explain to the rest of your company how to think about these tools, yes, and it’s what I’ll hand people now. It’s clear and short, and it’s honest about not knowing how things turn out. The introduction says getting to know these tools costs at least three sleepless nights, which I thought was too glib when I first read it. I’d count mine as the week Claude Code came out, in February 2025, and three is about right.