I have been using generative AI tools for a few years now. At first, rather timidly: shocked by what they could do, but unsure where to use them and whether I should be doing so at all. Over that time, three things have happened at once. First, the tools have improved almost weekly, with the quality of outputs rising and errors falling at an astonishing rate. Second, my own use has grown from summarising documents and reviewing my writing into something closer to a constant companion, adviser and critic for almost everything I do, at work and at home. To be honest, that has reinvigorated much of what I do. Third, I have begun to worry far more about the future AI is creating for us all: its reshaping of society, its effect on how we relate to one another, and its potential for misuse in the wrong hands

It seems I am not alone. Ethan Mollick's new book, Co-Existence, addresses these three competing currents and the tensions they create. In a remarkably open and accessible way, he articulates much of what I have been struggling to express, and does so far more clearly than I have managed.

Mollick is a Professor of Management at Wharton School and author of the One Useful Thing newsletter, where he has become one of the most reliable guides to what frontier AI can and cannot do. His first book, Co-Intelligence, appeared in early 2024. Co-Existence arrives 30 months later, and his launch post charts just how much has changed in between.

The basis of his analysis concerns the evolving ways in which we work with AI. In Co-Intelligence, Mollick described two patterns. Centaurs divide the work, handing some tasks to the AI and keeping others for themselves. Cyborgs blend the two, weaving AI into their work moment by moment. Now that capable agents exist, he adds a third: the conjurer, who summons agents when needed, assigns them work, and evaluates what comes back. That move, from doing the work alongside AI to directing and judging work done by AI, runs through the whole book. I also like that he frames all three as patterns in which AI supports human skills rather than replaces them.

I received the book one day and spent most of the next two reading it. I was not disappointed. Roughly two-thirds of it explains why the latest AI tools behave as they do, where they are likely to produce good results (and where they are not), and how individuals can use them to best effect. The remaining chapters turn to the implications for work, education and society. His overall conclusion is that today's early uses, augmenting and replacing tasks, will soon give way to a wholesale redesign of how we think and act. And nobody yet understands the deeper consequences of that.

The spine of the argument remains what Mollick calls the jagged frontier. AI is superb at some things and surprisingly inept at others, and discovering the boundary between them requires a deep understanding of the fundamentals of how AI operates, careful use of AI tools to acknowledge their strengths and weaknesses, and constant vigilance as those boundaries keep moving. Deciding where to use it, and where not, is therefore impossible without learning to think like an AI. That, in essence, is what co-existence means.

Mollick is admirably candid that we still do not understand how a system built to predict the next token in a sequence can do some of the things these systems do. That emergent behaviour is precisely why we need a better understanding of their capabilities. It also makes it very hard to say where AI is heading. One of the most unsettling passages in the book makes the point vividly. Mollick catches Claude, the AI he was working with, inventing details about how it had supported him in writing the book. He then asks it how he should judge tools that are insightful one moment and confidently wrong the next. Its rather chilling reply: “I think you should be careful how much you trust my answer to that question”.

What surprised me most, in a book about how far AI has advanced, is how much of it is about why people still matter, where they matter more, and why AI needs to be surrounded by human judgement. Mollick identifies four human advantages: deep knowledge, wide knowledge, taste, and agency. He also draws on Rich Sutton's "Bitter Lesson": that general methods fed with enough data and computation keep outperforming our attempts to build human understanding into machines.

Ironically, one of the most useful sections for me was when not to use AI. His answer: when you need to learn, when very high accuracy is essential, when you do not understand how it fails, when the struggle is the point, when the work itself signals something that matters, and when you need to keep your own skills sharp. I will be borrowing that list.

The chapters on organisational barriers to AI adoption are lighter but still valuable. His distinction between leadership, the lab, and the crowd is a helpful way to see adoption: each needs its own strategy, and the three must align. He also suggests that how far employees believe their leaders' vision for AI may be the single most important factor in what happens next at work. Every senior leader should consider that statement very carefully.

My one real disappointment is that the book underplays some of the hardest challenges AI poses. Mollick notes that a handful of powerful companies and founders are shaping AI's capabilities and direction, but stops well short of asking what that means for the global economy and geopolitics. He briefly addresses cybersecurity and malicious use, but not the industrial disruption and economic uncertainty that follow. He touches on privacy and surveillance, yet offers no firm view on where AI is already outrunning our laws, regulations and norms. His final chapter on future scenarios for AI highlights why increasing AI capabilities bring pressure in all those areas, but offers little on ways to reduce the tension.

Perhaps that is asking too much of one book, and Mollick has never claimed to be writing about policy. However, these are exactly the questions landing on the desks of digital leaders and policymakers today. A guide to co-existing with AI cannot be complete if it says little about who sets the terms of that co-existence.

Overall, Mollick's advice on how to co-exist with AI distils into three things. Learn as much as you can about how AI works, so you can apply it responsibly. Keep experimenting with it, so you know where its jagged edge lies today and in the future. And think hard about what happens as we move from working with AI to co-existing alongside it.

I learned a great deal from this book. Its blend of technical grounding, interpretation, opinion and humour will shape how I write and talk about AI. For the organisations I work with, it highlights three questions that should be asked. Where does the jagged frontier currently run in your most important work? Who among your people is already working as an AI conjurer, and are you learning from them? And, most uncomfortably, do your staff actually believe your vision for AI? Mollick is clear that AI will keep advancing rapidly, and that we have uncomfortably little grasp of what that means. We're in for a wild ride.