Co-Intelligence — Living and Working with AI

The Most Useful Frame

There is a concept in Co-Intelligence that clarifies more about AI capabilities than any amount of benchmark testing or breathless reporting: the jagged frontier. Large language models are extraordinarily capable — often superhuman — at some tasks, and surprisingly poor at adjacent ones that seem simpler. The boundary between these capabilities is irregular, unpredictable, and specific. A model that writes sophisticated legal analysis can fail at basic arithmetic. One that generates strikingly creative prose can confidently assert false facts.

This concept, developed by Ethan Mollick with colleagues at Wharton, is the foundation on which the rest of Co-Intelligence is built. Understanding that AI capability follows a jagged frontier — rather than being uniformly excellent or uniformly limited — is the prerequisite for using AI tools intelligently rather than being misled by their apparent confidence.

Who Mollick Is

Mollick is a Wharton professor who became one of the earliest serious academic researchers of AI tools in business contexts. His prior work examined the effects of AI assistance on professional tasks across multiple domains; the consistent finding — that AI raised the performance of lower-performing workers significantly while having smaller effects on the best performers — challenged both utopian and dystopian narratives about AI’s workforce impact.

He writes with the specific authority of someone who has actually used these tools extensively, not merely observed them. This gives Co-Intelligence a quality of practical grounding that most books on AI — whether enthusiastic or alarmed — lack. He is describing experiences he has had, not futures he is imagining.

Working With the Alien

Mollick’s central metaphor for AI is “alien” — not in the sense of hostile or extraterrestrial, but in the sense of genuinely different from human intelligence in ways that require adjustment. AI systems do not think like humans; they do not understand the world like humans; they do not have the situated context, the embodied experience, the genuine comprehension that human cognition rests on. Working with them well requires grasping this difference rather than projecting human-like understanding onto their outputs.

This is a more honest framing than either “AI is just a tool” or “AI is like a new employee.” It acknowledges the genuine weirdness of the technology — the fact that it can produce outputs that are astonishingly good and also wrong in completely undetectable ways — without either catastrophising or dismissing the weirdness.

The Four Principles

Mollick structures the practical advice around several principles, of which the most important is: always be aware of the AI’s hallucination potential and verify accordingly. AI systems generate plausible text, not necessarily accurate text. The confidence of the output is not a signal of its accuracy. A person who trusts AI output without verification is not working with a capable tool — they are outsourcing judgment to something that doesn’t have it.

The second major principle is about the nature of effective prompting: the best AI users are not the most technically expert but the most skilled at specifying what they want, directing the conversation productively, and evaluating what they receive. This is a human skill — clarity of thinking, specificity of requirement, quality of judgment — that becomes more valuable in the AI era, not less.

The third principle concerns the redesign of work: organisations that simply add AI as a layer on top of existing processes will gain less than organisations that rethink the processes themselves in light of what AI makes possible. This requires organisational courage and intelligence that many organisations will find difficult to summon.

What the Book Does Well

Co-Intelligence is unusually balanced on a topic that generates extreme positions. Mollick neither dismisses AI’s transformative potential nor predicts specific timelines for scenarios that are genuinely uncertain. He maintains a quality of epistemic humility about what AI will and will not do while being concrete and specific about what it currently does.

The book is also honest about failure modes. Mollick discusses AI’s tendency to agree with users’ existing positions (the ‘yes-person’ problem), its potential to undermine the skill development of junior professionals who use it too early in their careers, and the specific cognitive traps that make AI-assisted work feel more reliable than it is. These are not minor caveats — they are structural features of the technology that anyone using it seriously needs to understand.

The Limitations

The most significant limitation of Co-Intelligence is endemic to all books on fast-moving technology: it dates. Specific capabilities mentioned in the text have already been superseded; specific limitations the book notes have in some cases been reduced. This is not a criticism of Mollick’s execution — it is a constraint no single publication can overcome.

The business focus of the book is also a choice that leaves other important dimensions underexplored. The social effects of AI — on inequality, on education, on the nature of expertise, on creative work — are present but secondary. Readers who want a broader frame for thinking about AI’s civilisational significance will need to supplement Mollick with other sources.

Why It Matters

Co-Intelligence is the book most likely to improve how a professional actually uses AI in the near term. That is a specific and valuable ambition, and it is achieved. For readers who want to navigate between the hype and the dismissal and find the practical ground where productive AI collaboration actually happens, this is the necessary starting point.

Our rating: 4.5/5 — The most practically useful book on working with AI. Read it before anything else on the subject.