Co-Intelligence: Living and Working with AI
The real argument in *Co-Intelligence* isn't about whether to use AI — it's about what kind of relationship with it actually works, and why most people are getting that wrong.
Mollick, a Wharton professor who studies how people actually use tools rather than how they theorize about them, brings something rare to AI discourse: he's done the experiments. His starting point is the "jagged frontier" — a term for the maddening inconsistency of what AI handles well versus what trips it up. It can draft a credible legal brief but mangle a simple word count. It can generate novel research directions but confidently fabricate citations. The shape of that frontier matters more than abstract capability claims, because you can't know where the edges are without getting your hands dirty. This is Mollick's deepest argument: the only people equipped to use AI well are people who have already used it badly.
Humans, walking and talking bags of water and trace chemicals that we are, have managed to convince well-organized sand to pretend to think like us
— Mollick, *Co-Intelligence*, p. 193
The four principles he structures the book around — use AI for everything you ethically can, keep a human in the loop, treat the model like a person while knowing it isn't one, and expect rapid improvement — read like common sense until you consider how systematically most organizations violate them. Experiments he cites, including his own field study at Procter & Gamble, show that people working actively alongside AI outperform those who either ignore it or hand off entire workflows and walk away. He names these modes: cyborgs integrate continuously, centaurs divide the work strategically. Both beat the people who treat AI as an autonomous subordinate and trust its output unreservedly. The danger isn't anthropomorphizing AI too much — it's anthropomorphizing it just enough to drop your critical eye while retaining enough credulity to paste in whatever it generates.
If you ask an AI to give you a citation or quote, it is going to generate that quote or citation based on the connections between the data it learned, not retrieve it from memory.
— Mollick, *Co-Intelligence*, p. 95
Where the book strains is in its optimism about the pace of institutional adaptation. Mollick's take on education is sharp: the old apprenticeship model, where junior employees learned by doing tedious work, just broke — AI does that work now, and neither the junior nor the manager benefits from pretending otherwise. He calls this clearly. But his proposed fix — redesign assignments, experiment, adapt — assumes institutional flexibility that most schools and companies don't have. The audience most likely to read this book, curious professionals already experimenting on their own, probably doesn't need convincing. The audiences who most need it, managers who treat AI adoption as IT procurement, educators who frame chatbots primarily as cheating tools — they won't be reached by a book that assumes openness as its baseline.
AI could catalyze interest in the humanities as a sought-after field of study, since the knowledge of the humanities makes AI users uniquely qualified to work with the AI.
— Mollick, *Co-Intelligence*, p. 116
*Co-Intelligence* is best read as a calibration tool rather than a manifesto. If you've been avoiding AI out of anxiety, it removes the excuses. If you've been using it carelessly, it gives you a mental model for why your results are inconsistent. For people already deep in these tools, Mollick's jagged frontier framing earns its keep even when the surrounding chapters feel like a longer version of his newsletter. The book doesn't break new ground so much as lay it out clearly enough that you can finally see where you're standing.
Read the longer summary
The central claim
Mollick’s argument in one line: stop treating large language models as either saviors or threats, and start treating them as a strange new kind of colleague. He calls this framing “co-intelligence,” and insists it is more than a metaphor. LLMs are not tools in the hammer-sense (pick up, use, put down), and not agents in the science-fiction sense (autonomous minds with goals). They occupy a weirder middle zone: systems you collaborate with, delegate to, argue with, and sometimes catch in a lie.
The book’s real target is the reader who has not actually used AI seriously yet. Mollick is trying to drag that reader into the chair. His thesis is almost pedagogical — you cannot understand what this technology is, or what it is bad at, until you have asked it to do your actual job and watched what happens. Everything else in the book services that claim.
Four rules that do most of the work
The durable part of Co-Intelligence is the four rules Mollick gives for working with these systems. They read like common sense until you notice how many people — including smart ones — still don’t follow them.
Rule one: bring AI into everything you legally and ethically can. Not because it will be good at everything, but because you won’t learn its edges from reading about it. Mollick’s own sonnet-vs-fifty-words example captures this well. These models can pull off something genuinely hard (a competent sonnet) while failing at something trivially easy (counting to fifty words). You do not get that intuition from a tutorial.
Rule two: be the human in the loop. This sounds obvious and is almost never obvious in practice, because the tempting move is to paste output straight into the thing you’re shipping. Mollick is particularly good on the taxonomy here — people tend to slot into one of three modes. “Cyborgs” stay in continuous back-and-forth with the model, blending their work with its. “Centaurs” split the task up: human does this piece, AI does that one. “Self-automators” hand over the whole thing and walk away. The self-automators, unsurprisingly, learn the least and ship the worst work.
Rule three: treat the model like a person, but tell it which person. Context matters enormously. “Act as an expert marketer reviewing this pitch” consistently beats “review this pitch.” Mollick flags the obvious trap — once you start thinking of the system as a person, you start trusting it the way you trust people, which is fatal because it hallucinates fluently. The rule is a useful lie. Useful because it produces better prompts. A lie because the model isn’t thinking the way you are.
Rule four: assume the model you’re using now is the worst one you’ll ever touch. Mollick puts the doubling cadence for capability at somewhere between five and nine months. Even if that’s optimistic by a factor of two, it still means any workflow you build around what the model cannot do today is a workflow with a short shelf life.
These rules are not profound. They are the kind of thing that, once you’ve used AI for six months, you have figured out on your own. Their value is as a shortcut for people at month zero.
The jagged frontier
The concept that has traveled furthest from this book — beyond the book itself — is what Mollick calls the “jagged frontier.” The idea is that AI’s capability profile is lumpy in ways that don’t map to human intuition. A task that seems hard to us (generating a plausible research synthesis in a field you don’t know) is sometimes easy for the model. A task that seems trivial (arithmetic, counting words, citing a real paper) is sometimes catastrophically hard.
This matters because our mental model of what computers are good at was built during the era of spreadsheets and search. That model says: computers are precise and fast, humans are creative and fuzzy. LLMs invert it. They are creative and fuzzy. They will invent a citation that looks exactly like a real one because invention is what they are structurally good at.
Mollick’s framing is useful because it gives you a reason to experiment instead of generalize. You cannot predict from first principles whether a given model will be good at the specific task in front of you. You have to try it and watch. And the frontier moves with each release, so the map you drew last year is already wrong.
Centaurs, cyborgs, and why management suddenly matters
The most interesting throughline in the book — and it is underweight relative to how much it should matter — is that working well with AI is mostly a management problem, not a prompting problem.
Mollick makes this explicit in interviews and hints at it in the book. Once you have agentic systems that can run for hours on subtasks, the question stops being “what did the model say?” and starts being “how did I structure the work?” You are writing briefs. You are deciding what to check and what to trust. You are delegating, receiving output, giving feedback. Those are the moves of a manager, not a typist.
This reframing is more radical than it looks. It says that the skills that get undervalued in individual-contributor-heavy engineering cultures — writing a clear spec, defining what “done” looks like, knowing when to inspect and when to let go — are precisely the skills that will separate people who are productive with AI from people who produce AI slop.
Mollick extends this to the org level with a framework he calls leadership-lab-crowd. The “crowd” is the employees, who need model access and permission to experiment. The “lab” is a dedicated team working on what AI can do for this specific company. The “leadership” sets the vision and aligns incentives so people are not afraid to automate themselves out of their own jobs. Most companies have the crowd but skip the lab, so good ideas never diffuse. That diagnosis rings true even two years after publication.
Where the book is strongest
Co-Intelligence is best when Mollick writes as an educator rather than as a futurist. The chapters on teaching are the liveliest in the book, and for good reason: he has been watching what happens when ChatGPT meets a classroom in real time, and he has changed his own syllabus in response.
The “homework apocalypse” — the collapse of the take-home essay as a reliable assessment tool — gets the right treatment. Mollick does not wring his hands. He points out that the take-home essay was a questionable signal even before LLMs, and suggests obvious fixes: more in-class writing, more oral defense, more assignments that are so ambitious the student needs AI to attempt them at all. His revised prompt to his entrepreneurship students — plan something impossibly ambitious, assume you’ll use AI, I won’t penalize you for failing — is the kind of move you wish more faculty would make.
He also resurfaces Benjamin Bloom’s old two-sigma finding: individual tutoring makes the average student outperform roughly 98 percent of classroom-taught peers. We never scaled tutoring because it was too expensive. LLMs make one-to-one tutoring effectively free at the margin. The implication — that schools which ban these tools are shutting students out of the single most effective intervention we have ever measured — is large, and Mollick lets it land.
Credit also for honesty about his own process. He shows AI-generated passages he considered and did not use. He credits his wife Lilach Mollick with building many of the prompts the book leans on. It is a small thing but it reads as adult, especially compared to the genre around it.
Where it falls short
The honest version of this review has to say: the book aged faster than Mollick could have realistically planned for. Written in 2023, landed April 2024. By now, the specific model behaviors he describes — ChatGPT’s quirks, Bing’s unhinged moments, GPT-4’s patterns of failure on basic math — are either fixed, worsened in new ways, or replaced by tool-using systems that do not exhibit the old limits at all. Agentic systems, which Mollick only gestures at, are now the main game. The book does not know this.
Second, Mollick’s prose is aggressively readable, which is both a strength and a limitation. He opts for accessibility over depth almost every time that trade is available. The chapter on alignment gives you one paragraph on the paperclip thought experiment and moves on. If you wanted to actually understand why researchers are worried, or not worried, about goal-directed behavior in these systems, you will leave the book underserved. Co-Intelligence is not the place to work through the doom-vs-boom debate. Mollick does not pretend otherwise, but it is worth saying clearly.
Third, and this is the criticism I suspect will bother his fans, the book’s pro-AI optimism occasionally crosses into a kind of gentle boosterism that skips past hard questions. When Mollick talks about AI “smoothing” the capability difference between technical and business employees, he presents it as mostly good news. But smoothing can also mean: the distinct expertise you spent ten years building is now approximated by a tool your new hire can use on day one. That is a bigger deal than a smoothed curve, and the book does not really sit with it.
Fourth, he is notably skeptical in interviews of “adaptability” as an educational pitch — he calls it a catch-all for “don’t worry, it’ll be fine” — but the book itself sometimes leans on the same comforting vagueness it critiques elsewhere. The “we will figure this out” move appears often enough to notice.
And finally, you will still need a second book after this one. You will need a better grounding in how these systems fail in ways that matter in production. Mollick’s explanation of hallucinations as a structural consequence of pattern-based storage is correct as far as it goes, but does not get you to the practical question of which hallucinations are most dangerous or how to design workflows that catch them before they ship. The Deloitte Australia episode — fabricated court cases and invented quotes in a government report — is exactly the kind of failure mode that deserves a chapter and does not get one. You will need more on agents, since Mollick wrote before agentic workflows were mainstream and his advice assumes the old chatbot pattern. And you will need the economics. Mollick flags that the entry-level apprenticeship model has broken because middle managers would rather hand work to the model than to a junior. That is a significant claim. The book acknowledges it and moves on, when a responsible treatment would spend more time there.
Who should read it
If you have never seriously used an LLM for a non-trivial task, start here. Mollick is the right guide — patient, practical, unembarrassed about the parts that still feel like magic. By the end you will have a reasonable map of what these systems can and can’t do, a workable framework for thinking about where to deploy them, and four rules that save you from the most common beginner mistakes. A few days of reading will compress maybe three months of trial-and-error.
If you already use AI daily, skip most of the book and read the education chapter. The rest will feel like refreshing someone else’s LinkedIn feed. The teaching material is where Mollick has the deepest experience and the sharpest takes.
If you are a manager or team lead trying to figure out how to roll AI out across an organization, the leadership-lab-crowd framing is worth the price of the book by itself. You will want to supplement it with something more current on agentic systems, but the underlying diagnosis — that incentive design and dedicated experimentation matter more than tool selection — is durable.
If you are looking for a book that will resolve your anxieties about where all this is going, look elsewhere. Mollick is an optimist, but he is an honest one. He will tell you that nobody knows how good these systems will get, including the people building them. He thinks we should experiment our way forward and figure it out in flight. Whether that is wisdom or wishful thinking depends on what the next few years actually do, and the book, written when it was, cannot tell you which.