'Co-Intelligence' Book Review: Mollick's Four Rules, Read Against the Study Behind Them

The rules came out in 2024. The experiment behind them is older, and the sequel is newer.


image

In 2023, consultants at Boston Consulting Group who used GPT-4 on a business task built to sit just outside its ability were 19 percentage points less likely to reach the correct answer than consultants working without AI. 

On a different task, inside the tool's ability, the same study found they finished more work, faster, and at higher quality.

That split sits at the heart of the co-intelligence book. 'Co-Intelligence: Living and Working with AI' is a 2024 book by Wharton professor Ethan Mollick that treats generative AI as a collaborator, and Mollick is one of the study's co-authors. This review reads his four rules next to that experiment, then covers his 2026 sequel, 'Co-Existence'.

Is a 2024 book about fast-moving software still a usable guide? The book predates today's AI agents, so this review marks which advice travels and which doesn't. You'll also get a 15-minute test to run on your own work, plus a plain answer on which of Mollick's two books to read first.

📘 Hear a short book summary on your commute on Headway!

image

Quick summary: Five claims in the 'Co-Intelligence' book about artificial intelligence

Mollick's argument comes down to five claims.

  • Mollick treats generative AI as a collaborator that extends human thinking.

  • AI's capabilities are uneven: strong at some hard tasks, shaky at simple ones.

  • Four rules turn that idea into working habits.

  • Human oversight stays necessary because AI can produce plausible falsehoods.

  • AI can act as a tutor and coach, but practice stays with the learner.

The jagged frontier experiment behind the book 'Co-Intelligence,' in plain numbers

The finding in one line: AI's impact flipped from big gains to losses depending on the task, and the consultants couldn't always tell which was which.

Mollick calls that uneven edge the jagged frontier. The jagged frontier is the uneven boundary between tasks AI does well and tasks it gets wrong, and it does not follow human ideas of task difficulty. Inside the frontier, the authors write, AI can complement or even displace human work. Outside it, AI output is less useful and can degrade human performance.

The idea comes from a field experiment run by researchers at Harvard Business School, Wharton, MIT Sloan, Warwick Business School, and BCG. Because Mollick co-wrote it, keep in mind that the author is testing his own idea.

The study sorted 758 consultants, about 7% of BCG's individual contributor consultants, into three groups: no AI, GPT-4, or GPT-4 plus a short prompt-engineering overview. Half of them (385 people) worked on a footwear product task chosen to sit inside GPT-4's ability. The other half (373) got a business case with a spreadsheet and interview notes, built so that GPT-4 would get it wrong.

Task typeWhat participants didResult with GPT-4 vs. without

Inside the frontier

Came up with a new shoe idea and worked through 18 steps, from brainstorming to a press release

12.2% more tasks completed, 25.1% faster, and higher quality scores from human graders

Outside the frontier

Read a spreadsheet and interview notes, then pick which of the three brands the CEO should back

Correct answers fell from 84.5% without AI to about 70% and 60% in the two AI groups

A few details sit behind those rows. People in the lower half of the skill range gained the most from AI on the first task: 43% in the 2023 working paper, against 17% for the upper half. The AI groups also produced less varied ideas than the no-AI group, a point the working paper raises in its Appendix D.

The second task holds the twist. The AI handled the spreadsheet fine but missed a detail buried in the interview notes, according to the HBS AI Institute's summary of the study. Even so, graders rated the AI users' recommendations higher in quality than the no-AI group's, including the recommendations that were wrong (section 3.2 of the working paper). A polished answer can hide a miss.

Read the numbers with their limits in mind. The tool was GPT-4, a general-purpose AI chatbot, in spring 2023, and the participants were BCG consultants, a highly skilled group doing tasks close to their real-world jobs. The figures above come from the 2023 working paper. The study has since been published in Organization Science (2026), and some numbers have changed. The HBS AI Institute's summary of the published version puts the quality gain at about 32% instead of "more than 40%," and the split by skill at 31% and 11%.

'Co-Intelligence': Living and working with AI, turned into four habits

One retailer's description says ChatGPT arrived in November 2022 as a general-purpose AI that could do creative, innovative work that only humans had done before. The publisher says Wharton professor Ethan Mollick immediately understood what ChatGPT meant: humans had developed a kind of co-intelligence that could augment, or even replace, human thinking.

The four rules fit on an index card. What matters is how you use each one, so each rule below gets its own angle.

Rule 1: Invite AI in early, and know when to keep it out

Invite AI to the table early on tasks where a rough first draft costs you little.

Because the frontier is jagged, you can't guess where the tool will shine, so trying it on many tasks is how you find out. This rule suits people who learn by experimenting.

It suits you less if you handle client files, medical notes, or anything under a confidentiality agreement. Keep that data out of public chatbots, and check your employer's policy before adopting AI for work like it. The rule also fits poorly with tasks shaped like the BCG business case, where a fluent answer can hide a missed detail.

Rule 2: Stay in the loop, because fluent answers can be wrong

Be the human in the loop: read every answer as a claim you still have to check.

The study shows the common mistake. In the working paper's retention appendix, most participants kept very high shares of the AI's wording. The most common score was about 0.87 on a scale from 0 to 1, a level you can reach by changing a few words. The authors caution that this can't separate blind copying from good prompting.

On the footwear task, keeping more of the AI's text went with better scores. On the business case, the consultants who did worse tended to adopt the AI's output without questioning it. The authors argue that validating and interrogating AI output plays a critical role when a task sits outside the frontier.

The takeaway: know which side of the frontier you're on, and stay an active participant when you can't tell. Checking answers is a critical thinking habit, and if it feels like a skill you haven't built yet, books that train critical thinking are a good place to practice.

Rule 3: Treat AI like a person, and give it a role

Treat AI like a person you're briefing: give it a role and some context.

The working paper's Appendix E lists "assigning a persona" among the practices of people who worked in close back-and-forth with the AI. One consultant told it to act as a consultant when answering. The paper calls those heavy back-and-forth users cyborgs, and people who split tasks between themselves and the AI centaurs.

Try this five-minute drill. Pick a task you do every week. Run it with no role, then with a role such as an editor checking for clarity, then with a role plus your audience and limits. Compare the three answers to what you'd write yourself, and save the prompt that came closest.

Rule 4: Plan for a better tool next quarter

Assume this is the worst AI you will ever use, and build habits that survive a model update.

Keep your prompts tool-agnostic: describe the task and who it's for, and skip tricks that only work on one of the AI tools. Save one task you care about, along with the answer you'd accept. Re-run it after each major update and write down the date and what changed. After a few rounds, you'll have your own record of where the frontier moved when new tools arrived.

One limit to know: the book came out in 2024, so it doesn't cover today's autonomous agents.

image

Don't want to read the whole book?

Headway turns nonfiction books into 15-minute summaries.

Read a summary

A 15-minute frontier test to run on your own work

The BCG study worked because one task held a planted trap. To solve it, participants had to read the interview notes against the spreadsheet. You can copy that design on a small scale. These are directions only, and this article reports no results from running them.

  1. Pick a task from your own work with an answer you can verify. Skip anything confidential, or strip out the details first.

  2. Plant or find one easy-to-miss detail, such as a number in the notes that contradicts the summary table.

  3. Give the AI the same materials you'd give a coworker, with no hints about the detail.

  4. Check the answer against what you know. Use critical reading habits: ask what each claim rests on and whether the AI caught the detail.

  5. Log the tool and version, the date, what it missed, and what it got right.

  6. Repeat after the next model update.

One run is a hint, not proof. Try a few different tasks before you decide where the tool's frontier sits next to your own abilities.

Practice still has to be yours: The tutor and coach chapters next to deliberate practice

Mollick urges readers to work with AI as a co-worker, co-teacher, and coach. The publisher adds that he assesses AI's profound impact on business and education. That framing is useful for learning, since an AI tutor can answer the same question five different ways without getting tired.

The case for tutors is old. In 1984, Benjamin Bloom reported that the average student tutored one-to-one did better than 98% of the students in a conventional class (Educational Researcher, 1984). The catch was cost, which is why AI tutors draw so much attention.

The study's authors add a warning that applies here. If people stop handing inside-the-frontier work to junior colleagues, training gaps follow, because those tasks are where beginners build skill. Learners face a version of the same problem. If the AI always does the first draft, you skip the part that teaches you.

Deliberate practice is focused practice at the edge of your current ability, with feedback on each attempt. An AI coach can supply the feedback, but it can't do the repetitions for you. Knowing how you learn helps you spot which parts of a task to keep for yourself, and books on metacognition cover that skill well.

Imagine a learner preparing for an English-language job interview. Instead of asking AI to write perfect answers, they respond to common interview questions aloud, then ask the AI coach to point out unclear phrasing, grammar mistakes, and weak examples. They try again, using the feedback to improve each answer. After several rounds, they can explain their experience more clearly without relying on a script. The AI provides guidance, but the progress comes from the learner's own attempts.

'Co-Intelligence' by Ethan Mollick or 'Co-Existence,' one of the new releases: Which to read first

Start with 'Co-Intelligence' if you're new to chatbots, and start with 'Co-Existence' if you already use AI daily and want the follow-up.

Among this fall's new releases, 'Co-Existence' came out on October 20, 2026, from the same publisher, Portfolio. The publisher describes it as a book about how to live alongside machines that are sometimes smarter than us. The audiobook listing adds that Mollick focuses on the practical aspects of how these new tools for thought can transform our world.

One endorsement on the publisher's page, from Angela Duckworth, says the sequel includes practical advice on protecting your family from deepfake scams, the kind of harm bad actors cause.

Mollick also writes the popular 'One Useful Thing,' a Substack newsletter Ethan Mollick writes about AI. This article describes the sequel at the publisher's level, since it's based on the listing and not a full read. Retail listings disagree on page counts and editions, so check the one you're buying.

Reader situationStart withWhy

New to chatbots

'Co-Intelligence'

It introduces the four rules and the jagged frontier from the ground up

Already use AI daily

'Co-Existence'

The publisher presents it as the follow-up on living alongside machines that are sometimes smarter than us

Managing a team

'Co-Intelligence'

The jagged frontier idea comes from a study of working consultants, which maps to judging where your team should use AI

Teaching or studying

'Co-Intelligence'

It frames AI as a co-teacher and coach

Short on time

'Co-Intelligence'

The four rules are the quickest takeaway, and you can test them before you finish the book

If the ethics questions, such as how to align AI with human values, pull you in more than the work habits, an AI ethics reading list is a good next stop after either book.

Personalized recommendations by use case: Five Headway summaries to read next

A short summary is a cheap way to preview a book before you commit to the full read. These take about 12 to 20 minutes in text or audio, and a team of writers and editors prepares them, which is different from AI-generated summaries versus human-written ones. Each pick below matches a different need.

'Thinking, Fast and Slow' by Daniel Kahneman

Read it if Rule 2's checking habit feels hard. Kahneman separates quick, automatic judgments from slow, effortful ones, and a smooth AI answer invites the quick kind. If you like his approach, there are more books like 'Thinking, Fast and Slow' on how people judge.

'Calling Bullshit' by Carl T. Bergstrom and Jevin D. West

This one gives you a method for claims that sound right. Give yourself five minutes per claim: find where the number came from, then see whether it says what the claim says.

'Peak' by Anders Ericsson and Robert Pool

'Peak' is the deliberate practice book from the tutor and coach section above. It suits you if you want to build a skill with AI feedback. Skip it if you only need quick answers, since its payoff comes from sustained practice.

'Rebel Ideas' by Matthew Syed

Watch for a common mistake: a whole team prompting the same tool the same way. Expect the same five answers. The study's finding that AI users produced less varied ideas points in the same direction, and Syed's case for diverse thinking is the counterweight.

'Life 3.0' by Max Tegmark

Pick this up for the big questions in the book's closing chapters about where AI could go and what goals we'd want it to have. If all you want is better work habits, you can skip it.

📘 Check a book's main ideas before you buy it on Headway!

Where the co-intelligence book earns its place, and which chapters to skim

Three questions settle most decisions.

Are you new to chatbots? Read the co-intelligence book from the start. It's a New York Times bestseller written for general readers, so it introduces the four rules and the jagged frontier from scratch.

Do you already use AI daily? Read for the rules and the frontier idea, and skim the tool-specific passages, since they date fastest. The publisher says the book uses dozens of real-time examples of AI in action, and examples like those age first. Rule 4 told you to expect that.

Do you want technical depth or a cultural history of AI? Look elsewhere. The Wall Street Journal's review at launch noted that Mollick teaches management, not computer science. This is a general-audience business book, written by a business school professor.

The publisher's copy is generous. It says 'Co-Intelligence' shows what it means to think and work with smart machines, that 'Co-Intelligence' reveals the promise and power of a new era, and that the book is wide-ranging and hugely thought-provoking. 

It also calls the book the definitive playbook for living and working in the new age of AI, and says Mollick challenges readers to utilize AI's enormous power without losing their identity and to create a better human future. Treat that as a sales pitch. Ethan Mollick's co-intelligence book is a useful general-audience guide, and it doesn't need the hype.

The main reservations are fair ones. The book predates today's agents; it was written in the GPT-4 era, and the BCG study it grows out of used one model and one firm. Mollick co-wrote that study, so weigh it with that in mind. Still, the habit of checking AI work against your own judgment is the part that doesn't expire. After it, 'Co-Existence' is the natural next read.

What Headway readers say about previewing a book first

Two Headway readers describe the preview habit this article recommends. Olena Khodos, a content marketing manager at MacPaw, starts each book with a pre-read and asks, "Is this book actually useful, or is it just a clever title?" Michael Cupps, a data intelligence strategist, calls Headway "the filter that helps me know which books deserve my full attention." Both stories were written in partnership with Headway by real users who received a free gift, so read them as personal experiences, not independent reviews.

Test one rule this week — then find your next read on Headway

Pick one of Mollick's rules and run the 15-minute frontier test on a real task this week. Then choose your next read: 'Co-Intelligence' if you're starting out, or 'Co-Existence' if you use AI daily. If you'd like to preview a related book first, Headway has short summaries on thinking, practice, and AI that take 12 to 20 minutes to read or listen to.

image

📘 Start upgrading your tech literacy today by downloading the Headway app!

FAQs about the 'Co-Intelligence' book

What is 'Co-Intelligence' about?

'Co-Intelligence' is Mollick's guide to working with generative AI as a collaborator. It opens with his reaction to ChatGPT, a section he titles "Three Sleepless Nights." Chapter 1, "Creating alien minds," traces how AI grew into its current form, and Chapter 3 lays out the four rules covered above. The rest of the book applies those rules to work, teaching, and coaching.

Is 'Co-Intelligence' worth reading?

Goodreads readers rate it 3.9 out of 5 across about 18,800 ratings (checked November 16, 2026), a solid score. It fits newcomers best, since it introduces the four rules and the jagged frontier from scratch. Daily AI users get more from the rules and the frontier idea than from the tool examples. The closing section lists the main limits, and the 15-minute test lets you check the advice yourself.

How many pages is 'Co-Intelligence'?

The Penguin Random House listing gives 256 pages for the hardcover, and Porchlight Book Company lists the same count. Other retailers show lower numbers, such as 242 on the Margins reading tracker, because counts shift between hardcover, paperback, Kindle edition, and other ebook formats. Check the edition you plan to buy before you compare lengths.

Is there a 'Co-Intelligence' audiobook?

Yes. Audible lists an audiobook edition of 'Co-Intelligence,' and its page showed a 4.5 out of 5 rating from about 1,230 ratings on November 16, 2026. Many public libraries lend audiobooks through apps, so check yours before you buy. The sequel has an audiobook too, read by Mollick and running 5 hours 25 minutes, according to its Penguin Random House page.

What is 'Co-Existence' about?

The publisher describes 'Co-Existence' as a guide to living with AI that sometimes outperforms people. Its listed promises are: work with AI while keeping your judgment, see where human insight matters most, avoid over-dependence on systems built to persuade, adapt as AI reshapes education and business, and build a lasting human advantage. Those are the publisher's claims, not a review.

What is the best book to learn about artificial intelligence?

No single list settles that, because the right pick depends on your question. Three titles besides Mollick's two take different angles. 'The Coming Wave' by Mustafa Suleyman asks how societies should handle powerful new technology. 'AI 2041' by Kai-Fu Lee and Chen Qiufan sketches ten visions of the future with AI. 'A Brief History of Intelligence' by Max Bennett traces how brains developed and what that suggests for AI.

What is the best book on AI ethics?

There's no single best, because the field splits into angles. Two titles show the range. 'The Alignment Problem' by Brian Christian (W. W. Norton, 2020) looks at machine learning systems that don't do what we expect, such as résumé screeners with hidden gender biases. 'Atlas of AI' by Kate Crawford (Yale University Press, 2021) takes the social side, tracing the minerals, labor, and data that AI runs on.

What is Nick Bostrom's definition of superintelligence?

In 'Superintelligence: Paths, Dangers, Strategies' (Oxford University Press, 2014), Nick Bostrom describes a superintelligence as an intellect that far outperforms the best human minds in nearly every field. He then asks how such a system might arrive, what dangers it could bring, and what strategies could keep people in control, which matches the book's subtitle.


254K+ reviews worldwide

Your growth, your rules

Get the #1 most downloaded book summary app.