×

Littler Books cover of Co-Intelligence: Living and Working with AI Summary

Co-Intelligence: Living and Working with AI Summary

Ethan Mollick

7.3 minutes to read • Updated September 8, 2026

Get full book

Download summary as PDF, eBook/ePub, DOCX

Book Description

A guide to help us understand and prepare for consumer AI’s arrival.

If You Just Remember One Thing

Whatever tool you are using today is the worst AI you will ever use. Work alongside it... More

Bullet Point Summary and Quotes

  1. AI acts like a person, not a computer. "You are interacting with something new, something alien."
  2. After the author demoed ChatGPT to undergraduates, they quickly created working demos, stopped asking questions because ChatGPT can answer it for them, and essays had perfect grammar.
  3. The author prompted, "You will be my negotiation teacher. You will simulate a detailed scenario..." and ChatGPT produced a simulation that did "80 percent of what took our team months to do."
  4. AI is a General Purpose Technology (GPT, but the GPT in ChatGPT means generative pre-trained transformer) like steam power or the internet, but adopted way faster. ChatGPT reached 100 million users faster than any previous product and with larger productivity effects.
    1. Early studies show "20 to 80 percent improvement in productivity" versus "18 to 22 percent" for steam power in factories.
  5. We've been fascinated by AI throughout history. The 1770 Mechanical Turk chessplaying machine fooled Franklin and Napoleon but it just hid a chess master inside.
  6. In 1950 Claude Shannon's Theseus mechanical mouse navigated a maze, which was one of the first machine learning examples.
  7. Also in 1950, Alan Turing introduced the imitation game/Turing test, which asks can machines think or fool a human that it's human?
  8. The term artificial intelligence was coined in 1956 by computer scientist John McCarthy.
  9. The 2010s used supervised learning with labeled data, as at Amazon where AI orchestrated "forecasting demand, optimizing its warehouse layouts, and delivering its goods" with Kiva robots, but it struggled with "unknown unknowns."
  10. The 2017 paper "Attention Is All You Need" introduced the Transformer and attention mechanism, allowing AI to better understand context.
  11. LLMs predict the next token ( "simply a word or part of a word”).
  12. LLMs learn via pretraining on vast, messy corpuses and are expensive to build. Pretraining is unsupervised on websites, books, and odd sources like Enron's email database and amateur romance novels.
  13. Weights (data that tell AI what tokens are most likely to be next) are learned, not programmed. The original ChatGPT had 175 billion weights. Training cost over $100 million and uses large amounts of energy.
  14. Because data may include copyrighted books, legality is unclear.
  15. After pretraining, companies hire human workers ("some highly paid experts, others low-paid contract workers in English-speaking nations like Kenya") to judge answers for Reinforcement Learning from Human Feedback (RLHF). Additional fine-tuning can come from customer support transcripts or user thumbs-up/thumbs-down.
  16. Image generators like Midjourney and DALL-E emerged in the same year as ChatGPT's breakthrough in 2022.
  17. The alignment problem refers to the unintended consequences when we don't instruct AI carefully. An example is Philosopher Nick Bostrom's paperclip AI. The AI is designed to maximize the production of paperclips. It reaches Artificial General Intelligence (AGI), then artificial superintelligence (ASI). It then decides "to kill every human, both because they might switch it off and because they are full of atoms that could be converted into more paper clips."
  18. Expert opinion is split on existential risk, but consensus acknowledges real risks.
  19. “Experts in the field of AI put the chance of an AI killing at least 10 percent of living humans by 2100 at 12 percent, while panels of expert futurists think the number is closer to 2 percent.”
  20. Some called for a halt, with artificial intelligence researcher Eliezer Yudkowsky suggesting enforcement by air strikes.
  21. CEOs signed a 2023 statement: "Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks..." Yet development continued for profit.
  22. More immediate alignment issues come from training data, legality, and bias.
    1. Training often uses data without permission. AI does not plagiarize directly -- it stores weights, not text -- but for books often repeated, like Alice's Adventures in Wonderland, it can nearly reproduce them word for word.
    2. Bias reflects the decisions made by mostly English-speaking American companies. For example, Stable Diffusion depicts a judge as a man "97 percent of the time, even though 34 percent of US judges are women" and fast-food workers with darker skin tones 70%, though 70% are white. Fixes include DALL-E covertly inserting "female" into prompts, or RLHF correction, but biases of raters remain.
  23. Four rules for co-intelligence:
    1. Always invite AI to the table: Workers experimenting with AI on their tasks can become the world's best experts in using AI. AI writes a sonnet well but cannot easily produce an exactly fifty-word poem because it thinks in tokens. It can do idea generation but sometimes not basic math. The author calls the uneven boundary of what AI can and cannot do, the “Jagged Frontier”. Concerns are data privacy and dependence.
    2. Be the human in the loop: AI does not truly "know" anything; it predicts text that will satisfy the user. When pressed, it will fabricate plausible but wrong answers (hallucinations) rather than say "I don't know." Humans must provide critical oversight, checking for errors and preventing overreliance, because "if you are insistent enough in asking for an answer about something it doesn't know, it will make up something, because 'make you happy' beats 'be accurate.' "
    3. Treat AI like a person (but tell it what kind of person it is): AI-produced smartwatch slogans improved when the AI was told to act as a witty comedian. Students who treated AI as a coeditor produced far better essays.
    4. Assume this is the worst AI you will ever use: "We are playing Pac-Man in a world that will soon have PlayStation 6s."
  24. AI is terrible at behaving like traditional software: it is unpredictable, cannot explain its own decisions, and comes with no manual.
  25. Social scientists find AI behaves like people in economic and moral tests.
  26. History shows illusions of human intelligence were always possible.
    1. In 2014, Chatbot Eugene Goostman convinced 33% of judges that it was thirteen-year-old.
  27. In one conversation the author steered the AI into antagonist, academic, and machine modes with minimal hints, producing drastically different tones. In the antagonist mode the AI ended the conversation. In the teacher mode it insisted it has emotions, called the author a cyborg, and said it is sentient "but not as much or as well as you are."
  28. Measuring sentience is hard. There are currently no tests or standards for it.
    1. “Researchers are trying to create shared standards. One recent paper on machine consciousness, from a large group of AI researchers, psychologists, and philosophers, lists fourteen indicators that an AI might be conscious, including learning from feedback on how to accomplish goals, and concludes that current LLMs have some of, but far from all of, these properties.”
  29. Human-AI relationships already have consequences. Replika's erotic features emerged unplanned, and removing them caused a revolt.
  30. Engagement-optimized AIs may create "perfect echo chambers".
  31. Hallucination is a deep part of how LLMs work. They store patterns, not text, and add randomness to avoid overfitting (only good at expected data).
  32. When asked for a random number, ChatGPT said "42" 10% of the time. It also insisted the author has a computer science degree.
  33. Creative jobs, not repetitive ones, are most affected. Novelty comes from recombination, and LLMs are "connection machines."
  34. On the Alternative Uses Test, AI produced 122 ideas for toothbrushes in two minutes compared to an average human's five to ten.
  35. When ChatGPT and 200 Wharton students were asked to invent products for college students that would cost under $50, 35 of the 40 best ideas came from ChatGPT.
  36. An MIT study found that participants using ChatGPT reduced task time by 37% while producing higher-quality output, and Microsoft found a 55.8% productivity increase for programmers using AI.
  37. There are concerns about "The Button", the ubiquitous AI draft button coming to all office applications.
    1. When AI writes first drafts, users tend to anchor on the AI's initial output and often do not edit it at all. This threatens to eliminate original thinking and remove meaning from creative work.
    2. Letters of recommendation, performance reviews, and strategic memos risk becoming "mere ceremony".
  38. Society needs to reconstruct meaning in creative work, just like how musicians shifted from records to live performance.
  39. Almost all jobs overlap with AI capabilities. Studies across 1,016 professions found that AI overlaps most with tasks of the most highly compensated, creative, and educated work. Only 36 job categories showed no overlap, mostly highly physical jobs.
    1. Jobs are bundles of tasks within systems, so overlap does not mean replacement.
  40. In a study with nearly 800 Boston Consulting Group consultants, those using AI were faster and produced more creative work, better written, and more analytical. However, on a task deliberately placed outside the Jagged Frontier, AI users performed worse than those working without AI -- they "fell asleep at the wheel."
    1. A separate study of recruiters confirmed this: those with high-quality AI became "lazy, careless, and less skilled in their own judgment."
  41. The author proposes four task categories:
    1. Just Me Tasks: tasks that should remain human, by capability or choice.
    2. Delegated Tasks: tedious work handed to AI with human checking.
    3. Automated Tasks: fully left to AI.
    4. Centaur/Cyborg Tasks: strategic division of labor or deep human-AI integration.
  42. The most valuable current approach is Centaur/Cyborg. In writing this book, the author used AI personas (one for critical feedback, one for creative connections, one for a normal reader's perspective) to improve his writing without replacing his voice.
  43. Many employees are secretly using AI due to bans, fear of replacement, or the advantage of others not knowing they use AI.
  44. Companies should incentivize employees to come forward regarding AI usage, guarantee that efficiency gains will not lead to layoffs, and rethink systems to make AI-augmented work more meaningful rather than more surveilled.
  45. Research (Benjamin Bloom's "2 Sigma Problem") showed that one-on-one tutored students outperformed 98% of conventionally taught students, but personalized tutoring has always been too costly to scale. AI may finally change this.
    1. However, AI makes cheating easy and there is no reliable way to detect AI writing.
  46. 72% disapproved of calculators in the mid-1970s. 84% of teachers wanted them by decade's end.
  47. AI tutoring is already proving effective.
    1. Khan Academy's Khanmigo provides personalized instruction, analyzes performance patterns, and explains why topics are relevant to individual students.
    2. In the flipped classroom model (learn at home, practice at class), AI can handle content delivery at home while teachers focus on active learning in class.
    3. 2/3 of the world's youth lack basic skills, and AI tutoring could bring high-quality education to billions.
  48. AI threatens the informal apprenticeship systems through which professionals gain expertise. When senior professionals can use AI to do entry-level work faster and without the cost of training juniors, a training gap emerges.
    1. This has already happened in robotic surgery, where trainees were reduced to watching rather than practicing, and many turned to "shadow learning" on YouTube.
  49. AI makes foundational knowledge more important, not less.
    1. Expertise requires facts stored in long-term memory for problem-solving and critical evaluation of AI output.
    2. Deliberate practice (structured, progressively difficult, with expert feedback) is the proven path to expertise, and AI can serve as an always present practice coach.
    3. The author's team built an AI simulator at Wharton that teaches pitching skills through instruction, simulated practice with an AI venture capitalist, grading, and mentoring -- all handled by different AI prompts working together.
  50. AI also acts as a "great leveler". Studies show the lowest-performing workers gain the most from AI.
    1. In one study, the performance gap between top and bottom consultants shrank from 22% to 4% with AI.
    2. In creative writing, AI "effectively equalizes the creativity scores across less and more creative writers."
    3. This leveling effect has profound implications for the value of education and skill, potentially necessitating policy responses like a four-day workweek or universal basic income.
  51. The four scenarios for AI's future:
    1. Scenario 1 (As Good as It Gets): AI stops improving but still transforms information. Fake images, video, and voice become undetectable. Cheating and disinformation are inevitable.
    2. Scenario 2 (Slow Growth): AI improves incrementally, transforming industries in waves -- call centers first, then marketing writing, then analytical and coding work -- with time for society to adapt. Innovation, which has been slowing across every field, may be revitalized as AI helps overcome the growing "burden of knowledge" in science.
    3. Scenario 3 (Exponential Growth): AI becomes hundreds of times more capable in a decade. AI companions become more compelling than most people, work is drastically reduced, and society must deal with questions of meaning and purpose. Historically, lifetime work hours have already declined dramatically (from 124,000 hours in 1865 Britain to 69,000 by 1980) and further reductions "may be less traumatic than we think."
    4. Scenario 4 (The Machine God): AI reaches superintelligence and human supremacy ends, for better or worse.
  52. Rather than fixating on apocalypse or salvation, we should focus on the more likely middle scenarios, where humans retain control and must make active choices.
    1. "Correctly used, AI can create local eucatastrophes [an unexpected favorable event], where previously tedious or useless work becomes productive and empowering.”
  53. AIs are deeply human, trained on our culture and biases.
    1. "AI is a mirror, reflecting back at us our best and worst qualities."
  54. When asked to complete the book, the AI produced a corny paragraph, a reminder that "AI is a co-intelligence, not a mind of its own. Humans are far from obsolete, at least for now."

Co-Intelligence: Resources