I’ve been a fan of Wharton professor Ethan Mollick ever since I began reading his Substack newsletter shortly after ChatGPT’s public launch in late 2022. When he published Co-Intelligence in 2024, I eagerly bought a copy, read it cover to cover, and reviewed it on this blog.
I joined the preorder list when I learned that he was writing Co-Existence: The Next Phase of AI. When it arrived on my doorstep, I texted a photograph of the book in my hand to a friend who replied, “I guess I know what you’re reading this weekend.”
As I usually do with nonfiction, I looked through the table of contents before beginning. When I saw that Chapter 8 was titled Education, I resisted the temptation to read it first. I’m glad I did. The earlier chapters establish why the questions facing educators are becoming more difficult.
What interested me most was a tension running through Mollick’s argument. AI can perform more of the work that once required substantial human expertise, yet using it effectively requires people with knowledge and sound judgment. How do we help students develop those qualities when AI can complete many of the assignments and entry-level tasks through which people have traditionally learned?
That question stayed with me throughout the book. It also extends beyond education. Employers need to decide how to organize work, individuals need to decide when to rely on AI, and society needs to consider who will exercise authority over increasingly capable systems.
Co-Existence did not disappoint me. Its value lies in the way Mollick connects practical advice about using AI with larger questions about what its development means for human beings. The book’s first part considers our individual relationships with AI. Its second part examines work, education, and society.
Authorship and Responsibility
Mollick begins by explaining how he used AI in writing the book. He wrote the first draft of each chapter himself, then used AI for feedback and other assistance. He has also described that process publicly, including checking the research and citations himself.
I appreciate that explanation. It gives readers a basis for understanding the author’s contribution and the assistance he received. It also illuminates a question that recurs throughout the book: which parts of an activity should we retain because they develop our abilities or express something that matters to us?
For an author, working through a first draft can be part of figuring out what they believe. For a student, struggling with an argument can be part of learning how to construct one. Finishing faster has value, but so does the thinking that occurs along the way.
Understanding the Jagged Frontier
One of Mollick’s most useful concepts is the jagged frontier of AI capability. A system may perform a task that is difficult for a person while failing at something that seems simple. Its capabilities are expanding, but they are not expanding evenly.
Mollick uses illustrations to show how that frontier has moved outward. The practical lesson is easy to understand: A successful result on one assignment does not establish that the same system will perform reliably on another assignment. Users have to learn where the boundaries fall in their own work.
The scale of adoption makes those questions consequential. More than two billion people are estimated to use AI weekly, though the exact figure depends on which products and users are included.
A September 2025 study of ChatGPT’s consumer use found that practical guidance, seeking information, and writing together accounted for nearly 80 percent of conversations. These are familiar activities that reach well beyond software development or other specialized technical work.
Scale also has physical costs. The International Energy Agency’s 2025 base case projected that electricity demand from data centers would more than double by 2030, with AI a major driver. Those figures include uses beyond AI, which matters when forecasts about AI are used to discuss electricity prices, infrastructure, and environmental consequences.
From Prompting to Delegation
Mollick describes a shift in how we work with AI. The early generative AI chatbots encouraged people to experiment with prompts. More capable reasoning models and agents have expanded the possibilities for delegating tasks.
Reasoning models are large language models trained to devote additional computational effort to a problem before answering. That can improve performance on demanding tasks, although it can also take more time. Agents enable faster pursuit of a goal through a sequence of actions, sometimes using tools to search the web, operate software, or run code.
Those developments do not remove the need to check results. More capable models and access to outside sources can improve accuracy, but hallucinations remain a concern. OpenAI’s own evaluations have shown that improved reasoning capability does not guarantee fewer factual errors across every task.
Mollick’s practical advice places increasing emphasis on providing context. What is the task? What information does the AI need? What would a successful result look like? What constraints must it observe? These questions should sound familiar to anyone who has supervised other people.
He describes three ways to work with AI:
Centaur work—a person maintains distinct boundaries between human and AI contributions.
Cyborg work—those contributions are woven together as the task develops. Both retain substantial human involvement.
Conjurer work—involves assigning tasks to AI agents and evaluating what they return. The human contribution moves toward defining the work, deciding what to delegate, and judging the outcome.
A useful way to evaluate delegation is to consider how long the task would take you, how likely the AI is to succeed, and how much time you will need to review its work. A fast answer is less valuable when checking it takes longer than doing the assignment yourself.
This is one of the book’s more practical observations. Access to powerful tools increases what an agent can attempt. The person delegating the work still needs a reasonable way to decide whether the result is acceptable.
Why Human Knowledge Still Matters
Mollick argues that AI can reduce the value of expertise used to produce outputs that are clearly specified and readily evaluated. That is a narrower claim than saying that expertise no longer matters, and the distinction is essential to understanding the rest of his argument.
He identifies four human advantages: deep knowledge, wide knowledge, taste, and agency. Deep knowledge supports informed intuition within a field. Wide knowledge helps us recognize relevant ideas outside it. Taste is our capacity to judge the quality and suitability of what we encounter. Agency is the willingness to act, experiment, and pursue a goal without waiting for instructions.
These advantages reinforce one another. A person may know enough to ask a useful question but lack the knowledge to recognize a flawed answer. Someone else may understand a field well but fail to consider how another discipline could help solve the problem.
For educators, I think the implication is substantial. AI may reduce the value of performing a particular task unaided while increasing the importance of knowing whether its result is correct, appropriate, and useful. Students cannot acquire that judgment simply by learning how to request an answer.
I also see a connection to the purposes of a strong college education. Disciplinary depth, intellectual breadth, and the development of judgment fit comfortably within that purpose. Initiative matters as well. Colleges should be able to explain how their programs develop these qualities and how they know students have acquired them.
Mollick is equally concerned about preserving our abilities. Some activities deserve our effort because we need to learn, because the struggle matters, or because doing the work ourselves communicates something important. An AI-generated expression of concern, for example, does not necessarily carry the meaning of a message someone took the time to compose.
I agree with his emphasis on deliberate choices. Using AI for every available task can deprive us of practice. Refusing useful assistance can consume time without improving either our understanding or the result. The purpose of the activity should guide the decision.
His recommendations about when to use AI are useful for the same reason. He suggests looking for patterns of work in which assistance adds value: generating alternatives when we need quantity to find quality, transforming material we already have, considering a different perspective, or helping us move past an obstacle. An expert who can evaluate an answer quickly is in a different position from a beginner who cannot recognize its weaknesses.
This is also where breadth can help. AI may introduce an unfamiliar method or connect ideas from different fields. The opportunity is to investigate those possibilities, while recognizing when we need someone with deeper knowledge to assess them. Work that stretches beyond our expertise can be valuable, but it also demands a plan to check the results.
Organizations Need More Than Individual Productivity
The chapter on work begins with a paradox: People can experience substantial changes in their own productivity while the organization around them continues to operate much as it did before. Faster individual work can disappear into an unchanged approval process or another bottleneck.
Mollick identifies three contributors to organizational change: Leadership, the Crowd (also known as Employees), and the Lab. Leaders establish a direction. Employees discover useful applications in their daily work. A central group tests promising approaches, develops them, and makes them available across the organization.
The relationship among those groups is important. Employees have little reason to share a useful workflow if they believe the reward will be the elimination of their jobs. Mollick suggests that leaders consider ways to help employees benefit from the improvements they identify, whether through sharing savings, redesigning roles, or providing credible assurances about how discoveries will be used.
That is an important management issue. Buying access to an AI product does not establish trust or determine how its benefits will be distributed. Leaders need to explain the purpose of the changes they seek and follow through on their commitments.
Mollick also describes more radical approaches. StrongDM’s three-person AI team developed a Software Factory in which specifications and scenarios guide agents that write code without human code review. Its published manifesto urges spending at least $1,000 in tokens per human engineer per day. That is a guideline about how aggressively to use AI, rather than an audited statement of daily spending.
Mollick presents the effort as a bet that a small team using agents could produce work comparable to that of 15 or 20 people. The comparison is striking, but its meaning depends on which work is being compared and how the results are evaluated.
An example closer to my own interests is Mollick’s account of asking Claude Code to create a book. He describes the system developing the content and cover, creating a website, and arranging how the book could be produced and sold, with human authorization needed for a payment account. The example raises questions about the kinds of publishing and business decisions people may increasingly delegate.
These examples deserve attention, but leaders still need to test whether a proposed approach works in their own circumstances. Producing an output, meeting a professional standard, and taking responsibility for its use are different parts of the job.
Learning With AI While Developing Independent Competence
Given that Mollick teaches at the Wharton School, I was surprised by the relative brevity of the education chapter. As a doctoral alumnus of the University of Pennsylvania and an educator, I wanted to read more about his experiences and thoughts. The chapter focuses on learning and raises several questions that institutions need to pursue further.
Mollick criticizes the early emphasis on detecting AI use in assignments. He points to other ways of evaluating students, including in-class work, live discussion, and proctored assessments. He asks educators to consider how AI can improve learning for the people involved.
His discussion of AI tutoring is particularly interesting. A system designed to help a student learn should behave differently from an assistant designed to finish a task. Providing the answer immediately may satisfy a request while undermining its educational purpose.
This helps resolve an apparent tension in the book’s advice about avoiding AI when we need to learn. Asking AI to complete an assignment can remove the intellectual work that the assignment was designed to develop. Asking for a hint, an explanation, or another practice problem can support that work.
The prospect of broadly available tutoring also raises questions about access and implementation. I would want to know whether a proposed system works for students with varying levels of preparation, how instructors will use the insights it provides about their learning, and what support students need to use it effectively. Making a tool available is the beginning of that educational work.
For educators, I would make the student’s subsequent performance central to the evaluation. Can the student explain the reasoning? Apply it to a different problem? Recognize when a proposed answer is wrong? An impressive submission does not answer all of those questions.
Mollick’s recommendations about what to teach deserve particular attention. He argues for more factual knowledge, broader learning, and practice exercising judgment under uncertainty. He also emphasizes helping students make meaning and keep human purposes central to their work.
These recommendations support a demanding curriculum. I would resist treating access to AI as grounds for expecting less of students.
I would like to see colleges assess both independent competence and the ability to work effectively with AI. One assignment might require students to develop and defend an analysis on their own. Another could require them to use AI, document their choices, identify weaknesses in the output, and explain the final decisions they made.
That would give faculty more useful evidence than the finished product alone. It would also help students describe what they contribute when they enter a workplace where AI is already available.
Who Will Prepare the Next Generation of Experts?
The connection between education and entry-level work may be the book’s most consequential issue for college leaders. Mollick worries that AI will absorb tasks through which beginners have traditionally developed the expertise needed for more advanced work.
He draws on Matt Beane’s research, described in The Skill Code, which emphasizes challenge, complexity, and connection. Beginners need difficult work that stretches them, exposure to the larger system in which that work occurs, and relationships with more experienced practitioners.
AI can threaten those opportunities if experienced workers complete tasks with a system that previously required them to involve a junior colleague. The organization may gain efficiency while a beginner loses access to practice, context, or mentoring. The long-term consequence deserves consideration alongside the immediate savings.
I have explored this concern in my writing about the first rung of the career ladder and in my work on The Community Premium, a book about the value of college communities in the AI era. If we expect graduates to arrive with more developed capabilities, we need to examine how they will acquire them.
I would place responsibility on both colleges and employers. Colleges should create more opportunities for students to apply knowledge to authentic problems, receive feedback, and demonstrate competence. Employers should recognize that eliminating opportunities for beginners can undermine the development of their future experienced workforce.
A degree cannot supply every form of judgment that develops through supervised professional practice. Colleges can strengthen preparation through projects, internships, simulations, and mentoring, but employers still have a role in helping new workers understand the circumstances in which their knowledge must be applied.
This is where I would have welcomed a more fulsome discussion of institutional resources. More supervised practice, stronger mentoring, and closer employer partnerships require decisions about faculty time and money. If colleges are expected to provide more of the preparation that employers once supplied, who will pay for it?
For private colleges considering how to respond to AI, that question belongs alongside curriculum design. Institutions need to identify what they can improve, what they should stop doing, and which partners can help. Mollick provides a useful framework for thinking about the problem. Translating it into an educational program remains our responsibility.
Society and the Question of Control
Mollick’s discussion of society moves beyond the workplace and classroom. He considers a world in which AI systems increasingly communicate with other systems on our behalf, influence the information we encounter, and become companions or advisers.
The growth of AI-generated material is one part of that change. A 2026 Graphite study estimated that roughly half of newly published English-language articles in its web sample were primarily AI-generated. That estimate depends on the sample and detection methods. It does not establish that half of everything people read online is generated by AI.
Mollick also discusses research suggesting that conversations with AI can influence political preferences. The possibility deserves attention, although we should distinguish a change in someone’s stated views during an experiment from evidence of a change in actual voting behavior.
Companionship raises a different concern. Mollick suggests that AI interaction can supplement a person’s social life but can become troubling when it substitutes for human relationships. The distinction is important even as questions remain about which users are most affected and under what circumstances.
For me, those questions reinforce the importance of human connection. Access to a responsive system does not remove the value of relationships in which other people know us, challenge us, and share responsibility for what happens.
The book also considers how little we understand about the internal operation of these systems. Anthropic’s research on introspection found evidence of limited and unreliable abilities in the models tested. The researchers explicitly cautioned against equating those findings with human introspection.
We should take such findings seriously without assuming that familiar human terms describe identical experiences in machines. The practical questions include how systems behave, what they can be trusted to do, and who is responsible when their behavior causes harm.
Mollick asks who controls AI: company leaders, governments, or perhaps no actor with sufficient authority over the whole process. I think the concentration of decision-making deserves sustained attention. As individuals and institutions delegate more work, they also need to consider how they retain the ability to question or reverse decisions made on their behalf.
Planning for an Uncertain Future
Mollick closes with two scenarios, the Escalator and the Eschaton. The Escalator describes ongoing improvement in AI capabilities, with the jagged frontier moving outward and institutions adapting at different rates. The Eschaton considers a more radical transition in which AI’s capabilities fundamentally alter the human role in economic and social life.
The more dramatic possibilities are difficult to evaluate. Improvements in AI-assisted research could accelerate further development, but the pace and consequences remain uncertain. Greater capability and a loss of human control are also separate questions, even when they appear together in a scenario.
I find the Escalator a sufficient reason for institutions to act. Colleges and employers already have decisions to make about training, assessment, and the organization of work. They can begin addressing those decisions without agreeing on a date for artificial general intelligence.
Mollick ends with a warning: “The worst thing you can do is nothing. And the second worst thing is to plan only for today’s AI.”
A Few Final Thoughts
I recommend Co-Existence to readers trying to understand what increasingly capable AI means for their work and institutions. Mollick combines practical advice with questions that deserve more attention than a discussion of which tool to buy or how to write a prompt.
Some of the book’s examples will inevitably age. Its questions about knowledge, judgment, responsibility, and the development of expertise will remain. I essentially read the book twice while preparing this review, and I found the connections among those questions worth the additional time.
I believe AI is the most important technological development of my lifetime. Its capabilities are advancing beyond many of our current practices and, in some cases, beyond what we have imagined. I see an opportunity for education, particularly along the path Mollick calls the Escalator.
That opportunity requires colleges to be explicit about what students must know well enough to do for themselves and what they should learn to do with AI. It also requires us to work with employers to preserve opportunities for beginners to acquire the experience that more advanced work demands.
Mollick’s book provides a useful starting point. The decisions about curriculum, assessment, faculty support, and institutional resources are ours to make. What we do with those decisions will help determine whether our graduates are prepared to influence the future they will enter.