A Review of Co-Existence: The Next Phase of AI

Co-Existence is Ethan Mollick’s follow-up to Co-Intelligence. Co-Intelligence already feels old, but it was published in 2024. AI has improved enormously since then. Mollick’s recent post titled 30 Months on the Exponential (in Charts) on his popular, free Substack, One Useful Thing, discusses that evolution, complete with an otter video, something he has become known for.

The book is excellent. Mollick is an engaging and often funny writer who explains ideas at the reader’s level while still writing as an expert. Much of Co-Intelligence was new to me when I read it, but little was new this time around. Still, the book does a good job of distilling a vast amount of information into understandable overviews. One drawback is that it offers few how-to tips or suggestions. Instead, it gives readers an intellectual framework for thinking about AI in work, education, and society, as of late 2026. Here are a few of the portions of the book that I found most interesting.

Three Ways to Work With AI

Mollick describes three approaches to working with AI. He introduced the first two, centaur work and cyborg work, in Co-Intelligence, and he confirms that both remain widely used. A centaur hands parts of a task to the AI at clear points, whereas a cyborg uses the AI throughout the task. Describing his own writing process, Mollick says that he combines centaur and cyborg approaches. He writes his own drafts, but he uses AI for research, editing, fact-checking, and critiques. This approach resembles my own use of AI in writing.

The third way of working with AI, conjurer work, is new in this book. Here, the human is no longer actively in the loop. Instead, as a conjurer, the human assigns a task to AI agents and reviews the agents’ output. Mollick says that conjurer work requires delegation and prompting skills. He also warns that directing agents is often tiring because one person can monitor only so many open loops.

What Is Left for Humans?

Mollick directly addresses the issue of what work will remain for humans as AI improves. It is an unsettling question. The Bitter Lesson, a concept that Mollick borrows from a 2019 essay by computer scientist Richard Sutton, holds that AI eventually outperforms humans, at least in certain realms of expertise. Given enough data and computation, AI also outperforms the systems that researchers build from human expertise. Moreover, AI reaches those results without relying on encoded expert knowledge, so its methods often look nothing like a human expert’s.

The Bitter Lesson applies to tasks with clearly defined, measurable outputs, such as chess or protein folding. It does not apply, however, when people disagree about what counts as good, or when process matters more than output. Those limits provide the opening for Mollick’s four areas where humans have an advantage: deep knowledge, wide knowledge, taste, and agency.

Two of Mollick’s four advantages come from knowledge. The first, deep knowledge, lets experts quickly evaluate AI output, manage work that AI performs, fix AI’s mistakes, and handle novel situations that challenge AI. Deep knowledge also improves experts’ ability to prompt well, producing better AI outputs.

The second knowledge-based advantage is wide knowledge. Because AI learns from much of humanity’s written output, it knows many frameworks and techniques, but it does not always offer them up unless asked. People with wide knowledge can identify which approaches to ask for.

Mollick calls taste, a person’s standard of judgment, the third human differentiator. AI produces vast amounts of mostly mediocre output, so people must separate good work from slop. Taste tells people “which AI outputs to keep, which to discard, and which to use as raw material for something that the AI would never have generated on its own” (p. 89). Importantly, Mollick explains that people are not simply born with taste. Instead, people develop taste by examining their own reactions to a wide range of others’ work, from the best to the worst.

The final advantage is agency, which activates the other three. Mollick defines agency as a willingness to experiment without waiting for instructions. He acknowledges that agency is not always in the individual’s control but instead can be impacted by organizational and situational factors, such as an employer’s support for AI use. He argues that by cultivating deep knowledge, wide knowledge, taste, and agency, people can use AI to expand their capabilities. But if people instead neglect those advantages, AI may atrophy their existing skills.

AI’s Impact on Work

Mollick sums up AI’s effect on the workplace as: “Everything has changed, but also nothing has changed” (p. 116). More concretely, evidence shows that AI makes individual workers more productive, yet organizations show no widespread productivity growth. Mollick traces part of the gap to “secret cyborgs,” meaning workers who hide AI use because disclosure can bring punishment, devalued contributions, and higher expectations (p. 124). He agrees that workers have good reasons for their cynicism, because employers often use technology to cut jobs or monitor workers.

Mollick does not predict what AI will do to employment. He treats each job as a bundle of tasks. Like many others, he sees junior workers, whose bundles are thin and full of vulnerable tasks, as most at risk. He provides some reassurance from history, because three earlier industrial revolutions ultimately benefited humanity. But those revolutions brought rough transitions, some lasting a worker’s lifetime. Somewhat ominously, Mollick also admits that three revolutions are a “sample size” that is “not a lot to bet civilization on” (pp. 137–38).

AI’s Impact on Education

Mollick largely sets aside the issue of academic integrity, predicting that education “will adjust as needed” (p. 142) through in-class assignments and proctored exams. He asks a more basic question instead: can AI help students learn? To answer it, he first confronts two recent failures of educational technology: One Laptop per Child and MOOCs. Despite that record, Mollick is optimistic, because AI tutors have succeeded in some studies, though not all. He notes that AI can also save teachers a lot of time, which they can then spend on more personal contact with students.

On curriculum, meaning what we should be teaching in response to the rise of generative AI, Mollick favors teaching facts. He argues that information students hold in their memory gives them scaffolding for evaluation. He expects the humanities to matter more, because AI draws on the range of human knowledge. Beyond facts and the humanities, he stresses teaching judgment under uncertainty. Finally, he argues that educators must “help students center their humanness” (p. 149), a goal that sounds good but can be a little fuzzy in practice.

Conclusion

Although Mollick suggests that conjurer work is the most powerful way to use AI, and one of the biggest innovations since his previous book, agentic workflows make me uneasy. Of course, Mollick is simply reporting what is happening, not necessarily advocating for an agentic future. Legal research in particular, like writing, requires judgment and active human engagement. Both the legal and librarian professions should advocate for recognizing and protecting writing and research as core professional skills, unsuited to fully agentic workflows. As I have written before, we need to teach students to see themselves as the human in the legal research loop. Overall, I think this book should appeal to a wide range of readers, and I strongly recommend it.

Leave a Reply

Your email address will not be published. Required fields are marked *

*