Co-Intelligence — key concepts
Wharton professor Ethan Mollick argues that large language models are already general-purpose collaborators, a co-intelligence, rather than single-purpose software, and that treating them like ordinary tools misses both their usefulness and their risks. Because their abilities are uneven and improving quickly, he argues the only reliable way to learn what they're good for is direct, repeated experimentation, guided by a small set of practical habits rather than a fixed manual.
Co-intelligence
Mollick's term for treating an AI system as a general-purpose collaborator rather than a single-purpose piece of software, closer to a new kind of colleague than a calculator or a search engine. Because it can converse, draft, critique, and switch domains fluidly, he argues it needs to be worked with through habits and judgment, not just configured once and left alone.
The jagged frontier
AI capability doesn't map cleanly onto how hard a task looks to a human: some tasks that seem difficult turn out to be easy for the model, and some that seem trivial turn out to be beyond it, with no obvious boundary between the two. Mollick argues this means you cannot reliably predict where AI will help just by reasoning about a task. You have to test it directly.
The four principles for working with AI
Mollick's practical habits for day-to-day use: always invite AI to the table for a given task, stay the human in the loop rather than accept its output unchecked, treat it like a person but tell it what kind of person to be, and assume this is the worst AI you will ever use. They're offered as habits to build, not a technical manual to follow once.
Centaur vs. cyborg working styles
Two ways to divide labor with AI. A centaur style keeps a clean split: the human handles one part of a task, the AI another, with a clear boundary between them. A cyborg style blends the two together at a finer grain, going back and forth within a single piece of work until it's hard to say which parts came from which. Mollick treats both as legitimate, task-dependent choices.
Automation of tasks, not jobs
Because AI's capability is jagged rather than uniform, Mollick argues it is more likely to absorb specific tasks within a role than to replace an entire job outright, at least for now. That reshapes what a job consists of (which duties remain human, which get delegated) faster than job titles, training, or policy tend to track, without necessarily eliminating the role itself.
Answer these before you check.
These are the kind of open questions Recall asks in a real review session — no multiple choice, no re-reading the highlight first. Try answering out loud or on paper, then open the reveal to self-grade.
01What does Mollick mean by the 'jagged frontier' of AI capability, and what follows from it practically?
The jagged frontier describes how AI capability doesn't line up with human intuitions about task difficulty: a model can be excellent at something that looks hard and fail badly at something that looks trivial, with no smooth or predictable boundary between the two. Practically, this means you can't reason your way to knowing what AI will be good for on a given task; you have to actually try it and check the output, which is why Mollick treats hands-on experimentation as the core skill for working with these systems.
02What is the difference between a centaur and a cyborg working style, and why does Mollick treat neither as universally better?
A centaur style keeps a clean division of labor: the human does one part of a task, the AI does another, and the boundary between the two contributions stays clear. A cyborg style interleaves the two much more tightly, going back and forth within the same piece of work until the contributions blend together. Mollick treats the choice as task-dependent rather than a matter of one style being more advanced: some work benefits from a clear handoff, other work benefits from constant back-and-forth.
03Why does Mollick argue that AI is likely to automate tasks rather than entire jobs, at least for now?
Because AI's capability is jagged rather than uniformly strong across a domain, it tends to handle some of the specific duties inside a job well while remaining weak at others, rather than being able to perform the full role end to end. That means the more likely near-term effect is a reshaping of what a given job consists of (certain tasks delegated, others kept) rather than the wholesale replacement of the job itself, and Mollick argues this restructuring is already outpacing how fast job descriptions and training adapt to it.
This is a one-time snapshot. Recall keeps testing you on it.
Import your own highlights from Co-Intelligence (or any book) and Recall generates fresh open questions, grades your free-text answers, and schedules the next review with spaced repetition — so the concepts above don't fade in a month.