Co-Intelligence

In the [YOU] on AI Field Guide

[YOU] on AI asks whether the person doing the amplifying is worth amplifying. Co-intelligence takes that question and turns it operational: given a competent, tireless, occasionally brilliant, occasionally delusional collaborator inside the laptop, what practices allow the partnership to compound rather than corrode? The cycle's beaver metaphor—build dams, maintain them, direct the flow toward life—is a metaphor for co-intelligence: the human is the one who decides what dams to build and whether the resulting pond is good for the ecosystem; the machine is the one who can build faster and at greater scale than any human working alone.

Co-intelligence is also the frame that explains why the BCG study's most consequential finding—the skill leveler effect, in which the worst performers gained the most from AI access—is not straightforwardly good news. A skill leveler compresses the talent distribution by lifting the floor toward the ceiling; it democratizes capability. But it also erodes the premium that expertise commands, since the gap between the seasoned professional and the well-equipped novice has narrowed. Co-intelligence as a practice is the response to this compression: deepening the distinctly human contributions that the machine cannot supply—judgment, accountability, the hard private work of figuring out what one actually believes—so that the partnership has something valuable on its human side to compound.

Origin

The concept emerged from Mollick's sustained personal experimentation with AI systems beginning in late 2022 and was developed through the newsletter One Useful Thing before being gathered and extended in Co-Intelligence: Living and Working with AI (2024). The title was a deliberate provocation against both the tool and the rival framings, insisting that neither captured the actual phenomenology of working with these systems—the sense that something like a personality was present, however much the practitioner knew it was an artifact, and the sense that the collaboration produced things neither party contained independently.

The empirical ground for the concept is the BCG study, which provided controlled conditions in which the partnership's effects could be measured. The study confirmed both the productive core of co-intelligence—the enormous gains inside the jagged frontier of AI capability—and its vulnerability: the falling-asleep-at-the-wheel failure mode where human oversight relaxed precisely at the moment the machine moved outside its reliable zone. The study was Mollick's demonstration that co-intelligence is a skill requiring cultivation, not a feature that switches on.

Key Ideas

The inversion of competence. Every prior machine was good at what humans found hard (the precise, the computational, the procedural) and useless at what humans found easy (the intuitive, the creative, the linguistic). These systems invert the pattern: they handle expressive and creative tasks with a fluency that most humans cannot match across the domains they have been trained on, and they stumble at the mechanical tasks that defined previous computing. This inversion scrambles every intuition inherited from the prior era and is the foundational reason why the tool and rival frames misdescribe the encounter.

Centaurs and cyborgs. Two modes of structuring the partnership. The centaur maintains a clean division of labor, assigning whole tasks to whichever party is better suited, keeping the boundary between human and AI contribution deliberate. The cyborg integrates continuously, weaving human and machine effort at the level of the individual sentence, the tight back-and-forth in which the two kinds of mind are inseparable in the output they produce. Each mode has advantages and vulnerabilities: the centaur's boundary is a safeguard against falling asleep at the wheel; the cyborg's integration is the experience that practitioners describe as thinking with a larger mind. The skilled co-intelligence practitioner moves between modes depending on the texture of the problem.

Eyes-open anthropomorphism. Mollick recommends treating the AI as if it were a person because doing so produces better results and a more intuitive working relationship, while insisting in the same breath that it is not a person and that forgetting this is dangerous. The recommendation is not a metaphysical claim but an interface technique: these systems were trained on human language and respond to the conventions of human communication because those conventions are woven into everything they learned. The anthropomorphism is a deliberate, eyes-open instrument, maintained with full awareness that it is a fiction, because the fiction is the most effective interface available to a kind of mind that resembles ours in form and not at all in nature.

The deepest question. Mollick ends Co-Intelligence not with an answer but with a structured uncertainty: four scenarios for where the technology is going, and the observation that the appropriate human response depends on which scenario obtains, and we must choose without knowing. His counsel is to attend to the two middle scenarios—slow continued improvement and continued rapid improvement—because the extremes (the technology has essentially peaked; the technology reaches and surpasses human-level general intelligence) are either trivial to prepare for or beyond preparation. The deepest question is not about the machine but about the human: whether this technology will be used to become more fully oneself or less, to liberate the judgment and creativity the machine cannot touch or to surrender them through disuse.