Cyborg vs Centaur: how to work with AI

There are three ways people work with AI, and the research says one of them clearly wins. A Centaur keeps a sharp division of labor, deciding the strategy and judgment themselves and handing the model only the parts it does well, with firm human control throughout. A Cyborg blends with the AI in a continuous back-and-forth, weaving its output into their own thinking turn by turn. A Self-Automator simply hands over whole tasks and lets the model run. A 2026 study that tracked 244 consultants across roughly 5,000 AI interactions found that the Centaurs, just 14 percent of the group, achieved the highest accuracy and deepened their own expertise, while the Self-Automators became, in the researchers’ phrase, passive conduits who gained neither AI skill nor domain skill. The lesson is not which tool to use but which posture to take, and being a Centaur depends on holding a strong enough First Brain to keep control. Here is the difference, the evidence, and how to be the kind that wins.

Where the terms come from

The Centaur and Cyborg labels come from chess, then crossed into work. After Garry Kasparov lost to Deep Blue, he helped popularize a format sometimes called advanced chess, in which a human and a computer play as a team, and people noticed that the best results came not from the strongest human or the strongest engine alone but from a human who knew exactly when to trust the machine. A Centaur, named for the half-human half-horse, keeps a clean division: the human supplies strategy and judgment, the machine supplies calculation. A Cyborg is more fused, integrating the machine’s suggestions moment to moment.

The terms entered the study of knowledge work through the influential 2023 Harvard and BCG experiment, “Navigating the Jagged Technological Frontier”, which found AI helped on tasks inside its capabilities and hurt on tasks just outside them, where workers trusted it too far. That study framed the Centaur and Cyborg as the two productive ways to divide work with AI, and the 2026 follow-up put numbers on which one holds up.

The 2026 study: three modes, three outcomes

The follow-up is where it gets concrete. Tracking 244 consultants through about 5,000 interactions, the research identified three distinct ways of collaborating with AI, and the outcomes diverged sharply. Cyborgs, around 60 percent, kept up a continuous iterative dialogue with the model and developed genuine new AI-related expertise. Centaurs, about 14 percent, used AI selectively while keeping firm human control, and they achieved the highest accuracy while deepening their domain expertise. Self-Automators, roughly 27 percent, delegated entire workflows and ended up gaining neither, becoming passive conduits between the prompt and the deliverable.

Mode	Share	How they use AI	Outcome
Centaur	~14%	Selective, clear division of labor, firm control	Highest accuracy; deepened domain expertise
Cyborg	~60%	Continuous, integrated back-and-forth	Built new AI fluency; strong results
Self-Automator	~27%	Delegates whole workflows	Passive conduit; gained neither AI nor domain skill

The headline is the gap between the bottom group and the rest. The Self-Automators were not lazy or unintelligent; they simply handed over too much, and the act of delegating the whole task removed both the practice that builds domain skill and the engagement that builds AI skill, the endpoint of cognitive offloading taken to its extreme. They got output without growth, which is the worst long-run position to be in.

The jagged frontier, and why trust is the hard part

The original study’s most useful idea is the shape of AI’s competence. It is not a smooth line where the model is good up to a point and weak beyond it; it is jagged, strong on some tasks and surprisingly weak on others that look almost identical, with no obvious marker for which is which. The danger is that confident, fluent output looks the same whether the task sits inside the frontier or just past it, so a user who cannot independently judge the work cannot tell when they have crossed into the territory where the model quietly fails.

This is why trust, not capability, is the hard part of working with AI. The skill that matters most is not prompting; it is knowing when to believe the answer and when to check it, and that knowing comes only from your own understanding of the domain. A Centaur navigates the jagged frontier by holding enough expertise to sense when a task is on the wrong side of it. A Self-Automator cannot navigate it at all, because they offloaded the very understanding that would warn them, so they accept the fluent wrong answer along with the fluent right ones.

The practical implication is that the value of your own knowledge goes up, not down, as the tools improve. The better the model, the more convincing its mistakes, and the more it takes a genuinely informed human to catch them.

Why Centaurs win

The Centaur advantage is control held at the right level. By deciding the strategy and the judgment themselves and delegating only execution, Centaurs keep doing the cognitively valuable part, which both produces better work and keeps their expertise sharpening rather than fading. Crucially, retaining the judgment is what lets them catch the model when it is confidently wrong, the failure mode the original jagged-frontier study warned about, where trusting AI on tasks just past its competence backfires. A Centaur can tell which side of the frontier a task sits on. A Self-Automator cannot, because they offloaded the understanding that would tell them.

This maps directly onto what the offloading research finds elsewhere. The Self-Automator is the BCG study’s name for the person accumulating the most cognitive debt: maximal delegation, minimal practice, deepening dependence. The Centaur is the disciplined opposite, and the reason they stay accurate over time is the same reason they stay capable, which is that they never stop doing the thinking that matters, the pattern behind whether AI makes us dumber.

How to be a Centaur

The posture is learnable, and it comes down to deciding what you keep. Keep the strategy: define the goal, the constraints, and what a good answer looks like before you involve the model, rather than asking it to figure out what you want. Keep the judgment: read the output critically, check it against what you know, and own the decision about whether it is right. Delegate the execution: drafting, formatting, generating options, summarizing, the parts where the model is strong and the stakes of an error are low or easily caught.

A simple test separates a Centaur move from a Self-Automator one: could you do this task yourself, just slower, and would you be able to tell if the model got it wrong? If yes, delegating it is a Centaur move that saves time without costing skill. If no, you are not delegating, you are abdicating, and you are building dependence on a tool you cannot supervise. The difference is your own understanding, which is exactly what a First Brain provides, the same structure behind keeping the human synthesis that AI cannot replace.

Why this needs a First Brain

Being a Centaur is not a willpower trick; it requires having something to keep control with. The strategy, judgment, and supervision a Centaur retains all run on a connected internal model of the domain, a biological knowledge graph dense enough to set the right goal, recognize a wrong answer, and know which side of the frontier a task sits on. Without that internal structure, you cannot be a Centaur even if you want to be, because you have nothing to direct the AI with and no basis to judge it, so you default to Self-Automator by necessity.

That is First Brain before Second Brain applied to working with AI. The 14 percent who win are not the ones with the best prompts; they are the ones with the most structure, the understanding that lets them keep the valuable half of the work and delegate the rest safely, the deeper argument in cognitive augmentation that starts with your own biology. Building that structure is what makes the Centaur posture available to you, and it is the core of Building Your First Brain, free for the first 1,000 readers.

Key takeaways: Cyborg vs Centaur

People work with AI in three modes, and the evidence favors one. Centaurs keep a clear division of labor, deciding strategy and judgment themselves and delegating only execution; Cyborgs blend continuously with the model; Self-Automators hand over whole tasks. A 2026 study of 244 consultants found Centaurs, just 14 percent, reached the highest accuracy and deepened their expertise, while Self-Automators, around 27 percent, became passive conduits who gained neither AI nor domain skill. The terms come from chess and the 2023 Harvard and BCG jagged-frontier study. Centaurs win because retaining judgment lets them both produce better work and catch the model when it is wrong. Being a Centaur requires a strong First Brain to keep control with. The honest limit: Cyborgs also do well, so the real warning is against full delegation, not against integrating AI closely.

Frequently asked questions

What is the difference between a Cyborg and a Centaur in AI work?

Both keep a human meaningfully in the loop, but they divide the work differently. A Centaur keeps a clear separation, deciding the strategy and judgment themselves and handing the model only execution, with firm control throughout. A Cyborg fuses with the AI in a continuous back-and-forth, weaving its output into their thinking turn by turn. Both can work well; the mode that loses is the Self-Automator, who delegates whole tasks and gains nothing. The terms come from chess and a 2023 Harvard and BCG study, and being a Centaur depends on a strong First Brain to keep control with.

Is it better to be a Cyborg or a Centaur?

The research favors the Centaur for accuracy and skill growth. A 2026 study of 244 consultants found Centaurs achieved the highest accuracy and deepened their domain expertise, while Cyborgs built strong AI fluency and also did well. The clear loser was the Self-Automator, who delegated entire workflows and gained neither AI nor domain skill. So the safest posture is the Centaur, keeping strategy and judgment while delegating execution, but the more important takeaway is to avoid full delegation, which both Centaurs and Cyborgs do.

What is a Self-Automator and why is it the worst mode?

A Self-Automator hands entire tasks to AI and lets the model run, keeping little control or judgment. In the 2026 BCG study, this group, about 27 percent of consultants, became passive conduits who gained neither AI skill nor domain expertise. It is the worst mode because full delegation removes both the practice that builds your own ability and the engagement that builds AI fluency, so you get output without growth and accumulate dependence on a tool you can no longer supervise. It is the endpoint of cognitive offloading.

How do I become a Centaur with AI?

Decide what you keep and what you delegate. Keep the strategy: set the goal, constraints, and standard of a good answer yourself before involving the model. Keep the judgment: read the output critically and own the decision about whether it is right. Delegate the execution: drafting, formatting, generating options. A quick test: could you do the task yourself, just slower, and could you tell if the model got it wrong? If yes, delegating is safe; if no, you are abdicating. The understanding that lets you answer yes is a First Brain, which is what the Build First Brain approach develops.

Where do the terms Cyborg and Centaur come from?

From chess. After Deep Blue, formats pairing a human with an engine showed that the best results came from a human who knew when to trust the machine, and a Centaur, the half-human half-horse, came to mean a clean division of labor between human strategy and machine calculation, while a Cyborg meant a more fused integration. The terms entered the study of knowledge work through the 2023 Harvard and BCG experiment on AI’s jagged frontier, and a 2026 follow-up measured which posture produces the best outcomes.