Centaur AI vs Cyborg AI: what's the difference?

10th February 2024October 4th 2026Flavien BessetInformation

Centaur and Cyborg describe two ways of working with generative AI. In the Centaur model, humans and AI divide tasks according to their respective strengths, with a clear boundary between what each does. In the Cyborg model, collaboration is continuous: humans and AI work together, across sub-tasks, in a constant back-and-forth. The first model primarily raises the question "who does what?", while the second focuses on "how do we work together?" These two terms originate from a field study conducted in 2023 by researchers from Harvard, MIT, Wharton, and Warwick with the consulting firm BCG, and popularized by Ethan Mollick. A second study by the same team, published in late 2025, added a third profile, the Self-AutomatorThis article, which delegates everything to AI, has shown that each approach produces different results and requires different skills. It defines these approaches, their origins, provides concrete examples, highlights pitfalls to avoid, and offers sources of information.

Article published on February 10, 2024, enriched and sourced on September 27, 2026, updated on October 4, 2026 with the study "Cyborgs, Centaurs and Self-Automators".

Where do the terms Centaur and Cyborg come from?

The image of the centaur originates from chess. In June 1998, in León, Spain, Garry Kasparov faced Veselin Topalov in a new kind of match, which he dubbed "Advanced Chess": each player had a computer during the game. The idea, he later explained, came to him from Moravec's paradox: in chess, as elsewhere, the machine excels where the human is weak, and vice versa. Rather than pitting man against machine, why not have them play as a team? This format became known as "centaur" chess: a horse's body (computing power) surmounted by a human torso (judgment).

In 2005, a "freestyle" tournament open to human-computer teams produced a surprising result: the winner was not a grandmaster equipped with the best software, but a duo of American amateurs piloting three computers in parallel, using a more efficient working method. Kasparov derived a now-famous formula from this:

"A weak human + a machine + a better process prevailed over a powerful computer alone and, even more remarkably, over a strong human + a machine + an inferior process."
Garry Kasparov, “The Chess Master and the Computer”, The New York Review of Books, 2010 (our translation)

In other words, what makes the difference is neither the strength of the player nor that of the software, but the quality of the collaborative process between the two. This is precisely the question posed, twenty-five years later, by the Centaur and Cyborg models applied to generative AI.

The two terms, applied to generative AI, appeared in September 2023 in a Harvard Business School working paper entitled "Navigating the Jagged Technological Frontier" (Dell'Acqua, McFowland, Mollick, Lifshitz-Assaf, Kellogg, Rajendran, Krayer, Candelon, and Lakhani). The researchers observed two distinct ways of working with AI among 758 consultants at the Boston Consulting Group (approximately 7% of its consultants) who were equipped with GPT-4. They called these terms Centaurs and Cyborgs. Ethan Mollick, a professor at Wharton and co-author of the study, then widely disseminated these terms in his blog post "Centaurs and Cyborgs on the Jagged Frontier" and later in his book. Co-Intelligence (2024, translated into French in 2025).

The jagged border, a prerequisite for understanding the two models

The Harvard and BCG study is based on a simple idea: AI capabilities do not form a neat circle, but a "jagged boundary" (jagged frontierMollick compares it to a fortress wall, some towers of which extend far into the countryside while others retreat towards the center. Two tasks that appear to be of equivalent difficulty may be found, one inside, the other outside this boundary, and nothing, on the surface, allows one to guess which is which in advance.

The experiment's figures illustrate both sides of the divide. Across 18 consulting tasks located within the boundary (writing, analysis, creativity, persuasion), consultants assisted by GPT-4 completed an average of 12,2% more tasks, 25,1% faster, and with a quality level over 40% higher than the control group. The gain was particularly dramatic for the lowest performers: +43% for consultants below average, compared to +17% for those above. AI levels the playing field.

On a task chosen to fall outside the scope of the analysis (a strategic recommendation requiring the cross-referencing of financial data and interview transcripts), the result was reversed: consultants assisted by AI were 19 percentage points less likely to produce a correct answer. Without AI, 84% found the right answer; with AI, only between 60 and 70%, because many repeated a plausible but incorrect analysis. The researchers also noted a more subtle effect: the AI-assisted outputs were of higher quality, but more consistent with one another.

The Centaur and Cyborg models are precisely two ways of navigating along this boundary: the Centaur maps it in advance and decides which side to place each task on; the Cyborg tests it continuously, task by task, and corrects the trajectory as it goes.

Centaur AI vs Cyborg AI: Comparison

Criterion	Centaur AI	Cyborg AI
Principle	Humans and AI share the tasks	Humans and AI work in an intertwined way.
Central question	Who does what ?	How do we work together?
Pace of exchanges	Alternating work blocks	Continuous back-and-forth, on the scale of a paragraph or an intermediate decision
Typical example	The AI summarizes 40 pages of documentation, the human writes the recommendation	The human writes, the AI criticizes, the human redirects, the AI proposes a variant
Strength	Clarity of responsibilities, easy control, better accuracy on tasks where AI makes mistakes	Broader exploration, increased productivity on tasks near the frontier
Main risk	Underutilizing AI on tasks it would do well	Losing sight of who is responsible for the outcome, falling asleep at the wheel
Skills developed (2025 study)	In-depth professional expertise (upskilling)	New skills for working with AI (newskilling)
Suitable for	Structured, sensitive, or regulated tasks	Creative, writing, and exploratory tasks

The Centaur model: intelligently distributing the work

In Mollick's definition, Centaurs "practice a strategic division of labor: they alternate between tasks assigned to AI and human tasks, allocating responsibilities according to each person's strengths and capabilities." Like the mythological creature, the Centaur has a clear dividing line between its human and machine parts.

In practice, a manager preparing a quarterly review using the Centaur method entrusts the AI with summarizing activity reports, formatting tables, and creating a first draft of the slides; they retain responsibility for interpreting discrepancies, making decisions, and preparing difficult messages. Each person works within their area of expertise, and then the components are assembled.

The analysis of BCG consultants' conversations, published in 2025, reveals what the Centaurs are actually doing with AI. They are using it for three specific purposes: map the domain (request general information about the sector to better understand the problem), to have a method explained to them which they then apply themselves, and have a draft proofread and improved that they wrote. They deliberately limit the rest so as not to depend on the machine.

Task allocation: entrust each part with what it does best, based on an explicit mapping of what AI does well (and badly) in your business.
Agility: reassign roles when the situation changes, for example when a new model shifts the boundary.
Decision: combine machine-generated analysis and human interpretation, without the former replacing the latter.
Responsibility : to maintain human involvement in sensitive decisions, confidential data and everything that commits the organization.
Increase : use AI to enhance team capabilities rather than to automate tasks without rethinking them.

The Cyborg model: continuous collaboration with AI

Cyborgs, Mollick writes, “blend machine and person by deeply integrating them. They don’t simply delegate tasks: they intertwine their efforts with AI, moving back and forth across the jagged boundary.” The line between human and AI tasks becomes blurred: humans produce, AI completes or critiques, humans redirect, AI proposes a new version. The process is iterative and tightly interwoven.

In practice, a consultant drafting a business proposal using a "cyborg" approach writes the issues paragraph himself, asks the AI to make it more impactful, keeps two out of three formulations, generates three possible angles for the methodology section, selects one and rewrites it, then asks the AI to play the role of a skeptical client to test the argument. In the end, he would be unable to tell which sentences came from him and which came from the machine, but each sentence has been reviewed by him.

The 2025 study identified the actions that characterize Cyborgs, used at every stage of the work, from problem framing to the final deliverable: assigning a role to the AI ("act as a consultant"), breaking the work down into sub-steps, asking it to check its own calculations, contradicting it and asking it to reconsider its answer, adding data along the way, demanding justifications, pointing out its contradictions, and refocusing the analysis on a specific point. In other words, the effective Cyborg doesn't just discuss with the AI: it constantly challenges it.

Integration: Humans and AI contribute to the same process, on the same document, in real time.
Granular collaboration: The alternation can take place on the scale of a sentence, an analysis, or an intermediate decision.
Cognitive enhancement: AI helps to explore more options, process information quickly, and obtain instant contradictions.
Adaptation: The duo adjusts the work as they go, which makes Cyborg mode effective near the frontier, where it is not known in advance whether the AI will be good.
Vigilance: A highly integrated collaboration requires clarifying who is responsible for the final result, and maintaining enough distance to spot a plausible error.

The third profile: the Self-Automator, who delegates everything

In December 2025, part of the 2023 team (Steven Randazzo, Hila Lifshitz, Katherine Kellogg, Fabrizio Dell'Acqua, Ethan Mollick, François Candelon, and Karim Lakhani) published a detailed analysis of 4,975 exchanges between the AI and 244 BCG consultants. The task: to recommend to the CEO of a fictitious company which of its three brands to invest in, based on interviews and financial data, in a report of no more than 500 words. A task of the same type as the one placed outside the company's borders in 2023, and therefore a tricky one for the AI.

The researchers identified the two known profiles and a third:

Cyborgs (60% of consultants): an ongoing dialogue with AI, from framing to final drafting.
Self-Automators (27%): Literally, "those who automate." They condense all the work into one or two queries and reuse the response with few modifications. Their deliverables are quick and polished, but lack depth, like work done for them rather than with them. Researchers believe this proportion likely underestimates the reality across all companies.
Centaurs (14%): targeted use of AI on a few sub-tasks, while maintaining control over reasoning and decision-making.

The results clearly differentiate between the three modes. The Centaurs produced the most accurate recommendations. : on a task where AI easily misleads, maintaining control of the reasoning pays off. Centaurs and Cyborgs wrote more convincing notes than the Self-Automators. And above all, each mode develops different skills: Centaurs deepen their business expertise (upskilling), the Cyborgs acquire new skills for working with AI (newskilling), the Self-Automators make no progress on either one (no skill).

For an organization, the lesson goes beyond day-to-day productivity: an employee who delegates everything to AI delivers quickly, but does not build the expertise the company will need tomorrow, nor the expertise that would allow it to spot the machine's errors.

The common pitfall in both models: falling asleep at the wheel

Regardless of the method chosen, the main risk has a name in the literature: "falling asleep at the wheel." In a previous experiment, Fabrizio Dell'Acqua asked 181 professional recruiters to evaluate 44 CVs for a software engineering position, using AI with varying degrees of reliability (approximately 99%, 85%, or 75% accurate recommendations) or without AI. The counterintuitive result: recruiters using the least reliable AI performed better than those using the most reliable. The former spent more time on each application and verified the recommendations; the latter followed the machine without deliberation. An overly convincing AI lulls the user into complacency.

This trap is compounded by a problem of perception. In a study published by the METR institute in July 2025, sixteen experienced developers, working on their own open-source projects with the best tools available at the time, took an average of 19% longer to complete their tasks when using AI. They expected to save 24% of their time and, once the experiment was over, were convinced they had saved 20%. In February 2026, METR acknowledged that this protocol was no longer valid: more and more developers were refusing to work half the time without AI, which excluded the most advanced users from the study. This result therefore says nothing about today's tools; it does, however, show that the feeling of going faster is not a reliable measure. This is why the second of the four principles that Mollick proposes in Co-Intelligence is "be the human in the loop": regardless of the degree of integration, someone must remain awake.

Centaur or Cyborg: which way of working should you choose?

The two approaches coexist, often within the same day. A sensitive, regulated, or highly structured task benefits from an explicit division of roles, like the Centaur model: everyone knows who did what, each step can be controlled, and the process can be documented. A creative, writing, or exploratory activity is closer to the Cyborg model, with numerous back-and-forth exchanges between the user and the AI. The Self-Automator mode, on the other hand, is only justified for simple tasks, well within the boundaries, where the result can be verified at a glance. Four questions help to decide:

Where is the task located in relation to the border? If you know from experience that the AI succeeds (or fails), Centaur mode is sufficient. If you don't know, Cyborg mode allows you to find out at a lower cost.
What is the cost of a plausible error? The higher the cost, the more a clear boundary, formalized human control, and Centaur-type traceability are required. This method proved to be the most accurate for a tricky task.
Should the result carry your voice? A text, a proposal or a decision that involves your personal judgment is better worked on in Cyborg, in small steps, than in block delegation.
What do you want to learn? To develop expertise in a specific field, Centaur mode helps you progress; to learn how to get the most out of AI, Cyborg mode helps you progress. Self-Automator mode doesn't help you progress in any area.

One final point: the boundary is shifting. Mollick's fourth principle, "assume that the AI you use is the worst you will ever use," encourages us to regularly retest what we had previously deemed unattainable. A Centaur map created in 2024 is likely inaccurate by 2026.

And since 2023? The cybernetic teammate and the delegation to agents

The same research team followed up with an experiment at Procter & Gamble, published in 2025 under the title "The Cybernetic Teammate." Out of 776 professionals working on real product innovation challenges, a single person equipped with AI achieved the performance level of a pair without AI. AI also broke down silos between functions: without it, R&D profiles proposed primarily technical solutions, and sales profiles primarily commercial solutions; with it, everyone produced balanced solutions. Finally, participants equipped with AI reported more positive emotions, as if AI were fulfilling part of the social and motivating role of a teammate. AI is no longer just a tool to be piloted like a Centaur or a Cyborg; it takes on the role of a teammate.

The arrival of agents capable of working independently for several tens of minutes on a single file introduces a dynamic that the 2023 study couldn't capture: delegation with checkpoints. A brief is given, the agent is allowed to complete the steps, and the deliverable is then reviewed. This is an extreme form of task allocation, but it demands cyborg-like vigilance during the review phase, otherwise it risks slipping into self-automation. We detail the variations in our article on... Four ways to delegate work to AI and in the one devoted to Deep Research.

F.A.Q

What is an AI centaur?

An AI centaur is someone who works with AI by explicitly dividing tasks: the AI handles what it does well, while the human retains what requires context, judgment, or responsibility. The term originates from the "centaur" chess game launched by Garry Kasparov in 1998 and was applied to generative AI by Ethan Mollick and his co-authors in 2023.

What is an AI cyborg?

An AI cyborg is a person who continuously integrates AI into their work, with no clear boundary between what they do and what the machine does: they write, get criticism, rewrite, and propose variants, on the scale of a paragraph or an intermediate decision.

What is the difference between a centaur and a cyborg?

The difference lies in the granularity of the collaboration. The centaur alternates tasks in blocks, with a clear division of responsibilities; the cyborg constantly intertwines its efforts with AI. The former is suited to structured or sensitive tasks, the latter to creative and exploratory ones. In both cases, the human remains responsible for the outcome.

What is a Self-Automator?

This is the third profile identified in 2025 by the "Cyborgs, Centaurs and Self-Automators" study: a person who delegates most of the task to AI in one or two queries and then reproduces its response almost verbatim. The work is quick and well-presented, but less accurate and less convincing, and the person develops no new skills. This profile applied to 27% of the 244 BCG consultants studied.

Is it better to be a centaur or a cyborg?

Neither approach is superior in all circumstances. In the 2025 study, centaurs made the most accurate decisions on a task where AI is prone to error, while cyborgs developed the best skills in working with AI. The best approach is to switch between the two approaches depending on the task, and to avoid delegating without reviewing the work in all cases.

Sources

Fabrizio Dell'Acqua, Edward McFowland III, Ethan R. Mollick, Hila Lifshitz-Assaf, Katherine Kellogg, Saran Rajendran, Lisa Krayer, François Candelon and Karim R. Lakhani, “Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality”Harvard Business School Working Paper No. 24-013, September 2023. Abstract available at the Harvard Business School website.
Ethan Mollick, “Centaurs and Cyborgs on the Jagged Frontier”, One Useful Thing, September 16, 2023.
Steven Randazzo, Hila Lifshitz, Katherine C. Kellogg, Fabrizio Dell'Acqua, Ethan Mollick, François Candelon and Karim R. Lakhani, “Cyborgs, Centaurs and Self-Automators: The Three Modes of Human-GenAI Knowledge Work and Their Implications for Skilling and the Future of Expertise”Harvard Business School Working Paper No. 26-036, December 2025. Summaries on the Harvard Business School website and Mit sloan (March 2026).
Ethan Mollick, Co-Intelligence: Living and Working with AIPortfolio/Penguin, 2024. French edition: Co-intelligence: Living and working with AI, translated by Marc Rozenbaum, preface by Luc Julia, Éditions First, March 2025.
Garry Kasparov, “The Chess Master and the Computer”, The New York Review of Books, February 11, 2010 (review by Diego Rasskin-Gutman, Chess MetaphorsMIT Press).
Fabrizio Dell'Acqua, “Falling Asleep at the Wheel: Human/AI Collaboration in a Field Experiment on HR Recruiters”Working document, 2022. Presentation on the Harvard Business School website.
Fabrizio Dell'Acqua, Charles Ayoubi, Hila Lifshitz, Raffaella Sadun, Ethan Mollick, Lilach Mollick, Yi Han, Jeff Goldman, Hari Nair, Stewart Taub and Karim Lakhani, “The Cybernetic Teammate: A Field Experiment on Generative AI Reshaping Teamwork and Expertise”, NBER Working Paper No. 33641, April 2025.
METR, “Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity”July 10, 2025, and its methodological update of the February 24 2026.

To learn more: read our article on when AI transforms knowledge-based professions and our analysis AI is transforming work more than it is replacing us.