Literature review · Part 2 of 2
Centaur and Cyborg Modes of Human-AI Interaction
An evidence assessment of the jagged-frontier framing, its 2025 revision, and the meta-analytic case against it
Literature review · Part 2 of 2
An evidence assessment of the jagged-frontier framing, its 2025 revision, and the meta-analytic case against it
Part 1 · Primer›Part 2 · Literature review and evidence
This review asks one question: is the Centaur and Cyborg framing solid enough to teach, to mentor with, and to use as an advisory frame with executive audiences? The framing survives scrutiny, but only in its revised form and only with its boundary condition attached. Field-experimental evidence supports large productivity gains inside the AI capability frontier and measurable losses outside it. Meta-analytic evidence appears to contradict this, showing that human-AI combinations underperform the better of either alone. The contradiction resolves once task type is held constant. Anyone presenting this material without the 2025 three-mode revision, or without the meta-analytic counterweight, is teaching a superseded and one-sided model.
This review tests that question against the framing’s own primary evidence, against its 2025 revision, and against the strongest published challenge to it.
Two corrections govern the whole review, and both change what a practitioner should say in public.
The taxonomy is not Mollick's alone. It emerged from a nine-author Harvard Business School and Boston Consulting Group field experiment, with Mollick as one contributor[1]. Mollick coined and popularised the labels through his own writing[2][3][31]. Attributing the research to him alone misstates the provenance.
The two-mode model is superseded. A December 2025 paper from the same research group replaced it with a three-mode model and inverted the assumed population split[4]. Teaching the 2023 version in 2026 teaches a model its own authors have moved past.
Generative AI raises knowledge-worker output and quality on tasks that sit inside the model's capability frontier, and lowers accuracy on tasks outside it.
The mode of human-AI interaction, not merely the fact of AI access, determines output quality and the direction of skill development.
Combining a human with AI outperforms the better of human alone or AI alone. This is the strong-synergy claim that most executive presentations of the framing assume.
Sources were retrieved through Perplexity for market and working-paper coverage, and through Scite.ai for peer-reviewed and citation-level evidence. Inclusion required one of four conditions: primary authorship of the framing, a pre-registered field or laboratory experiment on generative AI and work performance, a systematic review or meta-analysis, or foundational theory that the framing explicitly rests on. Commentary, vendor material, and secondary summaries were excluded. Publication status is recorded on every entry in the reference list itself, because several of the most-quoted items have not completed peer review; section 2.6 names the ones that bear on the argument made here.
The originating study ran a pre-registered experiment with 758 Boston Consulting Group consultants using GPT-4[1]. Across 18 realistic tasks selected to sit inside the model's capability range, consultants using AI completed 12.2% more tasks and finished 25.1% faster, with significantly higher assessed quality[1]. On one complex managerial task deliberately chosen to sit outside that range, AI users were 19 percentage points less likely to reach the correct answer[1].
That single reversal carries the whole theoretical contribution. The authors named the boundary the jagged technological frontier, because tasks of apparently equal difficulty fall on opposite sides of it without warning[1]. Users cannot see the boundary, so they cannot calibrate reliance to it. Centaur and Cyborg describe the two working patterns that navigated the boundary successfully[2].
Randazzo and colleagues studied 244 BCG consultants working a single strategic recommendation task and found three modes rather than two[4]. Those 244 are not a separate population: they are the AI-access participants from the outside-the-frontier task in the originating study[1] — the same firm, the same GPT-4 platform and the same brand-investment case — with their conversation logs re-analysed qualitatively[4]. One dataset viewed twice, not two independent findings. The finding matters commercially, because it converts a style taxonomy into a skilling-risk model.
| Mode | Co-creation process | Share (rounded) | Skill trajectory |
|---|---|---|---|
| Centaur | Directed. The human decides both what to do and how, using AI selectively for bounded support tasks. | 14% | Upskilling. Deepens existing domain expertise. |
| Cyborg | Fused. AI is woven through every stage. The human sets the goal, the AI often drives execution, and the two iterate. | 60% | Newskilling. Gains AI-specific capability, holds domain expertise. |
| Self-Automator | Abdicated. The whole task collapses into one or two prompts, with output accepted largely unedited. | 27% | None. Gains neither domain nor AI expertise. |
Centaurs recorded the highest accuracy on the business recommendation[4]. Cyborgs and Centaurs were roughly equal on persuasiveness, and Self-Automators trailed on both[4]. Of the Self-Automator group, 44% accepted AI output with no modification at all, and the remainder made only surface edits[4].
The organisational implication follows directly. The same workflow, described identically in a policy document as human-in-the-loop, is enacted three ways with three different consequences for capability. Governance that counts AI usage cannot see this difference. Only observation of the interaction pattern can.
A pre-registered experiment with 791 Procter & Gamble professionals moved the question from individuals to teams[5]. Individuals working with AI matched the output quality of two-person teams working without it[5]. AI also dissolved functional silos, producing balanced solutions regardless of whether the professional came from R&D or from commercial[5]. Participants reported more positive emotional responses, suggesting the tool absorbs part of the social function of a colleague[5].
An earlier experiment by the same lead author supplies the failure mode. Recruiters given a high-accuracy AI reduced their own effort and disengaged, which the author labelled falling asleep at the wheel[6]. Better AI produced worse human attention. That mechanism is the direct antecedent of the Self-Automator finding.
Vaccaro, Almaatouq and Malone conducted a pre-registered systematic review of 74 papers reporting 106 experiments and 370 effect sizes[7]. On average, human-AI combinations performed significantly worse than the better of human or AI alone, at a pooled effect of g = −0.23[7]. Against the weaker baseline of humans alone, the combination performed clearly better, at g = 0.64[7].
Two moderators drive the result and both matter for practice. Losses concentrated in decision tasks, while gains concentrated in content-creation tasks[7]. When the human outperformed the AI, the combination gained; when the AI outperformed the human, the combination lost[7].
Berger and colleagues then re-analysed all 74 papers in that corpus and identified a design artefact[8]. Most experiments gave participants no trial-by-trial outcome feedback, which prevented them from learning the AI's reliability[8]. Studies that supplied feedback showed higher synergy, and feedback combined with AI explanations tended to produce positive synergy[8]. Explanations without feedback correlated strongly with negative synergy[8]. The pessimistic headline therefore partly measures how the experiments were built rather than what humans and AI can achieve.
Four studies establish the general effect and its distribution across skill levels.
Noy and Zhang ran a randomised writing experiment with mid-level professionals and recorded roughly 40% faster completion with higher assessed quality[11]. Brynjolfsson, Li and Raymond studied customer-support agents in the field and found gains concentrated among novices, with minimal or slightly negative effects for experienced agents[12]. Peng and colleagues found a 55.8% speed improvement among developers using GitHub Copilot on a bounded implementation task, though with a wide interval around it (95% CI 21% to 89%)[13].
Otis and colleagues then produced the result that breaks the tidy story. In a five-month randomised trial with 640 Kenyan entrepreneurs given a GPT-4 business mentor, the authors could not reject the null hypothesis of no average treatment effect on revenues or profits at all[14]. That average concealed a split: low performers did nearly 10% worse because of the assistant, whereas high performers may have benefited by over 15%[14]. This directly contradicts the levelling effect reported at BCG, where low performers gained most[1]. The two findings are not reconcilable by appeal to sample size. They differ in whether participants could evaluate and act on the advice they received.
METR ran a randomised trial with 16 experienced open-source developers across 246 real issues[9]. Developers allowed to use frontier AI tools were 19% slower, yet estimated afterwards that AI had made them roughly 20% faster[9].
METR has since stepped back from that instrument. A February 2026 update covering 57 developers, 143 repositories and more than 800 tasks estimates a change of −18% for the returning developers (95% CI −38% to +9%) and −4% for newly recruited ones (95% CI −15% to +9%)[27]. It also reports that between 30% and 50% of developers declined to submit some tasks because they did not want to work on them without AI, which biases the estimate downward, and that METR is redesigning the experiment for that reason[27]. The 2025 trial stands as published and the finding drawn from it below does not depend on the size of the slowdown, but its authors no longer offer that figure as their current estimate.
The gap between measured and perceived performance is the single most useful finding in this review for an advisory context. Most organisational evidence for AI benefit is self-reported. This trial shows self-report inverting the true sign.
Kosmyna and colleagues at the MIT Media Lab studied 54 participants across four months in three writing conditions, using EEG[10]. The LLM-assisted group showed weaker neural connectivity and lower recall of their own written material, which the authors framed as cognitive debt[10].
Publication status. Eight of the twenty-eight sources have not completed peer review. Seven of those bear on the empirical argument: [4] is a Harvard Business School working paper, [6] a laboratory working paper, [8], [10] and [13] preprints, and [9] and [27] technical reports. Each is methodologically strong and each is widely cited, but none carries journal-level validation. Present them as working evidence, not settled evidence. This list is not the record — every entry in the reference list carries its own status, so the count above can be checked rather than taken on trust. The eighth, [20], is Engelbart's 1962 institutional report, cited in 2.7 for a position rather than for a finding; it is listed here for completeness and carries no evidentiary weight in this review. Item [26] is scheduled for October 2026 and cannot be assessed.
Design limits, which peer-review status does not capture. [4] observed the three modes rather than assigning them, so its mode differences are associations and the obvious rival explanation — that stronger professionals adopt Centaur behaviour — is untested; its skilling results come from interviews after a single task, not from measured skill change over time. [7]'s corpus largely lacked trial-by-trial outcome feedback, which [8] shows accounts for part of the negative result. Neither limitation makes either source weak. Both change what may be claimed from them.
The framing sits on an older split. Licklider argued for a co-equal partnership between person and machine[19]. Engelbart argued for a human-controlled system that amplifies its operator[20]. Cyborg descends from the first tradition and Centaur from the second, which is why the two modes produce different skill outcomes rather than merely different workflows.
Bainbridge established the core irony four decades ago. Automating the routine leaves the human only the hard residue, while removing the practice that builds competence to handle it[21]. Parasuraman and Riley catalogued misuse, disuse and abuse of automation[22]. Lee and See supplied the trust-calibration model that explains why reliance drifts away from reliability[23]. Self-Automator behaviour is a textbook instance of all three.
In management theory, Raisch and Krakowski frame the automation-augmentation paradox that the taxonomy sits inside[16]. Krakowski, Luger and Raisch analysed centaur chess directly and describe interrelated substitution and complementation dynamics, which together render traditional human capabilities obsolete while creating new ones at the human-machine intersection[15]. Reading those dynamics as a trajectory, in which substitution eventually displaces complementarity as machine capability rises, is this review's extension rather than the paper's claim — but it is the strongest theoretical ground for asking whether Centaur mode has a shelf life. Lebovitz, Lifshitz-Assaf and Levina show what engaged and unengaged use look like among radiologists facing opaque tools[17]. Anthony, Bechky and Fayard supply the systems critique of collaboration language itself[18].
The metaphor's origin is freestyle chess rather than any research paper. Cowen popularised it as a labour-market model[24], and Kasparov's account sits in his trade books[25].
| Prop. | Verdict | Reasoning |
|---|---|---|
| P1 | Supported | Supported by a pre-registered field experiment now published in a top-tier peer-reviewed journal, with the negative case measured in the same design rather than assumed[1]. Corroborated across writing[11], support work[12] and code[13]. The boundary condition is load-bearing, not a caveat. |
| P2 | Supported | Strengthened rather than weakened by the revision. Mode predicted accuracy, and predicted whether a professional gained domain expertise, AI expertise, or nothing[4]. Consistent with the disengagement mechanism observed among recruiters[6] and with unengaged use among radiologists[17]. Supported on an observational design: mode was observed, not assigned, so selection effects are untested, and the skilling finding rests on interviews after a single task rather than on measured skill change over time[4]. That is a limit on what may be claimed, not a reason to downgrade the verdict. |
| P3 | Not supported | Rejected as a general claim. Across 370 effect sizes the combination underperformed the better party at g = −0.23[7]. It holds only against the weaker baseline of humans alone, at g = 0.64[7]. Anyone asserting strong synergy in general is asserting something the evidence does not carry. |
The apparent conflict dissolves once task type is held constant, and this reconciliation is the most defensible position a practitioner can hold.
The meta-analysis found losses concentrated in decision tasks and gains concentrated in content creation[7]. The BCG tasks that produced large gains were creative and analytical production tasks[1]. The single BCG task that produced a 19-percentage-point accuracy loss was a complex managerial decision[1]. Both literatures therefore report the same underlying pattern. AI raises output on production work and degrades human judgement on decision work.
The second reconciling factor is relative capability. Gains appeared when the human outperformed the AI, and losses appeared when the AI outperformed the human[7]. That is the same mechanism as the recruiter study, where a more accurate AI produced more human disengagement[6].
The third factor is experimental design. Restoring outcome feedback moves synergy toward positive[8]. Real workplaces supply feedback continuously. Most laboratory studies did not. The meta-analytic pessimism is therefore probably an underestimate of workplace performance, though the size of that correction is not yet established.
One conflict does not resolve. BCG reported that low performers gained most[1], and P&G reported that AI narrowed expertise gaps[5]. Otis and colleagues found the opposite, with high performers gaining and low performers declining[14]. Brynjolfsson and colleagues sit between them, with novice gains and negligible expert gains[12].
The most plausible explanation is the capacity to evaluate advice. Consultants and P&G professionals possessed enough domain knowledge to judge AI output and reject weak suggestions. Low-performing entrepreneurs did not, so poor advice passed through unfiltered. If that explanation holds, levelling is not a property of AI. It is a property of the baseline competence of the population using it. State this as a hypothesis, because no study has yet tested it directly.
The evaluation-capacity explanation in 3.2 is this review's inference from four studies with different populations and tasks. It has not been tested as a hypothesis in any retrieved source. Present it as a proposed reconciliation, not as a finding.
Teach three modes, not two. The Self-Automator category is where organisational risk concentrates, and it accounted for 27% of the sample[4]. A two-mode presentation omits the only group that gains nothing.
Lead every presentation with the boundary condition. The 19-percentage-point accuracy loss outside the frontier[1] is what separates an analyst from an enthusiast. Present the negative result before the positive one.
Cite the journal version. The originating paper completed peer review and appeared in Organization Science in March 2026[1]. Citing the 2023 working paper now signals that the reader has not tracked the literature.
Carry the meta-analysis into every briefing. An audience that later encounters the g = −0.23 result[7] without having heard it from you will discount everything else you said. Present it, then present the task-type reconciliation.
Separate measured performance from perceived performance. The METR inversion[9] is the strongest available argument against accepting internal AI-benefit surveys as evidence. Present it with its update rather than without: METR is redesigning the experiment and no longer offers the 2025 figure as its current estimate[27], so lead with the inversion itself rather than with the 19%. The direction is the finding, and the direction is what survives. For any client claiming productivity gains, ask what was measured rather than what was reported.
Treat mode as an observable governance variable. Policy language cannot distinguish the three modes, because all three satisfy a human-in-the-loop requirement[4]. Any assurance regime that counts AI usage rather than observing interaction pattern will miss the abdicated case entirely. This is the point where the framing converts into something an executive can act on.
Hold a position on shelf life. Krakowski, Luger and Raisch show substitution and complementation operating together, with new human-machine capabilities emerging as traditional ones are made obsolete[15]. Expect the question of whether Centaur mode survives. Answer it from what that paper establishes, and mark the step from those dynamics to a trajectory as your inference rather than its finding.
The claim[1], the correction[4], and the challenge[7]. These three establish the position and its strongest opposition.
The recruiter disengagement study[6], the design-artefact re-analysis[8], and the METR trial[9]. Together these explain why combinations fail and why self-report cannot be trusted.
The Kenya field experiment[14] against the customer-support field study[12]. This pairing produces the levelling contradiction described in section 3.2.
Bainbridge[21] and Lee and See[23] for why oversight degrades. Raisch and Krakowski[16] and Krakowski, Luger and Raisch[15] for the management theory. Lebovitz and colleagues[17] for what engaged use looks like in a regulated profession.
The peer-review status of the seven unreviewed sources that carry empirical weight — [4], [6], [8], [9], [10], [13] and [27] — which is the same list section 2.6 gives, so the two cannot drift apart. [20] is unreviewed too and is deliberately not monitored: a 1962 institutional report will not change status. METR's redesigned experiment[27]. The publication of [26] in October 2026. Any replication of the Kenya result, which would settle section 3.2.