As generative AI rapidly proliferates throughout organizations, executives are faced with a seemingly simple question: How should humans work with AI? A common answer is “keep humans informed,” which sounds reassuring.
But new research reveals that this answer is dangerously incomplete. What appears to be the same “human-involved” approach actually manifests itself in three fundamentally different ways, with very different impacts on performance and skill development.
To understand how companies can derive real value from human-AI collaboration, we conducted a field experiment with 244 consultants using GPT-4 for complex business problem-solving tasks. With the help of scholars from Harvard Business School, MIT Sloan School of Management, The Wharton School, and Warwick Business School, the experiment analyzed nearly 5,000 human-AI interactions to answer the important question: “What are humans actually doing and what should they be doing when they collaborate with GenAI?”
Three hidden patterns in collaboration between humans and AI
The most striking finding of our experiment was that experts using GenAI naturally categorized themselves into three different collaboration styles, each of which produced dramatically different results.
Cyborg (60% of participants) We are engaged in what we call “fused knowledge co-creation,” or continuous, iterative interaction with AI throughout the workflow. They used it in different ways for each subtask of their workflow. We assigned personas to the AI, divided complex tasks into modules, pushed back on AI output, uncovered inconsistencies, and validated results with dynamic interactions. For Cyborg, the line between human and AI thinking has been deliberately blurred.
Centaur (14% of participants) We practiced “directed knowledge co-creation,” which uses AI selectively for specific subtasks while maintaining tight control over the entire problem-solving process. They leveraged AI to enhance their capabilities, map problem areas, gather methodological information, and refine their own human-authored content. But they were firmly in the driver’s seat, using AI as a targeting tool rather than a collaborative partner.
Self-automater (27% of participants) We are committed to ‘Abdicated Knowledge Co-Creation’, delegating entire workflows to AI with minimal repetition and critical involvement. We provided the AI with data and instructions to perform a subtask and accepted its output unchanged or with only minor edits. Their work was fast and elegant, but lacked depth and resembled a finished product. for rather they and they.
Notably, all participants had access to the same tools and the same tasks. We have not received any different instructions regarding the AI work process. However, their emergent/instinctual choices about when to collaborate with AI and how much authority to give it created fundamentally different collaboration dynamics.
A framework for understanding collaboration
To understand these patterns, we developed a framework built around two fundamental questions that structure the dynamics of collaborative problem-solving between humans and machines. Who chooses what to do? and Who specifies how that is done?
Cyborgs will give humans the power to drive what, but AI will have greater control over the how. Centaur maintains human control and leadership on both sides and uses AI only for targeted assistance. Self-automators cede control of both to the AI. In particular, the fourth theoretical possibility (if the AI takes the lead in task selection, but the human takes the lead in execution) remained empty in our study. When experts relinquish control over what they work on, they also tend to relinquish control over how they do it.
Hidden costs: What happens to your expertise?
Perhaps our most important finding concerns what happens to expertise in each collaboration mode. The effects vary widely.
Cyborgs have developed new AI-related expertise— what we call “Newskilling.” Through continuous experimentation with prompting strategies, they learned how to effectively communicate with AI, when to push back, and how to get the most value from collaboration. We also maintained our subject matter expertise by remaining actively involved throughout the process.
Centaurs deepened their domain expertise—Traditional “skilling up”. They used AI to accelerate learning about unfamiliar industries, glean methodological guidance, and refine their thinking to build stronger foundational capabilities. However, because our interactions with AI were limited and targeted, we were unable to develop significant AI-related expertise.
Neither self-automator was developed.—Experience what is called “no skills.” By delegating the entire cognitive process to AI, we missed an opportunity to build domain knowledge and AI fluency. Their productivity gains came at the expense of professional development.
This discovery should give executives pause. When employees default to self-automating behaviors (as more than a quarter of our highly trained consultants do), organizations may be inadvertently hollowing out the very expertise that creates competitive advantage.
Performance Impact: Who Gets It Right?
In our experiment, we evaluated the output on two dimensions: accuracy (did we recommend the right brand?) and persuasiveness (how persuasive was the CEO’s memo?). This result challenges simple assumptions about AI collaboration.
Centaur achieved the highest accuracy— outperforms both cyborgs and self-automators when it comes to getting the right answer. By maintaining control over the analysis process and evaluating AI inputs with our own judgment, we avoided being fooled by AI’s confident but sometimes inaccurate recommendations.
Both cyborgs and centaurs were very persuasive.—Produces more convincing output than Self-Automators. Depth of engagement through iterative refinement (Cyborg) or human-driven analysis (Centaurs) translates into higher quality deliverables.
Cyborgs, in particular, have sometimes fallen victim to the persuasive powers of AI. Even when they employed best practices like validation, which involves having the AI check its own work, they were sometimes convinced that the AI confidently justified incorrect answers. This highlights a significant risk. Advanced collaboration with AI does not guarantee immunity from its errors.
What should companies do now?
These findings have immediate implications for how organizations deploy GenAI.
First, let go of the myth of a single “human participatory” approach. Managers need to recognize that their employees already have vastly different collaboration styles and that these differences are important. Simply mandating “human oversight” without clarifying what that means will produce deeply contradictory results.
Second, align your collaboration style with your strategic goals. For tasks that require maximum precision in high-stakes decision-making, we encourage Centaur’s behavior, the selective use of AI with strong human judgment. For tasks that require rapid iteration and creative exploration, cyborg behavior may be more appropriate. Reserve the Self-Automator approach for truly mundane tasks, rather than core or high-risk tasks, and when skill development is not a concern.
Third, monitor automation complacency. The 27% rate of self-automation in our study among highly skilled and motivated professionals who know their performance is being evaluated suggests that the temptation to overdelegate is strong. Organizations must develop mechanisms to detect when employees are moving toward full automation in tasks that require human involvement.
Fourth, rethink how you measure the success of your AI deployment. It’s not enough to just use end results like edit rates or approval rates as metrics for engagement. A self-automator that accepts the output of an AI and a cyborg that accepts a widely iterated and refined version can appear identical in data. Companies need to track the quality of interactions throughout the workflow, not just the results.
Fifth, invest in developing AI fluency along with domain expertise. Our findings suggest that combining both is the most sustainable approach. Cyborg behavior builds advanced AI skills while maintaining expertise. Centaur behavior builds domain skills while providing baseline AI exposure. Companies need training programs that intentionally develop both abilities, rather than expecting employees to figure it out on their own.
The stakes: Expertise in the age of AI
The emergence of GenAI brings contradictions to organizations. This technology promises to improve human judgment, creativity, and speed, but it also comes with quiet risks. That is, by delegating more and more thinking to machines, experts may gradually relinquish the very capabilities that make them valuable. The same tools that hone expertise in some people can be completely superseded in others, allowing organizations to achieve better results in the short term, but diluting the core of human judgment. This isn’t just an efficiency tool; it’s a revolution. Fortunately, productive modes of collaboration exist. Cyborgs and centaurs demonstrate that humans can effectively collaborate with AI, building on rather than depleting human expertise. The challenge for managers is to create the organizational conditions that foster these productive patterns while discouraging the seductive but self-defeating path of full automation.
As the capabilities of AI continue to expand and improve, successful organizations will not only understand what AI can do; how Humans should work on it. The first step to building proficiency in it is understanding that “human engagement” is not one approach, but actually three fundamentally different modes of collaboration with radically different outcomes.
François Candelon is a partner at the private equity firm Seven2 and an executive fellow at Harvard University’s D^3 Institute.. read other luck Column by François Candron.
Katherine Kellogg is the David J. McGrath Jr. Professor of Management and Innovation at the MIT Sloan School of Management..
Hila Lifshitz is a professor of management at Warwick Business School, a faculty assistant at Harvard University’s Institute for Innovation Sciences, and co-director of the AI Innovation Network..
Steven Randazzo is a PhD student at Warwick Business School, a visiting scholar at Harvard University’s Institute for Innovation Sciences, and co-director of the AI Innovation Network..
