The conversation about AI in professional settings keeps circling the same question: should we use it or not? That framing is already obsolete. The real question, and the one Randazzo et al. (2025) tackle in a new Harvard Business School working paper, is how people use AI when they do use it, and what that choice costs them.
The study followed 244 BCG consultants as they worked through a realistic business problem using GPT-4. What emerged was not a single pattern of AI use but three distinct modes of co-creation, each with different implications for output quality and professional skill development. The researchers call them Cyborgs, Centaurs, and Self-Automators. The labels are memorable, but the substance behind them is what makes this paper worth reading carefully.
Three Modes of Human AI Collaboration
Randazzo et al. build their framework around two questions: who decides what needs to be done, and who decides how to do it? The answers split their 244 participants into three groups.
The largest group, 60% of participants, fell into what the researchers call Fused Co-Creation (Cyborgs). These professionals used AI across every phase of their workflow: framing the problem, analyzing data, generating recommendations, iterating, extending ideas, evaluating results, and drafting the final report. They assigned personas to the AI, broke tasks into modular steps, validated outputs, challenged recommendations, and fed in new data mid-conversation. Their engagement was constant and exploratory.
Directed Co-Creation (Centaurs) made up 14% of the sample. These consultants used AI selectively, treating it more like a reference tool than a collaborator. One asked for Excel formulas. Another requested examples of clothing retailers that had successfully pivoted. They gathered what they needed and then did the analytical work themselves. Crucially, Randazzo et al. found that Centaurs didn’t use AI at all during the iterating phase. They iterated on their own, relying on their own understanding and skills. They also evaluated their solutions independently.
The remaining 27% were Self-Automators (Abdicated Co-Creation). These professionals consolidated the entire problem-solving process into one or two prompts. They copied all materials into the AI, asked it to analyze, recommend, evaluate, and report, and largely accepted whatever came back. Of the 63 Self-Automators Randazzo et al. studied, 44% accepted the AI’s output with zero modifications. The rest made only cosmetic edits to the final memo.
The Skilling Problem
The most consequential finding in this paper is about what happened to the professionals’ skills after the task.
Randazzo et al. identify three skilling outcomes tied to the three modes. Cyborgs developed new AI-related expertise, what the authors call “newskilling,” through iterative experimentation with prompting strategies, validation techniques, and conversational approaches.
They also maintained their domain expertise because they stayed engaged with the substance of the work throughout. The Centaur group tells a different story: because they did the core analytical work themselves, they came away with stronger task-related expertise (“upskilling”). AI supported their learning curve but didn’t replace their thinking. And the Self-Automators? No AI skills, no domain skills. The authors call this “no skilling.”
As Randazzo et al. put it, “co-creation styles are not only about efficiency; they can be engines of skill transformation, determining whether professionals build AI fluency, deepen domain expertise, or risk losing both” (p. 39).
I’ve been writing about this pattern from other angles on this blog. Shaw and Nave (2026) named it cognitive surrender, the slow erosion of independent reasoning that happens when people cede decision-making to AI systems. The Self-Automators in this study are living proof of that concept in a professional context. Fan et al. (2025) documented a version of the same thing in student writing: AI improved the essays without improving the students. Randazzo et al. are now showing it happening with experienced consultants solving real business problems, which makes the pattern much harder to dismiss as a student-level concern.
When Validation Fails
One detail in this paper needs a closer look. A consultant named Felix Schneider, working in the Fused mode, asked GenAI to verify its own recommendation. The AI confidently confirmed the answer was correct, backed it up with reasoning, and Schneider accepted it. The answer was wrong. Randazzo et al. note this happened to multiple professionals in the Cyborg group.
This is a serious limitation of the Fused model. Even when professionals try to stay critical, asking the AI to check itself is a circular safeguard. Kosmyna et al. (2025) showed that ChatGPT reduced neural engagement during writing tasks at MIT, and I think a similar mechanism operates here. When you’re deep in a conversational flow with AI, validation can become performative. You go through the motions of critical thinking without actually exercising it. The Schneider example is a warning: engagement with AI is not the same as independent judgment.
Who Got the Best Results?
Centaurs produced the most accurate business recommendations, outperforming both Cyborgs and Self-Automators. Both Centaurs and Cyborgs produced more persuasive outputs than the Self-Automator group. The mode with the deepest human engagement in the analytical work delivered the strongest results.
That finding should complicate the productivity narrative around AI. The Self-Automators were the most efficient, finishing fastest and interacting the least. And their work was the weakest. More AI did not produce better output. The relationship between AI reliance and quality is not linear, and anyone designing AI training for their organization needs to understand that.
I think this connects to what Bastani et al. found in their study of nearly 1,000 high school students using ChatGPT for math: AI helped during practice but hurt performance once the tool was removed. The mechanism is parallel. Offloading the cognitive work produces results in the moment but weakens the capacity to produce them independently.
Randazzo et al. make a point that extends well beyond individual users. Organizations shape co-creation choices through their culture, incentives, and training programs. Firms that prioritize workforce development tend to encourage augmentation, the Fused and Directed modes. Firms that prioritize speed and efficiency may push their people toward abdication, hollowing out expertise across the workforce over time.
The authors warn that “whether professionals lean into AI as collaborators, treat it as a selective tool, or offload their judgment to it altogether will shape not only the outcomes of today’s projects but also the very contours of expertise, authority, and competitive advantage within their organizations” (p. 42).
The encouraging piece is that co-creation modes are not fixed identities. Randazzo et al. argue that the same professional might operate as a Cyborg on one project and a Centaur on the next. The key is matching the mode to the demands of the task, not defaulting to whatever feels most comfortable. That’s a trainable skill, and it’s one that professional development programs should be building right now.
A Note on Timing
The data in this study was collected using GPT-4 in April 2023. AI capabilities have changed significantly since then. Models are far more capable, tools are more deeply integrated into everyday workflows, and professionals have had years of experience adjusting to GenAI. The three modes Randazzo et al. describe may still hold, but the distribution across them could shift as familiarity grows. I’d want to see a longitudinal version of this study before drawing conclusions about durability.
The underlying logic, though, is solid. How you distribute decision-making authority with AI shapes what you learn and what you produce. That principle doesn’t expire with a model update.
References
- Bastani, H., Bastani, O., Sungu, A., Geb, H., Kabakcı, Ö., & Marimane, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122(26), e2422633122. https://doi.org/10.1073/pnas.2422633122
- Fan, Y., Tang, L., Le, H., Shen, K., Tan, S., Zhao, Y., Shen, Y., Li, X., & Gašević, D. (2025). Beware of metacognitive laziness: Effects of generative artificial intelligence on learning motivation, processes, and performance. British Journal of Educational Technology, 56(2), 489–530. https://doi.org/10.1111/bjet.13544
- Kosmyna, N., Hauptmann, E., Yuan, Y. T., Situ, J., Liao, X.-H., Beresnitzky, A. V., Braunstein, I., & Maes, P. (2025). Your brain on ChatGPT: Accumulation of cognitive debt when using an AI assistant for essay writing tasks. MIT Media Lab. https://www.media.mit.edu/publications/your-brain-on-chatgpt/
- Randazzo, E., Lifshitz, A. H., Kellogg, K. C., Dell’Acqua, F., Mollick, E. R., Candelon, F., & Lakhani, K. R. (2025). Cyborgs, centaurs, and self-automators: How professionals co-create with generative AI (Harvard Business School Working Paper No. 26-036). Harvard Business School. https://www.hbs.edu/ris/Publication%20Files/26-036_e7d0e59a-904c-49f1-b610-56eb2bdfe6f9.pdf
- Shaw, S. D., & Nave, G. (2026). Thinking fast, slow, and artificial: How AI is reshaping human reasoning and the rise of cognitive surrender. Working paper, The Wharton School, University of Pennsylvania. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6097646