The Busiest AI Users Were Not the Most Accurate

By Martin Bergmann | AI Project Lab | November 2026

Think about the last real piece of work you did with an AI. Which of these sounds most like you?

A. I gave it the task, took what came back, and tidied the wording.

B. I went back and forth with it the whole way: asking, pushing back, asking again.

C. I worked out the logic myself and used it for pieces: gathering, sorting, polishing.

Pick one before you read on.

Researchers from Harvard, MIT, Wharton, and Warwick Business School gave 244 junior consultants at Boston Consulting Group the same business problem and a GPT-4 chatbot, then read every conversation. They found the same three styles you just chose from, and they named them. A is the Self-Automator. B is the Cyborg. C is the Centaur.

About six in ten people were Cyborgs. A little over a quarter were Self-Automators. One in seven was a Centaur.

The consultant who checked.

The task was a fictional company with three brands. Recommend which one should get the investment, then write the CEO a memo. There was one right answer, the men’s brand. The authors describe the task as outside what the AI could reliably do at the time, so a wrong first answer was a real possibility.

One consultant got exactly that: a confident recommendation for the wrong brand. Then they did what the advice tells us to do. They told the AI to check the accuracy of its answer. The AI checked, confirmed its own findings, and cited a statistic in its own defense. Satisfied, the consultant moved on. The answer was wrong.

The authors call this "validating" and say it happened to many Cyborgs. I could not find a count in the paper.

The busiest users were not the most accurate.

Here is the part that surprised me. The Centaurs, the smallest group, had the highest accuracy. They kept the overall logic for themselves and handed the AI the jobs around it, like mapping the data and polishing the draft. The Cyborgs, who worked with the AI at every stage, ranked below them. The paper reports accuracy as a ranking, not as percentages.

Cyborgs and Centaurs both wrote more persuasive memos than the Self-Automators. And 44 percent of the Self-Automators passed the AI’s output along without changing a word.

Checking is not the same as talking.

My reading, not the paper’s: when you ask the same AI to review its own answer, you get one source with two turns. A long conversation can feel like rigor and still check nothing. It is last week’s problem in a new form. Rereading feels like learning, and the back-and-forth feels like checking. If that idea sounds familiar, my last post covered the first half.

The paper does quote Cyborgs who caught the same wrong answer. In the examples the authors show, what they did differently was bring their own facts. One spotted a growth figure that did not match the data, pointed it out, and the AI changed its answer.

One habit to try.

This is my suggestion, loosely modeled on what the Centaurs did. Before you open the AI, write two lines: what I decide, and what I hand over. The decision stays with you: which option, and why. Mapping, sorting, drafting, and polishing can be handled by the AI. When you finish, ask one question: could I defend this answer without the chat open? If you want a check, use something the AI did not write, like the source data or a colleague.

Keep the limits in mind. This was one session, one fictional task, GPT-4 from April 2023, and high-performing junior consultants. It is a working paper that has not been peer-reviewed, and the tools have improved since. I read it as a pattern to watch, not a verdict.

Now look at your pick from the start. Whichever it was, think of the last time the AI was wrong in your work. Who caught it?

---

Martin Bergmann is the Director of an IT PMO and founder of the AI Project Lab. He writes about AI, project management, and the future of intelligent work.