Stop Using AI as a Search Engine. Use It as a Sparring Partner.
By Martin Bergmann | AI Project Lab | September 2026
Ask most people what they actually use AI for when a real decision is on the line, and the honest answer is a gut check: here is my plan, does this make sense? What comes back is almost always encouraging, well-organized, and confident. It rarely tells them anything they did not already believe.
Agreement is not the same as insight.
AI is very good at making a plan sound more finished than it is. Ask it to review a proposal, and it will usually tighten a few word choices, maybe flag a risk, and then confirm the overall direction. That confirmation feels like progress. It is not the same thing as finding out what is actually wrong with the plan.
In the 1910s, Andrew Carnegie reportedly dared a young journalist named Napoleon Hill to spend two decades interviewing the most successful people of his era and find the single idea behind their success. Hill published the result in 1937 as Think and Grow Rich. Most of the book has not aged well, but one idea in it has: the “mastermind,” a small group assembled specifically to challenge each other’s thinking, not to agree with it. Hill’s own account of this was closer to two decades of personal observation than to a study, and it is worth reading it that way rather than as proof of anything.
The empirical backbone came later, from a different field entirely. Starting in the 1980s, researcher Charles Schwenk and colleagues ran a long series of studies on what happens when a group is deliberately assigned to argue against its own plan, a technique called devil’s advocacy, compared with a group simply asked to reach consensus. Across decades of replication and a later meta-analysis, the pattern held: groups forced into structured disagreement produced higher-quality recommendations and surfaced more of their own hidden assumptions than groups that just sought agreement. The exact technique that worked best varied from study to study. The core finding did not: disagreement, done deliberately, beats consensus.
Where the bottleneck actually was.
Assembling a group willing to argue against you, with enough relevant knowledge to do it well, was always the hard part. Most people do not have five qualified peers on call who are willing to tell them their plan has a hole in it. That scarcity, not a lack of interest in the technique, is likely why devil’s advocacy stayed a boardroom exercise instead of an everyday habit.
A 2025 study at the ACM Intelligent User Interfaces conference tested a version of this with AI directly: a large language model assigned to voice dissenting views inside a group discussion, specifically to reduce the pressure that keeps people from disagreeing with the loudest voice in the room. The system reduced that conformity pressure and improved the group’s critical thinking. That study was about groups, not individuals working alone, but the mechanism is the same one Schwenk identified decades earlier: someone, or something, willing to push back changes the outcome.
Where AI actually helps.
Most people already have access to this and are not using it this way. The habit worth building is small: stop asking AI whether your plan makes sense; instead, ask it to argue against it. What is the strongest case that this is the wrong call? What would someone who has watched this exact thing fail somewhere else say? Where is the assumption you are not questioning because you already believe it?
The answer will not always be right. That is not the point. A good counterargument, even a wrong one, forces you to check your own reasoning in a way that agreement never will.
A search engine tells you what is already known. A sparring partner tells you what you have not questioned yet. Most people are still asking it to be the first one.
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.
Comments