Insights
Why Teams Need Disagreement in the Age of AI
Consensus used to be expensive. That may have been the only thing keeping it honest. Getting a room to agree once required real work: circulating a draft, absorbing objections, revising, persuading. That cost acted as a filter. If a group finally aligned, the alignment usually meant something had been tested. Now a team can generate a coherent, defensible position in ninety seconds. Everyone arrives already aligned — not because they argued their way there, but because they consulted the same class of system and received the same reasonable answer.
Agreement has become cheap
Agreement has become cheap, and cheap agreement is hard to distinguish from correct agreement. I keep returning to a result from a paper on team production (arXiv 2512.22736, Dec 2025). Holding average optimism constant, a team’s expected output rises with the degree of disagreement among its members. When people hold different priors about which method will work, they work harder to demonstrate their own, and the team learns faster. The authors go further: a manager forming pairs from a large workforce maximizes output by matching beliefs negatively.
Deliberately pairing people who disagree
Deliberately pairing people who disagree. That is an uncomfortable finding for anyone who has ever built a team by hiring “culture fits.” The empirical work is less romantic about how disagreement behaves in a real room. In a June 2026 study of power-imbalanced groups, researchers gave dissenting minority members AI support. Model-generated counterarguments made the atmosphere more flexible and the dissenters more satisfied. But when the AI paraphrased and delivered their dissent for them, participation went up while psychological safety went down. People spoke more and felt less safe. A machine can carry your objection into the room. It cannot make the room a place where objecting is survivable. A Scientific Reports paper published this week points at what does. Across two experiments, activating a sense of secure attachment to the group made minority opinion holders more task-focused and their dissent more persuasive — and made the majority more willing to use it. The variable was not courage or training. It was whether the dissenter believed the group would still hold them afterward. Disagreement has economic value. AI can amplify its volume. Only the organization can make it usable. This is where one principle of Contribution Leadership does most of its work: optimization removes variation, and variation is where discovery lives. For two centuries we treated variation as noise — a defect in the process. That was defensible when execution was scarce and consistency was the advantage. It is much less defensible now. Consistency is what the machines are for.
Difference is what the humans are for
Difference is what the humans are for. Which suggests an odd new managerial duty. Not building alignment. Protecting the residue of disagreement that survives after the models have spoken. A few questions I would want a leadership team to be able to answer: When we agree quickly, do we know whether we reasoned or retrieved? Who in this room was wrong last quarter, and were they rewarded for having been interestingly wrong? When someone dissents, do they leave the meeting more attached to the group or less? None of this argues for manufactured conflict. Contrarianism as a personality is a cost, not a contribution. A 2025 study on collective decisions found that whether a group values agreement or dissent depends on how strong its consensus already is — dissent is precious to a divided group, threatening to a settled one. Timing matters. But I suspect the organizations that struggle most in the next decade will not be the ones that argue too much. They will be the ones that stopped noticing when they had stopped arguing — because the agreement arrived so quickly, so fluently, and so uniformly that nobody thought to ask where it came from.