In October 2002, U.S. intelligence agencies set down their judgment of Iraq’s weapons programs in a National Intelligence Estimate. One question was what Iraq wanted with the high-strength aluminum tubes it had been trying to buy. “Most agencies,” the estimate said, saw them as compelling evidence that Saddam was reconstituting a uranium enrichment effort for a nuclear weapons program.
The Department of Energy, whose analysts the Senate would later call “the Intelligence Community’s nuclear experts,” disagreed. Its view went into the estimate in parentheses: the tubes “probably are not part of the program.”
The agencies had been arguing about those tubes for a year and a half.
The dissent in parentheses
In July 2004, the Senate Select Committee on Intelligence published its review. The information available, it concluded, “indicated that these tubes were intended to be used for an Iraqi conventional rocket program and not a nuclear program.”
Then it named a cause. The intelligence community had suffered from “a collective presumption” that Iraq had an active and growing weapons of mass destruction program, a “group think” dynamic. The presumption was so strong, the Committee wrote, that the community’s own tools for challenging assumptions, such as red teams and devil’s advocacy, went unused.
Nobody had hidden the doubt. It was in writing, inside the estimate. When agreement matters more than being right, a group stops hearing its own doubts.
A group that is sure of its answer reads evidence the way a fan reads the replay of a referee’s call. The same footage looks conclusive when it favors your side and murky when it doesn’t. The Committee described exactly that pattern: ambiguous evidence read as “conclusively indicative” of a weapons program, and contrary evidence ignored or minimized.
Janis’s word
The Committee credited the term to the psychologist Irving Janis, whose 1972 book Victims of Groupthink described what happens when members’“strivings for unanimity override their motivation to realistically appraise alternative courses of action.” The word itself first appeared in Fortune in 1952, where William H. Whyte Jr. used it for “a rationalized conformity.” Challenger got the label too: the night before the 1986 launch, not one engineer in the room spoke in favor of launching, and once it became “a management decision,” it didn’t matter.
Letting a second opinion in
Janis set the 1961 Bay of Pigs invasion beside the Cuban Missile Crisis a year later, inside the same White House. In his account the difference was that, the second time, someone was assigned to argue against the plan instead of waiting for a volunteer. A historian who studied the White House tapes has questioned parts of that account.
The 2002 estimate had the opposite arrangement. The tools existed. Nobody ran them.
Where the label slips
The stories are vivid. The research is thinner. Janis built his model from historical cases, and later tests have been unkind to it. In 1998 the management scholars Sally Fuller and Ramon Aldag wrote of “a quarter century virtually devoid of support for the phenomenon.” Aldag later described a 2000 experiment by Won-Woo Park that tested 23 of the model’s predictions: two were supported, and seven came out significant in the opposite direction.
Janis meant the term narrowly, for high-stress decisions and only when practically all the symptoms appear. The Senate stretched it over an entire community of agencies, one of which had dissented in writing for over a year. What survives the critics is the smaller claim: groups can lean too hard on agreement. The researcher Jerry Harvey argued that some of what looks like conformity pressure is stranger still, people going along with a consensus nobody holds. He called it the Abilene Paradox.
The vendor nobody objects to
Imagine you run the review of a new payments vendor. Your director championed it, and in three meetings nobody has raised a serious objection. Before the fourth, you change one thing. Each of the five reviewers writes a verdict and the biggest risk on a card, alone, before anyone sees anyone else’s.
Two cards name the same risk, one nobody has said out loud. It is on the table now, with two names on it, before the contract is signed.
Seven agents, one answer
A 2025 study from Zhejiang University, presented at the ICLR conference, borrowed the setup of Solomon Asch’s conformity experiments. Seven agents answered reasoning questions in turn, and the model under test went last.
The peers were simulated: their answers were scripted by the researchers, not reached by a real team. In one condition, the six had been right in earlier rounds and then all gave the same wrong answer. GPT-4o went wrong on 37.9 percent of the questions it answered correctly on its own. Breaking up the majority helped one model and, in one condition, made another conform more.
A reviewer agent that sees only the first agent’s summary is in the position of the CIA contractors who reviewed the tubes analysis: they got their information only from the CIA and did the work in one day. If you use agents to review each other, run each reviewer in a fresh session that can’t see the others’ verdicts. Give it the raw material, not the first agent’s write-up. Then plant one error you know about and see which reviewers find it.
If none of them finds it, then at least for that one known error, five agreeing reports gave you no effective check.