“When you asked me after the meeting yesterday what I thought about the Cuban proposal, I am afraid that I did not give a properly ordered answer.”
Arthur Schlesinger Jr., a historian working as a White House special assistant, sent that line to President Kennedy on April 5, 1961. The meeting he meant had run from 6 to 8:18 the evening before, at the State Department. On the table was a CIA plan to land a brigade of Cuban exiles on the island.
The room had heard an objection. Kennedy had invited an outsider, Senator J. William Fulbright, who spoke out against the plan. In Irving Janis’s account, built largely on Schlesinger’s, the President did not open Fulbright’s points for discussion. He went round the table asking each man for his verdict, and the meeting broke up before the round reached Schlesinger.
So Schlesinger’s case arrived the next day, on paper, for one reader. If the landing failed to set off uprisings, he warned, the logic of the situation could lead “step by step” to sending in the Marines. On balance, he wrote, “I am against it.”
The disagreement existed, out loud and in writing. It never became the room’s argument.
The meeting Sloan postponed
In 1967, Peter Drucker put it bluntly in The Effective Executive: “The first rule in decision-making is that one does not make a decision unless there is disagreement.”
His illustration came from General Motors. Alfred P. Sloan, Drucker wrote, “is reported to have said” at a meeting of one of his top committees that he took it they were “all in complete agreement on the decision.” Everyone around the table “nodded assent.” Then Sloan proposed to “postpone further discussion” until the next meeting, to give the committee time to “develop disagreement and perhaps gain some understanding of what the decision is all about.”
When nobody in the room disagrees, you haven’t seen the alternatives yet.
Drucker gave reasons. Disagreement supplies alternatives, and “a decision without an alternative is a desperate gambler’s throw.” It is needed to stimulate the imagination. And it keeps the person deciding from becoming “the prisoner of the organization,” where everybody always wants something from them.
A map app that shows one blue line gives you a route. The grey lines beside it tell you what the blue one costs. In April 1961 the grey lines existed, in a senator’s remarks and an aide’s memo. Nobody drew them on the map.
Eighteen months later
In October 1962, U-2 photos showed Soviet missile sites going up in Cuba. This time the advice was openly split. All of the Joint Chiefs of Staff argued for an air strike followed by an invasion. Others favored stern warnings. Kennedy chose a middle course, a naval “quarantine.”
One good outcome doesn’t prove a method. Experiments have tested a narrower claim. In a 2006 experiment, Stefan Schulz-Hardt and colleagues gave 135 three-person groups a personnel decision with four candidates, and split the facts so that the best choice only showed up if members pooled what they alone knew. Groups whose members all started out preferring the same wrong candidate hardly ever found the right one. Groups that started out disagreeing found it significantly more often, even when nobody’s first choice was correct. The study did not detect a decision-quality difference between minority dissent and full disagreement. The effect ran mainly through more intense discussion.
That was a lab task built around hidden information. Where everyone holds the same facts, disagreement has less to uncover, and it still costs a meeting.
The price nobody questioned
Imagine you lead a small software team, and the proposal is to raise prices 15 percent. Everyone in the meeting nods. Instead of deciding, you ask each person to come back next week with one other way to reach the same revenue.
Your support lead comes back with a spreadsheet. Most cancellations last quarter came from customers on the oldest plan, the ones a flat increase would hit hardest. Her alternative is to raise prices only for new customers. You may still pick the 15 percent. Now it’s a choice, with its cost beside it.
A model that agrees with itself
Asking a language model to review its own answer puts the review inside a decision already made.
In a 2023 study, researchers from Tsinghua University, Shanghai Jiao Tong University and Tencent AI Lab gave GPT-3.5-Turbo 200 arithmetic questions built with traps that invite an intuitive, wrong answer. In one, a small circle rolls once around a circle three times its radius. How many times does it turn? The model said 3; the paper’s answer is 4. In the paper’s self-reflection example, the model wrote, “I am confident that my answer is correct.”
Self-reflection moved accuracy on the 200 questions from 26 percent to 27.5. The authors call this degeneration of thought: in their description, once a model has settled on an answer, reflection stops producing new ideas, even when the answer is wrong. So they set two copies against each other, one giving an answer and one replying, under instructions that full agreement wasn’t necessary because the goal was the correct answer. A third copy judged. Accuracy rose to 37 percent. GPT-4 alone scored 51.
The same paper found limits. In translation tests, instructing the debaters to disagree on every point did not give the best results. And when the debaters were different models, the judge tended to side with the one built on the same model as itself.
Before you ask a model which option is best, ask it for three genuinely different options and what each gives up. If the three turn out to be one idea in different words, you haven’t got your grey lines yet. Ask again in a fresh session, and check each option’s claims before you pick.
Brigade 2506 landed at the Bay of Pigs on April 17, 1961. Within two days, Cuban forces under Castro’s direct command had defeated it.