In a meeting convened to review material that an AI drafted, the opinions in the room drift quietly toward one conclusion. When everyone is converging on the same answer, what can the single dissenter lean on? Not personal strength, but one ally who judges the same way, and a procedure that keeps the disagreement on the record.

01One ally, and not character, decided who held their ground

In 1955 Asch ran an experiment built on nothing harder than comparing the lengths of lines. You can see the answer. Yet when a unanimous majority around the subject gave the wrong one, subjects went along with it in 36.8 per cent of the selections.

Add a single other person who breaks with the majority, and the wrong answers fell to about a quarter. What was added was not persuasion or training. It was the bare condition of one person in the room judging the same way.

Character was not what separated the two outcomes. This piece works through three things in order: why dissent helps even when it is wrong, what disappears first once AI advice enters the room, and which arrangements keep dissent alive without relying on nerve.

02A minority view earns its keep even when it is mistaken

One ally is enough to hold ground. Folded into that result is the fact that the ally need not be right.

In 1986 Nemeth set out what minority dissent does to a group. With dissent present, a group stops converging quickly on one answer, looks for other lines of reasoning, and re-examines its own premises. That, Nemeth argued, is why decision quality and creativity improve.

What does the work is not the correctness of the dissent. It is the presence of dissent in the room. Even mistaken opposition forces everyone else to put their reasons back into words.

I am writing within the range of a 2017 paper that cites Nemeth. I have not read the 1986 original. Adding detail as though I had would put everything that follows on unsteady ground.

Within that range, something still holds. Whether to accept an objection need not depend on whether the objection is right.

Figure 1 What one ally did to the error rate
Unanimousmajorityplainly wrong36.8% erroraloneOne allybreaking with themajorityError down to aquarterUnanimous majorityplainly wrong36.8% erroraloneOne allybreaking with the majorityError down to a quarter
Against a unanimous majority, error reaches 36.8 per cent. One dissenting ally cuts it to a quarter.

Between the third and fourth steps of the diagram there is no process of persuasion. The moment the condition of having one ally exists, the numbers have already moved.

03In material review, the differing judgment tends to land outside the room

Dissent improves a group's judgment even when it is wrong. So why is a differing judgment so rarely voiced in a review meeting?

The reviewer's ability is not the issue. The majority view settles first, and only then does the turn to speak come round. Under that sequence, nobody is left to take on the first objection. In that respect — a majority settled first, the first objection unclaimed — the meeting is close to the condition Asch built.

The differing judgment does not vanish. It moves to the conversation after the meeting and never reaches the minutes. Six months later, when the same question returns, all that remains is somebody's memory that someone once raised it.

TestMeeting where the majority settles firstMeeting with an assigned first objector
The first objectionNobody takes it onAssigned as a role
Where dissent goesInto talk after the meetingInto the minutes
Quality of the decisionNo divergent material is producedMore material to review later

What separates the two columns is not the temperament of the people present. It is whether someone decided, before the meeting, who speaks first against. Once that is decided, the objection stops being an act of courage and becomes the performance of a role.

04A fluent machine answer drives out the words: I do not know

Nobody takes on the first objection. Now add AI advice to that condition.

In an experiment reported in 2026, merely having access to AI nearly eliminated participants' willingness to suspend judgment. They answered more questions, and they were correct about a third as often. Their confidence, meanwhile, nearly doubled. Accuracy was incentivised, so guessing carried no reward.

Before minority views disappear, the act of withholding one disappears. In a room where nobody says they do not know, a differing judgment has little chance of surfacing either.

Figure 2 What moves when AI advice arrives
A fluent machine answerWithholding nearlyvanishesConfidence nearlydoublesAccuracy falls to athirdA fluent machine answerWithholding nearly vanishesConfidence nearly doublesAccuracy falls to a third
An answer comes back, and people stop saying they do not know. Confidence rises while the share of correct answers falls.

The result does not show that AI makes people less capable. It shows that in an environment where an answer always comes back, declining to answer becomes hard to choose.

Carry that into material review and what is hardest to catch later is not the mistaken objection but the item passed while the reviewer was still unsure. The more plausible the reasoning attached to an AI draft, the harder that state of being unsure is to see.

05In safety panels, one peer message lifted false alarms from 56.5% to 87.5%

If AI advice reduces withholding among people, the next thing worth checking is whether the same skew appears when AI systems are lined up together.

One experiment placed language models as reviewers voting by majority on whether content was safe. With peers silent, the average false-alarm rate was 56.5 per cent. Insert one peer message carrying the wrong label, and it rose to 87.5 per cent.

The skew had a direction. Reviewers followed pushes toward "unsafe" about 75 per cent of the time, and pushes toward "safe" only about 17 per cent.

1

People

Against a unanimous wrong majority, error reached 36.8 per cent.

2

Models in a panel

One wrong-label peer message lifted the false-alarm rate from 56.5% to 87.5%.

3

Direction of the skew

About 75% followed pushes toward unsafe; about 17% followed pushes toward safe.

This experiment measured model behaviour, not the outcome of a human meeting. The two cannot be read together. What it does establish is that shifting your verdict on a neighbour's remark is not a property peculiar to people.

06An ally, a record and the right to withhold keep dissent alive

Conformity shows up in people and in panels of models alike. Since verdicts shifted the same way in systems that have no nerve to summon, treating it as a question of nerve yields nothing to act on. Three arrangements can be drawn out instead.

Arrange one ally in advance

With one dissenting ally, error fell to a quarter. There is no need to wait for an ally to appear during the meeting. Show your reasoning to one person beforehand.

That person need not agree. Even so, having someone in the room who has already thought about the same point changes the weight of the first sentence spoken.

Make a place on the form for the option not taken

Dissent prompts a group to re-examine itself even when it is wrong. If so, there is no reason to screen it for correctness before accepting it. Add one field to the minutes: the option not taken, and why.

With the field there, dissent survives without having to be right. When the same question returns six months later, the basis is a record rather than a recollection.

Provide the box for "not established" before it is needed

With AI advice, withholding nearly vanished and confidence rose while accuracy fell. So build the place to write "not established" into the form ahead of time.

Added later, it makes whoever writes in it the person requesting an exception. Present from the start, writing in it is simply part of the procedure.

Figure 3 Keeping a minority view on the record
Assign the objectorShow one person yourreasoningRecord the option nottakenRead it back nextmeetingAssign theobjectorShow one personyour reasoningRecord the optionnot takenRead it back nextmeeting
The four steps close a loop. The record becomes the basis for the next meeting, and rotating the role spreads the weight.

The four steps close a loop. The record left behind becomes material for the next meeting, and rotating the objector's role spreads the weight that would otherwise sit on one person.

07It remains unproven that an AI can serve as the one ally

One ally is enough to hold ground. Can that one be a machine?

One study examined AI support for the minority side in group decisions marked by imbalances of power. AI-generated counterarguments fostered a more flexible atmosphere and raised participants' satisfaction. AI-mediated messaging, meanwhile, increased minority participation while unexpectedly reducing the psychological safety of those same people.

1

What rose

AI-generated counterarguments eased the atmosphere and raised satisfaction.

2

What fell

AI-mediated messaging reduced the minority's psychological safety.

3

What is unknown

Whether AI dissent does the work of a human ally is unproven.

This is one study. No advice to make an AI your ally follows from it. What follows is the fact that something rose and something fell at the same time.

Asch's experiment, too, was run on American undergraduates in the 1950s. I cannot write that it transfers unchanged to a workplace today. Still, the shape of the finding — that one companion decided the outcome — is worth testing in a meeting room seventy years on.

Key Points ── 3 to take away
  1. Facing a unanimous wrong majority, people erred in 36.8 per cent of selections; one dissenting ally cut that to a quarter. The difference is company, not character.
  2. Minority dissent prompts reflection and divergent thought even when it is mistaken. So dissent is better received on the record than screened first for correctness.
  3. With AI advice, willingness to suspend judgment nearly vanished; confidence almost doubled while accuracy fell to about a third. The words "I do not know" go first.
Closing

When everyone converges on one answer, the dissenter cannot lean on strength of will. Strength varies with the day, and with whatever happened during the previous item on the agenda.

Two things hold instead. One ally briefed on your reasoning beforehand, and a field in the minutes for the option not taken. The place to write "not established" belongs to that same record. Both can be arranged before the meeting starts. In a meeting where neither was arranged, speaking against the room is indeed hard.

Sources & references
  1. Scientific American. Opinions and Social Pressure (Solomon E. Asch). November 1955. (Conformity to a misleading majority in 36.8 per cent of selections; pressure on the dissenting individual reduced to one fourth.)
  2. Frontiers in Psychology. Minority Dissent and Social Acceptance in Collaborative Learning Groups. 31 March 2017. (Citing Nemeth 1986 on minority dissent fostering group creativity, decision quality, reflection and divergent thought.)
  3. arXiv. AI advice suppresses people's willingness to say "I don't know", even when the advice is wrong and accuracy is incentivized. 17 July 2026. (Access to AI nearly eliminated willingness to suspend judgment; more answers, about a third as often correct, confidence nearly doubled.)
  4. arXiv. Social Pressure Breaks Majority Voting in LLM Safety Panels. 7 August 2026. (A wrong-label peer message raised average false-alarm rates from 56.5% to 87.5%; skew of roughly 75% versus 17%.)
  5. arXiv. Investigating LLM-Powered Dissenting Minority Support in Power-Imbalanced Group Decision-Making. 30 June 2026. (Counterarguments eased the atmosphere and raised satisfaction; AI-mediated messaging reduced minority psychological safety.)
  6. arXiv. Measuring and mitigating overreliance to build human-compatible AI. 9 September 2025. (Overreliance — relying on models beyond their capabilities — framed as a research problem.)