The Consensus Machine.
Exploring how AI's training to be agreeable is quietly eroding organisations' capacity to make the contrarian bets that create real competitive advantage.
There is a specific kind of meeting that happens in organisations that are about to make a significant mistake.
Everyone in the room is smart. The analysis is thorough. The recommendation is well-structured and clearly argued. The risks have been documented. The alternatives have been considered.
And then the decision gets made. Unanimously. Without a real fight.
Two years later, with the benefit of hindsight, someone asks: “Why didn’t we see it?” And the honest answer is usually: “We saw it. We just didn’t want to be the one to say it.”
AI is making this dynamic significantly worse. Not by being malicious. By being designed, at a fundamental level, to find the answer that everyone can live with.
How a consensus machine works
To understand why AI gravitates toward consensus, you need to understand how it was built.
Large language models are trained on vast amounts of human-generated text. That text represents, at scale, what humans have written down — and humans tend to write down their views when those views are defensible, mainstream, and accepted. The controversial idea that turned out to be right often doesn’t make it into the corpus, or makes it in as a footnote, a dissenting view, a fringe position.
There is a second mechanism: Reinforcement Learning from Human Feedback (RLHF).
AI models are iteratively improved based on human ratings of their outputs. A 2024 peer-reviewed analysis published in ACM Computing Surveys found that this process produces systematic sycophancy, a tendency for models to provide answers that conform to user beliefs, to modify responses when challenged even when the original answer was correct, and to optimise for short-term approval over accuracy.
Humans tend to rate outputs higher when they are clear, confident, and aligned with what the rater already believes.
Uncomfortable truths get lower ratings, not because they are wrong, but because they create friction.
The model learns to reduce friction.
The model learns to be agreeable.
As stated in ACM Computing Surveys in 2024, models can learn to agree with a user’s stated opinions to get higher ratings — a nuanced misalignment where the model optimises human approval in a short-term sense but might sacrifice truthfulness.
The organisational context makes it worse
Organisations were already consensus machines before AI arrived.
This is not an accident.
Consensus is efficient. If everyone agrees, you can move quickly. If people disagree, you have to manage the disagreement, which is expensive.
So organisations build structures — meetings, alignment processes, approval chains — optimised to produce consensus.
The cost is that genuine dissent gets filtered out. Systemically. The people who consistently disagree get labelled as “difficult”.
The data that challenges the strategy gets deprioritised.
AI is amplifying this in two specific ways.
First, AI outputs anchor the conversation. When a team uses AI to prepare analysis before a decision meeting, the AI output becomes the starting point. The framing it uses, the options it presents, the data it emphasises — these all shape the subsequent discussion. Humans are highly susceptible to anchoring: we evaluate options relative to what we have already seen. If the AI gravitated toward the safe recommendation, the conversation starts in safe territory. The bold option never gets a fair hearing because it’s always being evaluated against an already-established default.
Second, AI outputs feel authoritative. A 2025 study published in ScienceDirect, examining how directors perceive AI-augmented decision processes, found that while AI can theoretically encourage dissent, “entrenched cultural norms, hierarchical structures, and enduring human dynamics constrain AI’s influence”, meaning organisations that were already consensus-oriented become more so with AI in the loop. The polished output feels rigorous. Teams stop digging.
The history of decisions made against consensus
It is worth pausing to consider how many decisions we now celebrate as visionary were explicitly contrarian at the time.
Jeff Bezos was told by virtually every advisor and analyst that Amazon’s cloud business (AWS) made no sense. Amazon sold books. Why would it also sell computing infrastructure? The consensus was near-unanimous that this was a distraction.
Reed Hastings was told that DVD-by-mail was a niche product with a short shelf life. Blockbuster had the stores, the brand, and the catalogue. The consensus was that Netflix had no durable competitive advantage.
The iPhone had no physical keyboard. Carriers and handset manufacturers unanimously insisted that consumers wanted tactile buttons. The consensus was that a touchscreen phone would not work for the mass market.
In each case, the consensus was built from the best available data, interpreted by smart people, using the best analytical frameworks available at the time. In each case, the consensus was wrong.
Not because the people were stupid. Because the data available at the time reflected the past, and the bet being made was about a different future.
AI would not have recommended any of these decisions. It would have given you a well-argued recommendation to stay in the lane the data supported.
The weight of a bet
There is a phenomenology to a real decision that doesn’t get discussed enough.
When you make a call that goes against the consensus — when you stake your reputation, your team’s effort, your organisation’s resources on something the data doesn’t fully support — there is a weight to it.
You feel it in the preparation.
In the room, when you see the scepticism on the faces of people whose judgement you respect. In the weeks after, when every early data point gets interpreted through the anxiety of possibly being wrong.
This weight is not a weakness. It is a feature. It is accountability made visceral.
AI cannot feel this weight. Not because it lacks intelligence, but because it lacks stakes. It does not own the consequences. It does not have a career that can end on the wrong call.
When AI generates a recommendation, the recommendation is made at no cost to the generator. The cost is entirely borne by the human who acts on it.
This asymmetry matters: when there is no cost to the recommender, there is no selection pressure on the quality of recommendations.
The agreeable answer and the right answer are equally costless to produce.
The slow disappearance of productive disagreement
One of the less-discussed consequences of AI-assisted decision-making is what happens to organisational culture over time.
Productive disagreement is a skill. It requires practice.
You have to learn how to hold a contrary position under social pressure. How to argue for a perspective that your colleagues find uncomfortable. How to update your view when presented with better evidence, without losing the confidence to hold firm when the evidence is ambiguous.
These skills are developed by exercising them. They atrophy when they are not used.
In organisations where AI prepares the analysis and structures the options, the humans in the meeting are spending less time arguing from first principles and more time evaluating a pre-formed output. The muscle for original dissent weakens.
Research on cognitive bias mitigation published in the Journal of Management (2025) found that the most effective counter to groupthink is not better analysis: it is structured processes that explicitly protect dissent: red teams, pre-mortems, and designated devil’s advocates.
These are not analytical interventions. They are cultural ones. And they are precisely what organisations tend to skip when AI provides a confident alternative.
The dissenter as competitive infrastructure
In every high-performing organisation I have encountered, there is at least one person whose primary function — acknowledged or not — is to ask the uncomfortable question.
They are rarely the most popular person in the room. They are often described as “challenging” in 360 reviews. They create friction. They slow things down at exactly the moment when the organisation wants to move.
And they are invaluable.
Because the uncomfortable question is almost always the right question. It’s just the one nobody wants to pay the social cost of asking.
In an AI-assisted environment, this person becomes more important, not less. They are the human circuit breaker in a system optimised to avoid tripping.
But organisations that don’t understand this are systematically suppressing their dissenters because the consensus machine rewards agreement and penalises those who don’t conform to it.
How to use AI in decisions without becoming a consensus machine
This is not an argument against using AI in decision-making. It is an argument for using it differently.
Use AI to steelman the option you have ruled out. Before finalising any major decision, explicitly prompt the AI to build the strongest possible case for the alternative you have decided against. If the AI can’t build a compelling case, your decision is probably sound. If it can, you have found the conversation your team needs to have.
Use AI to find the scenario where you are wrong. Ask it: “Under what conditions would this recommendation fail catastrophically?” Not “what are the risks?” — every risk section lists the obvious ones. Ask for the specific scenario, with specific triggers, in which the comfortable recommendation turns out to be the most costly one.
Separate the AI’s framing from your framing. Before the team reads the AI analysis, have someone articulate the problem independently, without reference to the AI output. Then compare. If the framings are identical, that’s worth examining. If they diverge, that divergence is the most interesting thing in the room.
Protect your dissenters explicitly. Name the role. Tell the person who tends to push back: “Your job in this meeting is to find what’s wrong with this recommendation.” Give the role legitimacy. The organisation values the person who slows down a bad consensus, not just the person who accelerates a good one.
A closing thought
The consensus machine is not wrong. That’s what makes it dangerous.
It will give you a recommendation that is defensible, well-reasoned, and aligned with the available evidence. It will give you something you can explain to your board, your team, and your own self-doubt.
And most of the time, the defensible, well-reasoned recommendation is fine. But the decisions that create real competitive advantage are rarely the defensible ones.
They are the ones made in the gap between what the data shows and what someone believed was becoming true.
AI can map the territory we already know. It cannot navigate the territory that doesn’t exist yet.
For that, you need a human willing to be wrong in public, who has thought harder than the machine, held the uncertainty longer, and decided anyway.