A Shared Mind Needs Dissent

The most dangerous AI system in a bank may not be the one that is obviously wrong.

It may be the one that is plausible, polished, and present in every meeting.

Imagine credit, risk, compliance, treasury, and the board using the same model family, the same retrieval layer, and the same approved enterprise assistant. The organization appears to have gained a shared mind. Information moves faster. Briefings become consistent. Decisions arrive with neat explanations.

But if the same system frames the problem, selects the evidence, summarizes the objections, and drafts the recommendation, apparent alignment can hide intellectual concentration.

A shared mind becomes groupthink when it eliminates the independence that makes collective intelligence valuable.

Banks therefore need more than human oversight. They need dissent by design: independent analysis, visible uncertainty, protected minority views, and people with the authority to stop a confident consensus.

Consensus is not collective intelligence

Collective intelligence is often confused with efficient agreement. They are not the same.

A group becomes intelligent when its members contribute different information, models, experience, and ways of testing a claim—and when the process can distinguish a better explanation from a louder one. If everyone begins from the same assumptions, more participants may simply reproduce the same error.

Large language models are unusually good at compression. They can turn thousands of pages and many contributions into a coherent narrative. That is valuable. It is also exactly where dissent can disappear. A caveat becomes one sentence. A minority hypothesis is labelled an edge case. Contradictory evidence is reconciled into a smooth answer before a decision-maker sees the conflict.

A multidisciplinary Nature Human Behaviour perspective on LLMs and collective intelligence identifies both sides of this transformation. Language models can lower barriers to participation, help coordinate knowledge, and synthesize views at scale. They can also reshape how information is aggregated and transmitted, creating risks that deserve examination rather than an automatic assumption that better synthesis means better judgement.

The goal should not be frictionless consensus. It should be productive disagreement that converges only after the evidence has been examined.

The same answer can spread everywhere at once

Model concentration is usually discussed as a vendor or operational-resilience problem. It is also an epistemic problem: many institutions can inherit the same blind spot.

Within one bank, a general-purpose model may support research, customer service, coding, compliance, and management reporting. Across the market, several banks may rely on the same foundation model, data provider, or implementation partner. Outputs will be adapted, but the underlying patterns can remain correlated.

That matters because financial risk is rarely independent. Institutions watch one another. Markets react to shared narratives. A model trained on the same history and optimized around familiar assumptions may reinforce the interpretation already embedded in prices, policies, and internal data.

The ECB has warned that excessive dependence on AI could create echo chambers, repeat existing biases, suppress originality of thought, and increase groupthink and confirmation bias. Its central-bank view of AI emphasizes comparing alternative interpretations and assessing risk from multiple perspectives.

This extends the question posed in When Every Bank Has the Same AI, What Still Matters?. Context and workflow still matter, but so does independence. If every team asks the same model to challenge the same model, the bank has not created a second opinion. It has created a second prompt.

Adding a human does not automatically add a different view

The standard answer to AI risk is to put a person at the end of the process. That person sees the recommendation and approves or rejects it.

This sequence can create deference rather than challenge. The AI speaks first. It frames the case, organizes the evidence, and anchors the expected answer. The reviewer must generate an alternative while explaining why the polished recommendation is wrong. Under time pressure, approval becomes easier than independent thought.

A 2024 systematic review and meta-analysis in Nature Human Behaviour examined 106 experiments that compared humans alone, AI alone, and human-AI combinations. On average, the combinations performed better than humans alone but worse than the better of the human or AI operating alone. Decision tasks showed performance losses on average. The authors caution that results varied widely and depend on task and design, but the finding punctures a convenient assumption: combining a person with a model does not itself produce synergy.

Real oversight begins before the AI's conclusion becomes the default. The reviewer may need to record an initial judgement first, see evidence before recommendation, or compare outputs produced by genuinely different methods. As “Human in the Loop” Is Not a Governance Model argues, oversight requires context, authority, triggers, and time. It also requires cognitive independence.

Banking already has a language for challenge

This is not a foreign concept. Model-risk management has long recognized that validation must be more than a ceremonial sign-off.

The Federal Reserve's guidance on model risk management calls for “effective challenge”: critical analysis by objective, informed people with appropriate expertise, independence, incentives, standing, and influence to produce change. The formulation is powerful because it combines intellectual and organizational conditions. A critic who is correct but powerless is not a control.

The same principle should govern AI-assisted decisions. Challenge should come from people who did not build the workflow, are not rewarded for its adoption, and can ask whether the system should be used at all. It should also come from evidence outside the model's preferred frame: customer complaints, operational exceptions, frontline experience, alternative data, scenario analysis, and deliberately different models.

NIST makes this explicit in the AI Risk Management Framework Core. It recommends diverse and multidisciplinary perspectives, a culture of critical thinking, engagement with people outside the development team, and independent assessors where appropriate. Diversity here is not decorative representation. It is a mechanism for exposing hidden assumptions and emergent risks.

Give dissent rights, not just airtime

An invitation to “speak up” is weak governance. The process must preserve disagreement and attach rights to it.

For a material AI-assisted decision, I would use a six-part dissent protocol:

Independent first views. Ask relevant experts to record their assessment before seeing the model's recommendation or the senior sponsor's position. Independence must exist at the start, not be reconstructed later.
Plural methods. Compare at least one materially different source of judgement: a rules-based benchmark, another model architecture, an external view, a manual sample, or a team using different data. Multiple agents built on the same foundation model do not count as diversity.
Visible uncertainty. Show confidence ranges, missing data, disputed assumptions, and plausible alternatives. Do not let the executive summary average away disagreement.
Named challenger. Assign someone responsible for building the strongest credible case against the preferred decision. Give that person access, time, status, and no adoption target that conflicts with the role.
Stop and escalate rights. Define which evidence, uncertainty, or minority objection pauses action and who adjudicates it. The EU AI Act's human-oversight requirements for high-risk systems specifically address awareness of automation bias and the ability, as appropriate, to disregard, override, or reverse output. Those abilities must be operational, not theoretical.
Dissent memory. Store rejected hypotheses and the reasons they were rejected. Revisit them when outcomes arrive. A minority view that repeatedly proves useful should change the process; one that never does should become more precise.

This protocol does not mean every decision becomes a debate. The intensity should follow materiality. A low-risk drafting tool needs lightweight checks. A system influencing credit, suitability, fraud intervention, capital, or market exposure needs stronger independence.

Measure convergence before celebrating it

Organizations measure AI accuracy and adoption. They should also measure whether AI is making their thinking dangerously uniform.

Track how often reviewers disagree before and after seeing an AI output. Compare error correlations across models and teams. Measure overrides, not only their frequency but their outcomes. Look for departments producing unusually similar language, sources, and risk estimates. Test whether the system surfaces minority evidence or systematically compresses it. Run scenarios in which the dominant assumption is deliberately wrong.

Most importantly, separate the quality of a decision process from the convenience of its output. A faster committee pack is useful. Ten executives repeating its conclusion is not evidence that the conclusion improved.

This is where an AI agent's job description should include epistemic boundaries. Is the agent collecting evidence, proposing a decision, critiquing another analysis, or acting? A system asked to do all four will tend to validate its own frame.

The shared mind should contain an argument

The promise of AI is not that every person in an institution will think the same thought more quickly. It is that the institution can connect more evidence, more expertise, and more perspectives than any individual could hold alone.

That promise depends on architecture. Preserve provenance. Keep alternative hypotheses visible. Prevent one model from becoming the author and judge of every decision. Give challengers authority. Learn from overrides and near misses. Reward people who improve a decision by disagreeing with it.

A bank does not become intelligent when every screen displays the same answer.

It becomes intelligent when it can hold several possible answers, test them honestly, and change its mind before reality changes it by force.