“Consensus is not truth: what medicine teaches us about building AI that protects human freedom”

Artificial intelligence is often presented as a story of speed, scale and efficiency. Faster answers. Faster administration. Faster workflows. Faster decisions. But in medicine, speed without judgement can be dangerous. Some decisions should not be accelerated until they have first been properly understood. This matters because AI is no longer confined to convenience or entertainment. It is entering medicine, law, education, finance, cybersecurity, public administration and political life. These are domains where confident mistakes can carry serious consequences, and where the structure of reasoning matters as much as the answer itself. As a surgeon, I approach this question from the operating room and from the multidisciplinary cancer meeting. Medicine has spent decades learning something that AI must now learn quickly: Consensus is useful. But consensus is not the same as truth. From one doctor to the MDT – and back to one robot? Modern medicine gradually moved from individual expert opinion toward multidisciplinary decision-making. That was progress. A single doctor, however experienced, may miss something. A surgeon may focus on whether something can technically be removed or repaired, while a generalist asks whether it should be removed at all. A radiologist may identify subtle imaging patterns that change surgical risk. A pathologist may be cautious about a borderline diagnosis. An oncologist may think in probabilities and long-term outcomes. A specialist nurse may understand the patient’s psychological resilience, family support structure and personal priorities. The multidisciplinary team, or MDT, exists because complex problems are better approached when multiple independent perspectives analyse the same question from different angles. Medicine moved from centralised reasoning toward decentralised reasoning. But this only works when disagreement is allowed to surface. The value of an MDT is not that everyone agrees. The value is that assumptions are challenged before irreversible decisions are made. If the future of medical AI simply moves from: one doctor → multidisciplinary team → one robot then that is not progress. It is regression disguised as innovation. A single large language model may sound balanced and authoritative. It may summarise guidelines beautifully and produce polished answers in seconds. But if it collapses uncertainty into one confident response, it risks recreating the very weakness that multidisciplinary medicine was designed to overcome. Medicine moved beyond the single expert for a reason. AI should not take us back there. Why false consensus is dangerous In high-stakes environments, the greatest danger is often not open disagreement. Open disagreement can be examined and challenged. The greater danger is false consensus. False consensus occurs when uncertainty remains present but becomes invisible. It happens when discussion moves too quickly, hierarchy silences doubt, guidelines are applied mechanically, or minority concerns are not voiced strongly enough. In medicine, this can change everything. A borderline lymph node. An indeterminate thyroid cytology. A frail patient with competing risks. A cancer that may require aggressive treatment – or may be safely observed. A guideline that fits the average patient but not the individual sitting in front of you. In these situations, truth is not found by averaging opinions. It is approached by testing assumptions. That is where AI must become more sophisticated. The question is not whether AI can produce answers. It can and always will. The question is whether AI can help humans understand the uncertainty behind those answers. Sovereignty is not only about data In the digital world, sovereignty is often discussed in terms of infrastructure: where data is stored, who controls access, which legal framework applies, and how systems are secured. Those questions matter. But sovereignty is not guaranteed simply because data is stored locally. A society may host its own servers and still remain intellectually dependent on opaque external models, imported institutional logic or centrally filtered representations of reality. True AI sovereignty therefore cannot only mean control over hardware and data. It must also mean sovereignty of reasoning. The purpose of digital systems should not merely be to process information eAiciently, but to protect the conditions under which human beings can think independently, disagree openly, preserve individuality and participate in reasoned debate. Freedom of thought, autonomy and structured disagreement are not obstacles to progress. They are foundations of progress itself. This matters because bias in AI is not always obvious. It does not only appear as factual error. It may appear as omission, over-smoothing, excessive deference to consensus, or the quiet disappearance of minority interpretations. Every large model reflects choices: what it was trained on, what was excluded, which values were embedded and which assumptions became invisible through repetition. A single model therefore does not only centralise computation. It may centralise worldview. The economist Friedrich Hayek warned that complex systems cannot be properly understood, let alone controlled, from a single central point. His insight was that knowledge in complex systems is distributed. No single institution, expert or authority possesses all the relevant information. Individuals respond to local realities, partial signals and constantly changing conditions that cannot be fully aggregated into one central intelligence. Attempts to centralise that knowledge may destroy the very diversity of information that makes intelligent adaptation possible. Nature offers similar examples. Watch a murmuration of starlings. Thousands of birds move together in fluid, coordinated patterns that appear almost choreographed. No bird directs the flock. No central intelligence issues commands. Each bird responds only to the movement of its immediate neighbours, following a small number of simple rules. The result is not chaos It is a form of collective intelligence that no single bird , and no external controller, could produce alone. Hayek saw similar principles at work in markets, institutions and societies. Order emerging from distributed independent reasoning is often more adaptive and more resilient than order imposed from above. A single AI model that resolves disagreement into one confident answer is not a murmuration. It is one bird flying alone, telling the others where to go. One way to reduce this risk is pluralism by design. A multi-agent reasoning system, particularly one in which independent agents reason from separate models, datasets or institutional contexts, can make bias more visible. If agents challenge one another and preserve disagreement, the system becomes less dependent on one model’s assumptions or one institutional worldview. This does not eliminate bias. But it can make bias more contestable, inspectable and difficult to hide behind a fluent answer. When disagreement is the signal In complex medicine, disagreement is not always a problem. Sometimes disagreement is the signal. This is the central principle behind multi-agent clinical reasoning. The purpose of such a system is not to replace clinicians or clinical governance. Its purpose is to make reasoning more visible, structured and honest about uncertainty. Most AI systems optimise for agreement. A more useful approach in complex domains is to preserve independent reasoning before it is synthesised. Multiple reasoning agents can analyse the same clinical case independently before seeing one another’s conclusions. They can then challenge assumptions, identify missing information, test weak logic and raise alternative pathways or overlooked risks. The outputs should not simply be averaged. There should be no automatic majority-vote rule. If three agents analyse a case and two agree, that does not necessarily make the majority correct. In medicine, a minority concern may identify the missing diagnosis, the unsafe assumption or the overlooked patient factor that matters most. This principle extends far beyond medicine. John Stuart Mill and Alexis de Tocqueville both warned that majority opinion can itself become a form of tyranny when it suppresses dissent and individuality. A majority can be useful. It can also be wrong. If a reasoning system treats convergence as truth, it may simply automate the tyranny of the majority. That is why future AI systems should preserve minority risk signals rather than smooth them away. The role of the moderator A multi-agent system is only useful if disagreement becomes structured. Unstructured disagreement is noise. Suppressed disagreement is dangerous. Structured disagreement can become intelligence. In a multi-agent reasoning system, the moderator should not choose a winner or hide disagreement behind a final central answer. Its role should be to organise reasoning. Where do the agents agree? Where do they disagree? Why do they disagree? Which assumptions are fragile? What information is missing? Which minority concern may represent a genuine safety signal? The moderator’s purpose is not to remove uncertainty. Its purpose is to make uncertainty usable. The goal is not an autonomous tumour board or a machine that replaces clinicians. The goal is better human reasoning. Truth approximation through structured disagreement In medicine, absolute truth is rarely available at the moment decisions must be made. We work with evidence, experience, imaging, pathology, patient preference, uncertainty and risk. Truth is approximated by testing these elements against one another. That is also how science progresses. Progress is not created by consensus alone. It emerges through hypothesis, challenge, contradiction, revision and refinement. George Bernard Shaw expressed this sharply when he wrote: “The reasonable man adapts himself to the world; the unreasonable one persists in trying to adapt the world to himself.  Therefore all progress depends on the unreasonable man.” That does not mean every contrarian is right. It means progress requires space for dissent. In medicine, this may mean challenging a diagnosis or questioning whether a guideline truly applies to the individual patient. A system that gives one answer may be useful for simple tasks. But in complex domains, the stronger system may be the one that can present several plausible interpretations and explain why they diAer. Truth approximation does not require universal agreement. It requires disagreement to remain visible, structured and accountable. What this means beyond medicine The same principle applies beyond healthcare. AI systems are entering public services, cybersecurity, education, business transformation and institutional decision-making. In each of these fields, the danger is not only technical failure. It is institutional overconfidence. An AI system used in public administration should show what data it relied on, where uncertainty remains and where human judgement is still required. A cybersecurity system should not suppress minority threat signals because most indicators appear normal. An educational system should not treat children as statistical averages. Small sovereign environments such as Andorra may have a particular opportunity here. Not because small countries automatically build better technology, but because governance, entrepreneurship, medicine and civic responsibility can sometimes remain closer together. Smaller jurisdictions can test carefully, adapt quickly and build governance alongside innovation rather than after it. The principles we should build into the foundation As AI becomes more deeply embedded in society, three principles should matter. The first is transparency over fluency. A polished answer is not enough. Systems should show what they know, what they do not know and where uncertainty remains. The second is human responsibility over machine autonomy. AI should support the judgement of clinicians, teachers, entrepreneurs, public servants and citizens. Responsibility should remain human and inspectable. The third is structured disagreement over false consensus. If systems hide disagreement, they may hide the signal that matters most. If they preserve disagreement intelligently, they can improve human reasoning. These are not cosmetic design choices. They are foundational choices. The cool-down period before action In medicine, there is often a moment before a major operation or irreversible treatment when discussion has taken place, evidence has been reviewed, and both clinicians and patients must ultimately commit. That moment matters. Not paralysis. Not indecision. A deliberate cool-down period before irreversible action. This pause does not appear by accident. It is created by the structure of reasoning that preceded it. When disagreement has been surfaced, examined and recorded – when minority concerns have been voiced rather than smoothed away – the people making the final decision know what they are committing to and what remains unresolved. Structured disagreement is not an obstacle to decision. It is what makes the decision honest. The most dangerous systems are not the ones that move slowly. They are the ones that move confidently into irreversible action without having preserved space for reflection. As AI enters more areas of human life, society itself is entering such a moment. Infrastructure matters. Data quality, cybersecurity, privacy and traceability all matter. But the deeper question is what kind of intelligence we choose to build. The future of AI should not be the replacement of human reasoning. It should be the disciplined expansion of it. The purpose of technology should not be to automate conformity. It should be to strengthen freedom, individuality, responsibility and the human search for truth. Dr Volkert Wreesmann MD PhD Dr Volkert Wreesmann MD PhD is a surgical oncologist and founder of ANDSURGEONS, an international cross-border surgical network headquartered in Andorra. He is also the founder of an AI startup exploring multi-agent reasoning and structured disagreement as a basis for improved truth approximation in medicine. His work focuses on complex surgical decision making, clinical governance, human judgement under uncertainty, and the design of AI systems that support rather than replace professional reasoning.The post “Consensus is not truth: what medicine teaches us about building AI that protects human freedom” first appeared on All PYRENEES.