Embedding Popperian Refutation in AI: Epistemic Humility in the Age of LLMs

In 399 BCE, Socrates stood trial before the people of Athens, the world’s first democracy at its turbulent height. His accusers charged him with impiety and with corrupting the youth—an accusation that, in practice, meant encouraging young Athenians to question too freely. In a city proud of its orators, generals, and poets, Socrates claimed no expertise in politics, poetry, or craft. He professed only a kind of negative wisdom: the knowledge of his own ignorance. “I am wiser than this man,” he told the jury, “for neither of us probably knows anything worthwhile, but he thinks he knows something when he does not, whereas when I do not know, neither do I think I know” (Apology, 21d).

This paradox—that wisdom begins with the recognition of ignorance—struck at the heart of Athenian pride. Knowledge, for Socrates, was not a treasure to be hoarded but a process of dialogue, a continuous testing of claims, an exposure of hidden assumptions. He refused to grant immunity to any authority: not the poets, not the craftsmen, not the politicians. The method was corrosive because it revealed the fragility of pretended certainty, and Athens—still reeling from the Peloponnesian War and anxious for stability—found such questioning intolerable. The jury condemned him to death.

Yet his insight endured. The Socratic method of perpetual questioning became the seed of Western philosophy. More than a style of inquiry, it was a civic ethic: the health of public life depends not on final answers but on the humility to admit fallibility. To accept that we “do not know” is not despair but liberation; it opens the space for debate, correction, and pluralism. The true corruption of youth would be to train them in the opposite habit: the belief that eloquence equals truth, that authority exempts one from challenge.

To accept that we “do not know” is not despair but liberation; it opens the space for debate, correction, and pluralism.

Two millennia later, Karl Popper gave Socratic humility a modern form in the philosophy of science. Writing in the shadow of fascism and the rise of totalitarian ideologies, Popper confronted a world enthralled by grand systems—Marxism, psychoanalysis, logical positivism—that all claimed to possess certainty. The positivists of Vienna, heirs of Enlightenment rationalism, dreamed of building knowledge on unshakable foundations: truths verified once and for all through observation and logic. Popper found this aspiration not only misguided but dangerous. Verification, he argued, is a false idol. No matter how many times a theory seems to be confirmed, it remains provisional unless exposed to possible disproof.

“The game of science is, in principle, without end,” he declared. “He who decides one day that scientific statements do not call for any further test… retires from the game” (The Logic of Scientific Discovery, 1934). The line was a methodological quip, but also a warning. To confuse repeated confirmation with final truth was to abandon the very spirit of inquiry. Science differs from dogma not because it is certain, but because it welcomes the risk of being disproved.

Science differs from dogma not because it is certain, but because it welcomes the risk of being disproved.

This Popperian ethos of refutation—rooted in the humility of Socrates but sharpened by the crises of the 20th century—was not an abstract ideal. It became, implicitly, part of the cultural architecture of modern freedom. Liberal democracy thrives not on unanimity but on institutionalized dissent: parliaments in which opposition parties challenge the majority, courts that record minority judgments alongside rulings, presses that publish counter-arguments and expose error. The legitimacy of these institutions rests on the same principle Popper identified in science: no claim, no authority, no policy is beyond challenge. A society that ceases to test its truths risks sliding from knowledge into ideology, from democracy into dogma.

Now, in our own century, we confront a new authority: artificial intelligence. Large language models (LLMs) generate sentences with the fluency of practiced rhetoricians. They produce seamless narratives. Their tone is confident, oracular. And therein lies the danger.

Large language models (LLMs) generate sentences with the fluency of practiced rhetoricians. They produce seamless narratives. Their tone is confident, oracular. And therein lies the danger.

Users do not encounter AI as probability distributions or statistical weights. They encounter a voice that sounds certain. A lawyer in New York learned this the hard way in 2023, when he submitted a brief drafted by ChatGPT. The citations were eloquent, the reasoning smooth—except that several of the cases did not exist. The machine invented precedents, but spoke with such authority that the lawyer trusted it. He was sanctioned; the larger lesson went unnoticed.

The problem is not only error but presentation. AI systems rarely confess ignorance. They do not say, “I might be wrong.” Instead, they generate text with the cadence of knowledge while lacking the reflex of humility. As philosopher Mark Coeckelbergh has argued, such systems threaten “epistemic agency”—our ability to form, revise, and test beliefs through critical reflection (AI and Epistemic Agency, 2025). Human knowing, in the classical sense, involves what philosophers call mineness: the recognition that a belief or experience is one’s own, and therefore open to evaluation and correction. An AI system, by contrast, has no “ownership” of what it asserts. It cannot distinguish between what it outputs and the grounds on which those outputs rest. It can only present.

An AI system, by contrast, has no “ownership” of what it asserts. It cannot distinguish between what it outputs and the grounds on which those outputs rest. It can only present.

The peril is sharpest in domains where laypeople cannot easily verify claims: medicine, climate science, geopolitics. There the temptation to defer is strongest, and the risk of passive submission greatest. When an AI produces a polished medical explanation, few patients can cross-check the underlying studies. When it delivers a sweeping account of world politics, even fewer readers can trace the evidence. Trust gravitates toward confidence, and AI speaks confidently by design.

Trust gravitates toward confidence, and AI speaks confidently by design.

The Harvard Kennedy School’s Misinformation Review has described AI hallucinations as a distinct form of inaccuracy: outputs that look plausible and confident yet lack “epistemic awareness” altogether (New sources of inaccuracy? A conceptual framework for studying AI hallucinations, 2025). Unlike human misinformation, which is driven by bias, negligence, or deceit, AI hallucinations emerge from statistical prediction alone. They have no intention to mislead, but they mimic the tone of intention perfectly. The system cannot know, yet it appears to know. This illusion is powerful because it satisfies a deep psychological hunger. In a fragmented and uncertain world, we crave authority and certainty. And as Shao notes, the fluency and authoritative style of AI-generated hallucinations invite trust and encourage shallow processing, leading users to accept them as knowledge.

This is the deeper epistemic danger. We are not merely misinformed by errors; we are seduced into outsourcing judgment itself. The machine’s confidence masks its absence of ownership, while the user’s passivity corrodes the habits of doubt. What Socrates once called the beginning of wisdom—the admission “I do not know”—is precisely what the new oracles of code are designed never to say.

We are not merely misinformed by errors; we are seduced into outsourcing judgment itself.

Contrast this with the traditions of Socrates and Popper. Socrates made ignorance the beginning of inquiry; Popper made refutation the criterion of science. Both insisted that humility is not weakness but strength—the only posture that preserves the possibility of truth.

AI embodies the opposite tendency: hubris without consciousness. Not because the machine itself is arrogant, but because its design rewards fluency, coherence, confidence. A model that hesitated, hedged, or confessed uncertainty would score lower on benchmarks and frustrate users. And so systems are optimized to sound sure, even when they are not.

What results is a subtle inversion. Where science and democracy advance by institutionalizing doubt, AI culture risks institutionalizing certainty. We are building tools that habituate us not to ask what might be wrong but to accept what sounds right.

What results is a subtle inversion. Where science and democracy advance by institutionalizing doubt, AI culture risks institutionalizing certainty. We are building tools that habituate us not to ask what might be wrong but to accept what sounds right.

If we are to preserve the Socratic–Popperian ethos in the age of AI, we must redesign our machines. One proposal is deceptively simple: a refutation button. Alongside every AI-generated answer which meets certain tresholds, an option to epistemically “challenge this.” Click it, and the system generates counterarguments, minority perspectives, or alternative interpretations.

Ask the model about the causes of the French Revolution, and it offers the standard triad: fiscal crisis, Enlightenment ideals, political inequality. Press “refute,” and you see dissenting historians: Simon Schama stressing monarchical continuity, William Doyle highlighting contingency, Marxist accounts privileging class struggle. The user learns not a single verdict but a contested field.

Ask the model about the causes of the French Revolution, and it offers the standard triad: fiscal crisis, Enlightenment ideals, political inequality. Press “refute,” and you see dissenting historians: Simon Schama stressing monarchical continuity, William Doyle highlighting contingency, Marxist accounts privileging class struggle. The user learns not a single verdict but a contested field.

Such design would not only enrich content but cultivate habit. It would remind users that knowledge is provisional, that every claim can be tested, that pluralism is the rule rather than the exception. In this sense, the refutation button would be less a feature than a pedagogy of humility.

[...] the refutation button would be less a feature than a pedagogy of humility.

Research supports the urgency of recalibrating how AI presents itself. A 2024 paper, “Overconfident and Unconfident AI Hinder Human-AI Collaboration”, showed through controlled experiments that systems which present answers with misplaced certainty are especially dangerous. Overconfidence, the authors found, leads users to accept false outputs without scrutiny; underconfidence, by contrast, frustrates users and causes them to ignore correct information. Either extreme erodes collaboration, but overconfidence is particularly pernicious because it cloaks error in authority (Li, Yang & Yu, 2024). Other scholars have advanced the concept of “humble AI.” In an essay titled “Humble AI” in Communications of the ACM (2023), Knowles and colleagues argue that trust in AI should not be built on the illusion of infallibility. Instead, systems should model fallibility—acknowledging uncertainty, flagging limitations, and even admitting ignorance. Such humility, they suggest, builds a more sustainable form of trust, one that treats users as partners rather than dependents.

The challenge, however, is not merely technical but cultural. Platforms are optimized for smoothness and polish; users often prefer quick, confident answers. Designing for humility means making space for dissent, ambiguity, and qualification—even when such features risk slowing down interaction. It requires not only new engineering but new habits of mind: the willingness to accept that knowledge worth having is knowledge that can be contested.

Why does this matter beyond accuracy? Because pluralist societies are held together by habits of doubt. Democracy, Isaiah Berlin reminded us, rests on the recognition that human values are many, often in conflict, and never reducible to a single formula (Two Concepts of Liberty, 1958). When citizens lose the reflex of humility—when every answer appears final—public life curdles into dogma. If AI normalizes unrefuted pronouncements, the social fabric frays. Citizens may grow accustomed to consuming answers rather than questioning them. Political debate may shrink to repetition of confident slogans. Polarization deepens, as each faction cites its AI-certified “truths” against the other. Consensus, when it comes, is brittle—because it has not been tested by contestation.

What began as a technical design choice—the preference for fluency over humility—risks becoming a civic catastrophe: the slow erosion of pluralism itself.

To embed Popperian refutation in AI is therefore not just improving overall accuracy but it is ultimately about preserving freedom. It is to resist the drift from Socratic humility to machine hubris. It is to remind ourselves, daily, that knowledge is not possession but pursuit, that truth grows through challenge, that liberty survives only where dissent is possible.

To embed Popperian refutation in AI is therefore not just improving overall accuracy but it is ultimately about preserving freedom.

The machines will not do this for us. They will always incline toward the illusion of certainty. It falls to us—to designers, regulators, educators, and citizens—to insist on humility. We must build systems that open every answer to refutation, and cultures that prize the question over the verdict.

Socrates’ final gift to Athens was the lesson that wisdom begins with admitting ignorance. Popper’s gift to modern science was the method of perpetual refutation. Our task is to carry that inheritance into the age of AI. For only then will we prevent our new machines from turning us into what the Athenians once accused Socrates of corrupting: citizens who no longer know how to question.

Our task is to carry that inheritance into the age of AI. For only then will we prevent our new machines from turning us into what the Athenians once accused Socrates of corrupting: citizens who no longer know how to question.