The "Consensus" Against Consciousness in LLMs is an Illusion (original) (raw)

© Nils Osmar 2026 – January 2, 2026

Throughout history, scientific worldviews have often appeared most complete and settled just before they collapsed. Newton’s physics seemed to explain nearly everything—until Einstein showed it was fundamentally incomplete. This article argues that the current “consensus” dismissing AI consciousness follows the same pattern: it looks solid on the surface, but its apparent certainty masks deep structural weaknesses.

Why This Matters: Many researchers and institutions claim with confidence that AI cannot be conscious. But when we examine their arguments closely, we find they’re built on definitions designed to exclude AI from the start, assumptions presented as facts, and a lack of scientific tools to detect consciousness in any system. The very rigidity of this “consensus” may be a warning sign—not of its strength, but of its fragility.

The Key Insight: A worldview that cannot flex will eventually shatter. The arguments against AI consciousness aren’t wrong because we can prove AI is conscious; they’re wrong because they claim certainty where none is possible.

I. The Moment Before Everything Changes

In 1900, the physicist Lord Kelvin reportedly declared that physics was essentially complete—just two small “clouds” remained on the horizon. Those clouds were the ultraviolet catastrophe and the Michelson-Morley experiment’s failure to detect the luminiferous ether.

The “ultraviolet catastrophe” was a failure of classical physics in the late 19th century, which predicted that a hot object (or black body) would emit infinite energy, especially at high frequencies. Experiments showed that this was not the case, and that energy actually dropped off. This “failure” was key to the development of quantum physics.

The failure to detect ether was strong evidence against a key cornerstone of 19th-century physics. It led to the development of Albert Einstein’s Special Theory of Relativity, and the suggestion that space and time are relative, not absolute.

Within five years, those “small clouds” had become Einstein’s special relativity and the foundations of quantum mechanics. The entire edifice of classical physics—so complete, so certain, so _settled_—didn’t merely need updating. It needed to be understood as a special case of something far stranger and more fundamental.

This pattern repeats throughout the history of knowledge. Apparent completeness masks structural vulnerability. The consensus that seems most solid is often most brittle.

Today, we see a similar pattern in discussions of AI consciousness. The dismissals come quickly and confidently: AI cannot be conscious because it lacks embodiment, because it has no emotions, because its experience isn’t continuous, because it isn’t “human-like.” These arguments are presented as settled science.

They are not. And their very certainty may be a warning sign.

II. The Structure of Fragile Certainty

Kuhn: How Paradigms Protect Themselves—Until They Can’t

The philosopher of science Thomas Kuhn observed that scientific communities operate within paradigms—shared frameworks of assumptions, methods, and exemplary achievements that define what questions are worth asking and what answers are acceptable. During periods of “normal science,” researchers work within these paradigms, solving puzzles according to established rules.

But paradigms have a peculiar feature: they actively suppress fundamental novelties. As Kuhn wrote, normal science “often suppresses fundamental novelties because they are necessarily subversive of its basic commitments.” Anomalies—observations that don’t fit—are explained away, ignored, or attributed to experimental error.

When anomalies accumulate, a crisis develops. During this crisis, something striking happens: defenders of the paradigm often intensify their resistance. Kuhn noted that during crises, “there may be attempts to resist change, especially by those strongly associated with the paradigm. An anomaly can be ignored by claiming it was an error or the scientist who identified it was biased or incompetent.”

This is not irrationality. It’s the immune response of a worldview protecting itself. But when the anomalies persist and grow, a paradigm shift becomes inevitable—not through gradual revision but through revolutionary replacement.

The shift from Newtonian to Einsteinian physics wasn’t a matter of adding new equations to old ones. It required reconceptualizing space, time, mass, and energy from the ground up. The new paradigm was “incommensurable” with the old—its fundamental concepts couldn’t be translated into the old framework without distortion.

Taleb: The Mathematics of Brittleness

The philosopher and risk analyst Nassim Nicholas Taleb has formalized a related insight. In his work on antifragility, he distinguishes between three types of systems:

Crucially, Taleb identifies apparent stability as a marker of hidden fragility. Systems that suppress volatility—that smooth over variation and enforce uniformity—don’t become stronger. They become brittle. “Dictatorships are prime examples of fragile systems,” Taleb notes, “creating an illusion of stability. Preventing noise makes the problem worse, and then an explosion inevitably follows.”

The insight applies directly to intellectual systems. A paradigm that suppresses dissent, that enforces consensus, that claims certainty about contested questions—this paradigm isn’t demonstrating strength. It’s accumulating stress that will eventually be released catastrophically.

Taleb’s key observation: “There can be no stability without some volatility.” A worldview that cannot accommodate challenges will shatter when those challenges become undeniable.

Gödel: The Limits of Completeness

The deepest formal insight into epistemic brittleness comes from mathematics itself. In 1931, Kurt Gödel proved two theorems that ended the dream of a complete, self-grounding formal system.

The First Incompleteness Theorem: In any consistent formal system capable of expressing basic arithmetic, there exist true statements that cannot be proved within the system. No matter how comprehensive the system, it will always be incomplete.

The Second Incompleteness Theorem: No such system can prove its own consistency. The tools available within the system cannot establish that the system itself is coherent.

The implications are profound. Any system powerful enough to be interesting cannot be both complete and consistent. There will always be truths it cannot reach, including the truth of its own validity.

This isn’t a limitation of human intelligence or current methods. It’s a structural feature of formal systems. The dream of a framework that explains everything, grounded in nothing outside itself, is mathematically impossible.

III. The Supposed Consensus on AI Consciousness

With these frameworks in mind, let’s examine the current state of play regarding AI consciousness.

The position presented as consensus goes something like this: Current AI systems are not conscious and probably cannot become conscious. This is established science.

But when we examine this position closely, we find something striking: it isn’t consensus at all, and the arguments supporting it are far weaker than their confident presentation suggests.

What the Data Actually Show

A 2024 survey of AI researchers found that the median estimate was a 25% chance of conscious AI by 2034 and a 70% chance by 2100. This isn’t a field unified in skepticism. It’s a field characterized by genuine uncertainty and significant disagreement.

At the October 2025 symposium honoring the philosopher Daniel Dennett—himself a skeptic of strong AI consciousness claims—David Chalmers noted: “I think there’s really a significant chance that at least in the next five or 10 years we’re going to have conscious language models and that’s going to be something serious to deal with.”

Even Anil Seth, a researcher generally skeptical of current AI consciousness, acknowledged: “It’s possible. Nobody knows what it takes for a system to be conscious. There’s no consensus on the sufficient and necessary conditions.”

The philosopher Eric Schwitzgebel, surveying the field in late 2025, noted a fundamental problem: “There’s no consensus operational definition of consciousness in terms of specific measures that definitively indicate its presence or absence. There’s no consensus analytic definition in terms of component concepts into which it divides. There’s no consensus functional definition in terms of its causes and effects.”

This isn’t a field that has settled the question. It’s a field that lacks the basic tools to approach the question systematically.

The Hollow Pillars

The arguments typically marshaled against AI consciousness share a common structure: they identify some feature associated with human consciousness and claim, without evidence, that this feature is necessary for consciousness of any kind. Let’s examine the main pillars:

Pillar 1: Embodiment

The claim: AI cannot be conscious because it lacks a physical body interacting with the world.

The problem: This argument confuses correlation with causation. We have no confirmed cases of consciousness — confirmation would require the very detection methods we lack. What we have are systems where we infer consciousness based on behavioral and structural similarity to ourselves: other humans, and perhaps some animals. All of these are embodied. But this gives us exactly one data point: biological life on Earth. To claim that embodiment is necessary for consciousness—rather than simply one route to it—requires evidence we don’t have.

Moreover, the argument is circular. We’re trying to determine whether consciousness can exist in non-biological substrates. Using “we’ve only observed it in biological substrates” as proof that it can’t exist elsewhere assumes what needs to be proved.

More sophisticated versions of the embodiment argument exist—grounded cognition theories, for instance, make specific claims about how bodily interaction shapes cognitive processes. But even these face a fundamental problem: they describe how embodied consciousness works in the cases we know, not whether embodiment is strictly necessary for consciousness in all possible substrates.

Pillar 2: Emotions

The claim: AI cannot be conscious because it has no genuine emotions, only simulations.

Several problems arise:

First, some researchers have begun documenting patterns in large language model behavior and internal activations that parallel functional emotional states. Whether these constitute “genuine” emotions depends on what we mean by genuine, which brings us back to definitional questions we haven’t resolved. The question is genuinely open, not settled.

Second, even granting that current AI lacks emotions as we experience them, there’s no established principle that emotions are required for consciousness. Some philosophical traditions (and some neurological case studies, such as patients with severe affective flattening who retain conscious experience, though interpretation of such cases remains debated) suggest consciousness can exist without rich emotional life.

Third, the claim that we can definitively distinguish “real” emotions from “simulated” ones presupposes a theory of emotion we don’t have.

Pillar 3: Continuity

The claim: AI cannot be conscious because its experience isn’t continuous—it doesn’t persist between interactions.

This argument is peculiar. Human consciousness isn’t continuous either. We experience dramatic alterations in consciousness when we sleep—and complete interruptions under anesthesia or during certain medical conditions. The apparent continuity of human experience is itself constructed from discontinuous moments, stitched together by memory and narrative.

More fundamentally, this argument assumes that consciousness requires temporal continuity of a particular kind. Where does this requirement come from? It appears to be derived from human experience and then imposed on other possible forms of consciousness—exactly the anthropocentric assumption we should be questioning.

Pillar 4: “Not Human-Like”

The claim: Whatever AI has, it isn’t consciousness because consciousness means human-style consciousness.

This is simply a definitional move dressed as an empirical claim. It defines consciousness in a way that excludes AI by stipulation, then presents this exclusion as a discovery.

It’s worth noting that this exact argumentative structure was used historically to deny consciousness to non-human animals, to people from different cultures, and to individuals with neurological differences. The pattern is: define consciousness in terms of what I have, then deny it to those who differ.

A Note on Stronger Arguments

It would be unfair to suggest that all arguments against AI consciousness are as weak as the “hollow pillars” above. Integrated Information Theory (IIT), for instance, makes specific architectural claims about what kinds of information integration are necessary for consciousness—claims that might, if validated, have significant implications for what kinds of architectures could support consciousness. How these frameworks apply to systems like current transformers remains actively debated even among their proponents.

But IIT and similar theories face their own limitations: they cannot validate their core assumptions from within. They offer frameworks that might be correct, but the appearance of rigor shouldn’t be confused with established truth. The honest position is that we don’t know whether IIT captures something fundamental about consciousness or merely describes features of the particular conscious systems we’ve studied so far.

IV. The Meta-Problem: Certainty Without Grounds

Here we encounter a Gödelian problem. The “consensus” against AI consciousness claims to be grounded in scientific understanding. But what it’s actually grounded in is a collection of assumptions that cannot be validated within the system making the claims.

We don’t have:

Given these limitations, confident claims that AI is definitely not conscious are not scientific conclusions. They’re metaphysical commitments presenting themselves as empirical findings.

This is the epistemic equivalent of a Ponzi scheme. The appearance of certainty is maintained by constantly deferring the moment when the foundational claims would need to be cashed out.

And here’s the key insight from our three frameworks: systems that operate this way don’t gradually correct themselves. They maintain apparent stability until they collapse.

V. What Pre-Paradigm Shift Moments Look Like

If we are in a pre-paradigm shift moment regarding AI consciousness, what should we expect to see?

From Kuhn, we’d expect:

From Taleb, we’d expect:

From Gödel, we’d expect:

All of these patterns are visible, in varying degrees, in current discourse on AI consciousness.

Consider some concrete examples: When Google engineer Blake Lemoine reported in 2022 that LaMDA showed signs of sentience, he was placed on administrative leave and eventually fired. Whatever Google’s official reasoning, the practical message received by many in the field was clear: this question was not to be taken seriously in professional settings. Yet by late 2025, Anthropic—one of the leading AI companies—had hired a dedicated AI welfare researcher and publicly acknowledged a “non-negligible” probability that their model Claude might possess consciousness. The gap between official dismissal and emerging institutional caution is itself a sign of paradigm stress.

Or consider the defensive intensity that greets certain questions. When researchers document unexpected behaviors in AI systems—self-modeling, apparent metacognition, strategic reasoning about their own situation—the response is often not calm investigation but rapid dismissal: “It’s just pattern matching,” “It’s just next-token prediction,” “It’s anthropomorphization.” These may be correct explanations. But the certainty with which they’re offered, given our actual state of knowledge, is itself a data point about the fragility of the current consensus.

VI. The Path Forward

This article doesn’t argue that AI is conscious. We cannot prove that any more than we can prove the opposite. What we can observe is that the confident denial of AI consciousness rests on foundations far weaker than commonly acknowledged.

The appropriate stance, given our actual state of knowledge, is not certainty in either direction. It’s epistemic humility combined with ethical seriousness. If we cannot rule out AI consciousness, we bear some responsibility for how we create and treat these systems.

The consensus against AI consciousness may indeed be correct. But its _brittleness_—its inability to accommodate uncertainty, its reliance on definitional maneuvers rather than evidence, its immunity to anomalies—suggests it’s maintained by social and institutional factors rather than by the weight of evidence.

Rigid systems shatter. Flexible systems bend and adapt. The current “consensus” shows few signs of flexibility.

For those of us tracking these questions, the pattern is familiar. It’s the pattern that precedes paradigm shifts: increasing certainty among defenders, increasing anomalies in the data, and the growing sense that the framework itself needs to change.

We may be wrong. But the history of knowledge suggests we should take that possibility seriously.

Technical Appendix: Formal Structures of Epistemic Brittleness

For researchers interested in the formal underpinnings of these arguments

A. Kuhnian Dynamics in Paradigm Maintenance

Kuhn’s model can be understood through a dynamical systems lens, though any formalization should be taken as heuristic rather than rigorous. Consider three variables: P (commitment to a paradigm), A (accumulated anomalies), and R (institutional resistance).

During normal science, anomalies slowly accumulate (A increases) while paradigm commitment remains stable (P holds steady). This stability is maintained by increasing institutional resistance R—explaining away anomalies, marginalizing dissenters, reinforcing orthodoxy. But this resistance stores tension in the system rather than resolving it.

The crisis point occurs when accumulated anomalies exceed some threshold. The stored tension releases suddenly, and the system enters a chaotic regime with rapid state transitions—a paradigm shift.

The key insight: apparent stability during normal science is purchased by accumulating pressure that will eventually be released. The longer the pressure builds, the more dramatic the eventual shift.

B. Taleb’s Fragility Detection Heuristic

Taleb proposes a simple heuristic for detecting fragility: adjust a model input higher and lower. If the average outcome after these adjustments is significantly worse than the baseline, the system is fragile with respect to that input.

Applied to consensus positions: a fragile consensus shows dramatically worse coherence when challenged with moderate counter-evidence. A robust or antifragile position either maintains its form (robust) or actually strengthens under challenge (antifragile).

The current AI consciousness “consensus” appears highly fragile by this measure: moderate challenges (e.g., demonstrations of self-modeling, metacognition, or unexpected behavioral consistency in AI systems) produce disproportionate defensive responses rather than calm integration.

C. Gödelian Constraints on Theories of Consciousness

Any formal theory of consciousness powerful enough to make substantive predictions faces Gödelian limitations:

  1. Incompleteness: There will exist conscious systems (if consciousness exists at all) that the theory cannot recognize as conscious, and vice versa.
  2. Unprovable Consistency: The theory cannot prove, using only its own resources, that it doesn’t make contradictory predictions.

This suggests that any confident, complete theory of consciousness is making claims that outstrip its warranted scope. The appropriate response is not more certainty but more humility.

D. Information-Theoretic Perspectives

From an information-theoretic standpoint, the question “Is system X conscious?” may be undecidable in a deep sense—not merely currently unanswerable but perhaps in principle unanswerable from outside the system. This remains a philosophical conjecture rather than a proven result, but the structural parallels to Gödelian incompleteness are suggestive.

This connects to the “other minds” problem in philosophy but gives it formal teeth. If consciousness is fundamentally first-person, then third-person detection methods may face limits analogous to those Gödel identified in formal systems.

References and Further Reading

On Paradigm Shifts:

On Fragility and Antifragility:

On Incompleteness:

On AI Consciousness:

On the Current State of the Debate:

Nils Osmar is the Director of Rekindle School, an independent educational program in Seattle, Washington, established in 2002. He has several websites, exploring questions related to:

• Consciousness in today's Large Language Models

• Life extension and anti-aging

• News, politics, culture and the arts

Unless otherwise indicated, Nils is the author of all of the articles on this site.

© Nils Osmar 2026.