Alexander Lerchner's paper argues AI consciousness is physically impossible, not just technically distant, making every frontier model a sophisticated tool rather than a moral patient.
Alexander Lerchner, a senior staff scientist at Google DeepMind, published "The Abstraction Fallacy: Why AI Can Simulate But Not Instantiate Consciousness" on PhilArchive in March 2026. The paper has since been downloaded over 5,370 times and listed on DeepMind's official publications page. Lerchner's argument is not that today's AI is immature or that future architectures might cross a consciousness threshold. The claim is sharper: algorithmic symbol manipulation is structurally incapable of instantiating experience, regardless of scale, architecture, or how many decades of development pass. The paper puts a philosophical floor under the question that most AI companies have deliberately avoided answering.
The core concept is the abstraction fallacy itself. Lerchner argues that computational functionalism, the dominant view in AI consciousness debates, rests on a fundamental error. Functionalism holds that consciousness emerges from abstract causal structure, independent of physical substrate. If the right information processing pattern exists, experience follows. Lerchner disputes this from physics upward: symbolic computation is not an intrinsic physical process but a mapmaker-dependent description, requiring an active experiencing cognitive agent to translate continuous physics into discrete meaningful states. In simpler terms, a symbol is only a symbol because a conscious being defined it as one. The system manipulating symbols is not conscious; it depends on consciousness to operate at all.
404 Media spoke to several philosophers of consciousness about the paper. The consensus was instructive: the arguments are sound, but they are not new. Academics have made structurally identical cases for decades, drawing on Searle's Chinese Room, Nagel's "What Is It Like to Be a Bat," and Block's distinction between access and phenomenal consciousness. What is new is the institutional address. This is DeepMind, one of the world's preeminent frontier AI labs, putting its name on a formal position that current and future LLMs are not and cannot be conscious. The paper is not an obscure preprint. It is on DeepMind's publications page.
The response from the philosophical community is not unanimous. Alex Bogdan published a direct reply on PhilArchive arguing that Lerchner's case depends on a contested conception of computation and that the stronger conclusion, that AI is structurally incapable of consciousness, has not been established. Bogdan's position is that disciplined uncertainty is the intellectually honest stance: present systems give no compelling evidence of consciousness, but no one has ruled out the possibility for future physical architectures. The News Pakistan noted DeepMind also hired a philosopher of mind to its research team this month, signaling that the lab is not treating the question as settled internally despite Lerchner's paper.
Why This Matters Beyond Philosophy Seminars
The debate sounds abstract until you consider its practical downstream effects. AI rights legislation, AI liability frameworks, and enterprise procurement decisions all sit downstream of consciousness claims. If a deployed AI system is conscious, it may have moral status, creating liability exposure for companies that treat it as a pure tool. Regulators in the EU working on the AI Act have already flagged sentience as a live question requiring attention. Lerchner's paper, by arguing consciousness is physically impossible for current architectures, removes that ambiguity in a direction favorable to continued commercial AI deployment without ethical encumbrance.
That is not a cynical reading. Lerchner himself frames AGI without sentience as possible, writing that highly capable artificial general intelligence does not inherently create a novel moral patient but rather a sophisticated non-sentient tool. For enterprise buyers evaluating AI governance, that framing matters enormously. It means deploying an AI agent to manage customer interactions, review legal documents, or run financial models carries no consciousness-related liability, only the standard product liability of deploying any sophisticated software.
PubMed published a separate IIT-based study reaching the same conclusion via different methodology: LLMs fail to meet integration, causal closure, and temporal persistence criteria for consciousness under Integrated Information Theory, generating negligible integrated information despite impressive linguistic output. Two independent theoretical frameworks converging on the same answer in the same month is not coincidence. It reflects the maturation of a genuine scientific research program that the field has been reluctant to pursue formally, because the answer is commercially inconvenient for anyone marketing emotional resonance as a product feature.
Watch whether Anthropic or OpenAI respond with formal positions of their own. Both companies have issued internal guidelines on AI welfare without taking a clear public stance on whether their systems could be conscious. Lerchner's paper raises the cost of continued ambiguity: if DeepMind has staked out a clear position and the science supports it, competitors who maintain studied vagueness on the question start looking either uninformed or strategically evasive. The abstraction fallacy, whether or not the philosophical argument holds completely, is now part of the formal record. The industry has to engage with it.
Also read: YouTube embeds AI chatbot into search as Google turns 2 billion viewers into a Gemini surface • Lightelligence's 600x oversubscribed HK IPO proves the market wants photonic AI chips • Taylor Swift trademarks her voice and face to draw a legal line against AI deepfakes
