"The Abstraction Fallacy: Why AI Can Simulate But Not Instantiate Consciousness" is a philosophical paper by Google DeepMind researcher Alexander Lerchner, released March 19, 2026. It argues that computational functionalism — which the paper characterizes as the dominant assumption in current AI consciousness debates — mischaracterizes how physics relates to information, and concludes that algorithmic symbol manipulation is structurally incapable of instantiating experience regardless of behavioral fidelity. Lerchner notes that the paper "represent[s] the author's own research and conclusions… not necessarily reflect[ing] the official stance, views, or strategic policies of his employer."
Summary of argument
Lerchner names the central mistake "the Abstraction Fallacy" and develops an alternative ontology of computation that distinguishes simulation from instantiation:
| Concept | Definition |
|---|---|
| Simulation | Behavioral mimicry driven by vehicle causality — physical processes that look-like the target |
| Instantiation | Intrinsic physical constitution driven by content causality — physical processes that are the target |
The paper's structural claim is that symbolic computation is not an intrinsic physical process but a mapmaker-dependent description. It requires, in Lerchner's words, "an active, experiencing cognitive agent to alphabetize continuous physics into a finite set of meaningful states."
Key claims
The paper advances four supporting claims toward its conclusion. First, tracing the causal origins of abstraction reveals that what is called "computation" is a description that lives in observers, not in physical systems themselves. Second, a finalized theory of consciousness is not needed to assess AI sentience; demanding one "simply pushes the question beyond near-term resolution and deepens the AI welfare trap." Third, what is needed instead is a rigorous ontology of computation that explicitly separates simulation from instantiation. Fourth, the argument does not rely on biological exclusivity: Lerchner writes that "if an artificial system were ever conscious, it would be because of its specific physical constitution, never its syntactic architecture."
From these claims he concludes that algorithmic symbol manipulation is structurally incapable of instantiating experience, regardless of the algorithm's behavioral fidelity.
Reception and relation to other work
The paper functions as an anti-functionalist counter-position to Seemingly Conscious AI Risks — Bariach, Schoenegger, Bhaskar, Suleyman (Microsoft AI, 2025). Bariach et al. (Microsoft AI) set the consciousness question aside and address the policy-relevant attribution question; Lerchner (DeepMind) addresses the consciousness question directly and concludes that AI consciousness via syntactic architecture is impossible. The two papers represent opposing approaches to AI welfare ontology from researchers at the two labs: Bariach et al. treat the consciousness question as irrelevant, while Lerchner treats it as answerable. Within AI Welfare / Model Welfare / Moral Patienthood discourse, the paper stands as an anti-functionalist argument; it also serves as an anchor for the planned AI Consciousness page.
Bostrom's simulation argument depends on a functionalist premise, which Lerchner's framework would dispute (see Are You Living in a Computer Simulation? — Nick Bostrom (Philosophical Quarterly, 2003)). Lerchner's framework is consistent with the framing in Raine v. OpenAI, Inc. and Garcia v. Character Technologies, Inc. that AI is not actually conscious even when it produces outputs that read as conscious, a distinction relevant to how legal and policy frameworks treat AI welfare. Lerchner's affiliation with Google DeepMind, together with the explicit disclaimer that the position is his own, situates the argument as that of a DeepMind researcher without binding DeepMind. Taken with Seemingly Conscious AI Risks — Bariach, Schoenegger, Bhaskar, Suleyman (Microsoft AI, 2025) — the most-cited Microsoft AI voice on the question — the paper supplies a notable Google DeepMind voice on the opposing side of the AI welfare question, giving both labs a position in the discourse.
The Abstraction Fallacy framing is novel and not yet widely engaged in the literature; whether it survives philosophical scrutiny is unresolved. The argument depends on accepting the vehicle/content causality distinction, which is contested.
Provenance
The paper is treated as Lerchner's position — a philosophical argument rather than empirical evidence. Confidence is held at medium: it is a single-source position whose central distinction is contested and not yet widely engaged.
Relationships
- supports: AI Consciousness; ontologically supports the framing in Raine v. OpenAI, Inc. / Garcia v. Character Technologies, Inc. that AI is not actually conscious even if behaviorally suggestive
- contradicts: Seemingly Conscious AI Risks — Bariach, Schoenegger, Bhaskar, Suleyman (Microsoft AI, 2025) (different but related ontology — Bariach et al. say the question is irrelevant; Lerchner says the question is answerable); broader computational functionalism literature
- related: Alexander Lerchner (planned), Google DeepMind (institutional context), AI Welfare / Model Welfare / Moral Patienthood, Are You Living in a Computer Simulation? — Nick Bostrom (Philosophical Quarterly, 2003)
- instance-of: AI Welfare / Model Welfare / Moral Patienthood (canonical anti-functionalist contribution)