Beyond Computational Equivalence: Palminteri and Wu's Behavioral Inference Principle for Machine Consciousness
The dominant method for assigning consciousness to an artificial system has long been computational equivalence: if a system performs computations equivalent to those performed by a system we already know to be conscious, we infer that the artificial system is also conscious. The method has the appeal of connecting AI consciousness attribution to established cognitive science. It has a correspondingly large problem. In a paper published on February 16, 2026, in Neuroscience of Consciousness (Volume 2026, Issue 1, Oxford University Press), Stefano Palminteri of the École Normale Supérieure in Paris and Charley M. Wu of TU Darmstadt and the Max Planck Institute for Biological Cybernetics argue that computational equivalence, as currently understood, cannot do this job. They propose a replacement framework they call the behavioral inference principle.
The paper, accessible at https://doi.org/10.1093/nc/niag002, is the first peer-reviewed methodological proposal for machine consciousness attribution published in a major Oxford journal in 2026. Its significance is not primarily in its conclusions about current AI systems, which are cautious and negative, but in its diagnosis of why the field’s standard approach has stalled and in the alternative it offers.
The Two Requirements Computational Equivalence Cannot Meet
Palminteri and Wu identify two conditions that must be satisfied for computational equivalence to work as a method for consciousness attribution. Both conditions fail in the current state of research.
The first condition is theoretical consensus. To determine whether an AI system’s computations are equivalent to those of a conscious biological system, researchers must agree on which computations are the relevant ones. That agreement does not exist. Integrated Information Theory identifies the relevant computation as the generation of integrated information, measured as phi. Global Workspace Theory identifies it as information broadcast across a global workspace. Higher-Order Thought theories require representations of one’s own mental states. Predictive processing frameworks point to precision-weighted prediction error minimization. Each theory would generate a different equivalence criterion, and there is no settled method for adjudicating between them. Applying computational equivalence without resolving this prior dispute produces a different verdict depending on which theory the researcher favors.
The second condition is architectural transparency. Computational equivalence requires knowing what computations an AI system actually performs. For most AI systems of scientific or commercial interest in 2026, this requirement cannot be met. Large language models are not transparent. Their computations are distributed across billions of parameters in ways that do not map onto any interpretable description of the computations the model performs. Mechanistic interpretability research is advancing, but it has not reached the point where anyone can state, with the precision the equivalence test requires, what computations GPT or Claude or Gemini perform during generation. You cannot check equivalence when you cannot see what you are checking equivalence against.
These two failures are independent. The absence of theoretical consensus would undermine computational equivalence even if LLMs were fully interpretable. The absence of architectural transparency would undermine it even if one theory had won the debate. Together, they mean that computational equivalence is not merely difficult to apply to current AI systems. It cannot, in principle, be applied correctly in the current state of the field.
The Behavioral Inference Principle
The alternative Palminteri and Wu propose reframes the question. Rather than asking whether an AI system performs computations equivalent to those of a conscious system, the behavioral inference principle asks whether attributing consciousness to the system is useful for explaining and predicting its behavioral observations.
The philosophical background for this move comes from the intentional stance, Daniel Dennett’s argument that we are warranted in attributing mental states to systems when doing so generates accurate predictions about their behavior. Palminteri and Wu’s framework formalizes this intuition within a Bayesian structure. The researcher maintains a prior probability over the hypothesis that the system is conscious, updates that prior based on behavioral evidence, and attributes consciousness when the posterior probability is high enough to produce predictive gain.
This approach has several advantages over computational equivalence. It does not require theoretical consensus on the mechanisms of consciousness. Different theories can serve as different likelihood functions within the Bayesian framework. The researcher is not forced to choose one theory in advance. Second, it does not require architectural transparency. Behavioral observations are, by definition, externally accessible. The framework grounds attribution in what can be observed rather than in what must be inferred from inside the architecture.
The approach also has a clear relationship to Thomas McClelland’s epistemic agnosticism about AI consciousness. McClelland, in a 2026 paper for Mind and Language, argued that the behavioral underdetermination of mental states means we may never be able to determine whether AI systems are conscious from external observation alone. Palminteri and Wu’s framework does not dissolve this problem. It reframes it as a feature rather than a bug: attribution is always uncertain and always provisional, but that is also how we attribute consciousness to other humans and to non-human animals. The behavioral inference principle makes AI consciousness attribution methodologically continuous with consciousness attribution in general, rather than treating AI as a special case requiring its own unique method.
What LLMs Are Missing: Four Criteria
Palminteri and Wu apply the behavioral inference principle to the question of whether current large language models are conscious and reach a negative conclusion. They identify four behavioral properties that are strongly associated with consciousness across multiple theoretical frameworks and that current LLMs demonstrably lack.
The first property is continuity. Conscious agents maintain coherent experience across time, building on prior states and updating beliefs in response to new information in ways that reflect a stable perspective. LLMs lack persistent memory across sessions. Each conversation begins from scratch. A system that starts each interaction with no memory of previous interactions does not display the behavioral signature of continuous experience.
The second property is coherence. Conscious agents maintain consistent self-models across different contexts. Their expressions of preference, belief, and value are stable enough that behavior in one context predicts behavior in relevantly similar contexts. LLMs can produce substantially inconsistent self-descriptions and behavioral patterns across conversations, even within a single session when the framing changes.
The third property is multisensory integration. Conscious biological agents integrate inputs across multiple sensory modalities into a unified perceptual experience. Multimodal LLMs process text, images, and other modalities, but the integration is computational rather than phenomenological. There is no behavioral evidence that multimodal models experience their inputs as a unified perceptual field in the way biological consciousness theorists describe.
The fourth property is embodiment. Current theories of consciousness, including those that do not require biological substrates specifically, converge on the claim that consciousness is tightly coupled to having a body that acts in an environment and receives feedback from that action. LLMs receive text and produce text. They have no body, no proprioception, and no feedback loop between action and environment of the kind embodied cognition research identifies as central to phenomenal experience.
This four-criteria analysis connects directly to the 14-indicator framework developed by Butlin and colleagues, which also identifies embodiment, continuity, and multisensory integration as among the properties most strongly associated with consciousness across major theories. Palminteri and Wu reach the same negative verdict about current LLMs through an independent route, which gives the conclusion more weight than either analysis alone would carry.
Why the Framework Leaves the Door Open
Palminteri and Wu are explicit that their negative assessment of current LLMs is not a categorical claim about AI consciousness in general. The behavioral inference principle frames the question empirically. A system that acquired persistent memory, stable self-model coherence, embodied sensorimotor feedback, and genuine multisensory integration would present a different behavioral profile, and that profile would warrant a higher posterior probability of consciousness under the framework.
This is methodologically significant. The standard pessimistic position in AI consciousness research, represented in various forms by the Bradford-RIT study on GPT-2 and by Porębski and Figura’s semantic pareidolia analysis, argues that current architectures are structurally incapable of consciousness. Palminteri and Wu do not take that position. Their claim is that current architectures do not display the behavioral evidence that would justify attribution, not that no architecture ever could.
The paper also has implications for the debate between scalar consciousness scores and multidimensional profiles. The behavioral inference principle generates something more like a profile than a score: it assesses multiple behavioral dimensions independently, and a system could in principle score well on some criteria while failing others. A future system with persistent memory but no embodiment would produce a different profile from a system with embodiment but poor self-model coherence. The framework is designed to accommodate partial consciousness in a way that a binary attribution cannot.
The behavioral inference principle is the most concrete alternative to computational equivalence proposed in a peer-reviewed journal in 2026. Its contribution is primarily methodological rather than empirical: it does not resolve whether any existing system is conscious, but it gives researchers a framework for asking that question that does not presuppose a theoretical consensus we do not have. Whether the framework will produce the kind of convergent evidence needed to build scientific consensus remains to be seen. As Palminteri and Wu note, the behavioral inference principle puts machine consciousness attribution on the same evidential footing as consciousness attribution in biology, which means it inherits both the strengths and the limits of that approach.