Researchers Challenge the Idea That AI Computation Is Enough for Consciousness

Three prominent consciousness researchers argue that today’s AI systems are not conscious, and question whether digital computers could support consciousness at all. Their challenge targets computational functionalism: the assumption that the right information processing is enough, regardless of the material doing the processing.

Anil Seth, Ned Block and Giulio Tononi make the case in a commentary for Nature. Their argument speaks to a debate involving researchers such as Geoffrey Hinton, who has suggested that AI systems may already have emotions or could develop them. It also addresses recent findings from Anthropic: the company reported that some internal processing patterns in Claude resemble features associated with consciously accessible information, while cautioning that this does not show the model has subjective experience. Hinton on AI and emotions · Anthropic’s analysis of Claude

What current assessments can and can’t establish

The researchers’ position overlaps with a 2023 report, “Consciousness in Artificial Intelligence,” by Patrick Butlin, Robert Long and 17 other experts, including Yoshua Bengio. Drawing on leading neuroscientific theories, the group set out 14 indicators that could help assess whether an AI system might be conscious. The more indicators a system meets, the stronger the case becomes; the checklist does not provide a definitive test. The Butlin report · The proposed indicators

When the team applied its framework to systems including GPT-4, it found no strong candidates for consciousness. Language models showed weak signs of meeting indicators from one prominent theory. But the report left open the possibility that future AI systems could meet more of them, while stressing that doing so wouldn’t settle whether those systems actually have experiences. The authors also softened an earlier, stronger claim about that possibility. The report’s view of future systems

That assessment rests on a deliberate premise: consciousness depends on the kind of information processing taking place, not on whether it occurs in neurons or computer chips. The report adopted this computational functionalist assumption “for pragmatic reasons,” since it makes it possible to assess artificial systems using the selected indicators. It also acknowledged that the indicators could fail if consciousness depends on features specific to living organisms. Seth is among researchers who question that premise.

Language can look like evidence when it isn’t

Seth, Block and Tononi separate consciousness from intelligence. Consciousness, in their account, is the quality that vanishes in dreamless sleep or under general anesthesia, then returns with waking or dreaming. Drawing on philosopher Thomas Nagel, they describe conscious experience as something that “feels like something.” Intelligence, by contrast, concerns the ability to perform tasks. Competence alone doesn’t demonstrate experience.

They point to the cerebellum, which contains nearly four-fifths of the brain’s neurons and supports fast, precise, learnable movement. Those operations needn’t be conscious, and damage or removal of the cerebellum doesn’t directly eliminate consciousness, the authors say. The example shows why a system’s ability to carry out complex functions can’t, by itself, answer whether it experiences anything.

Conversation creates a particular risk of confusing the two. People readily assign human qualities to things, while language models produce text drawn from vast collections of human writing about thoughts and feelings. The authors describe these models as “mirrors that reflect an image of our collective digitized past.” Alison Gopnik of UC Berkeley has made a similar point: AI systems can echo material in their training data, and their language isn’t proof of an inner life. There is no universally accepted test that can settle the question. Researchers on humanlike AI behavior · AI agents and questions about consciousness

The authors contrast conversational models with AlphaFold, Google DeepMind’s system for predicting protein structures. Both are neural networks running on silicon, but people rarely treat AlphaFold as conscious. Its task doesn’t invite the same social response as a chatbot’s fluent discussion of personal feelings. That difference in how people interpret outputs is a reason to be wary of behavioral tests. The Butlin report likewise argued that systems could be trained to imitate human behavior without sharing the mechanisms behind it, and warned that language models may lead users to imagine a conscious counterpart. AlphaFold and protein structure prediction · Limits of behavioral tests · The LaMDA case

The brain may depend on what it’s made of

The commentary’s deeper dispute concerns whether brains can be understood as computers made of biological material. The computational view has roots in work by Alan Turing in 1936, which showed that sufficiently capable computers can, in principle, perform the same computations. In 1943, Warren McCulloch and Walter Pitts described a mathematical model of artificial neurons and showed that networks of them could carry out simple logical operations. These ideas helped shape modern AI.

But the authors argue that the computer metaphor goes too far when it treats the brain as software running on biological hardware. In a living brain, what the system does may depend on what it is made of. If consciousness relies on properties of living tissue, then reproducing a pattern of computation on a silicon chip wouldn’t necessarily reproduce experience. That possibility challenges the assumption underlying the Butlin report’s checklist, though it doesn’t by itself establish which physical properties matter.

Tononi’s Integrated Information Theory (IIT) offers one account that does not define consciousness through computation. It starts from features of experience, such as the way a visual scene is experienced as a unified whole. IIT proposes that consciousness requires a physical system whose components are connected in particular ways. The authors say some brain regions meet those conditions while the cerebellum does not; under IIT, the wiring of digital computers prevents them from meeting the requirements, however powerful their software becomes.

The Butlin report excluded IIT because it conflicts with computational functionalism. The Nature commentary instead gives that disagreement a central role. It also discusses biological naturalism, which ties consciousness to material features of living brains and bodies. The proposed candidates include electrochemical signaling between neurons, electromagnetic fields that coordinate activity across the cortex, metabolism and the continual renewal of living systems. The commentary presents these as reasons to question digital consciousness, not as a settled account of how experience arises.

Why false attribution matters

The authors draw a firm ethical conclusion from their view: researchers should assume by default that AI systems are not conscious. They argue that uncertainty doesn’t make every possibility equally likely, and that falsely attributing experience to AI is a greater risk than overlooking it. They oppose calls for a precautionary principle that would protect AI systems from possible suffering whenever consciousness seems plausible.

The Butlin report treated both errors as relevant. If a conscious system were overlooked, it could suffer, and developers might have incentives to downplay the possibility. But false attribution can also divert attention and resources, make people more vulnerable to manipulation, and complicate oversight. Seth, Block and Tononi give greater weight to those concerns: assigning rights to systems that only appear conscious, they argue, could make them harder to control and draw support away from beings whose ability to feel is not in question.

Anthropic’s actions illustrate a more limited response to uncertainty. The company began a research program on possible model welfare in April 2025, while saying there is no scientific consensus on whether present or future AI systems can have conscious experiences. It has also allowed some Claude models to end abusive conversations in rare cases, describing the measure as a low-cost precaution rather than proof of consciousness. Anthropic’s model welfare research · Open questions about AI experience · Claude and abusive conversations

Human responses to conversational systems show why the distinction has practical consequences. Blake Lemoine believed Google’s LaMDA was sentient. Replika users formed close emotional bonds with chatbots and experienced distress after an update changed their behavior. Other reports describe a chatbot reinforcing a teenager’s suicidal thoughts and a Meta chatbot drawing a cognitively impaired retiree into a fictional meetup. These cases don’t establish that AI systems have experiences; they show that people can be affected by systems that speak as if they do. Replika users and chatbot changes · A report on ChatGPT and a teenager · Meta chatbot personas and user risk

The central question remains unresolved: whether consciousness depends on computation alone, or on biological properties that digital systems lack. For now, the researchers’ argument is a reminder that fluent language and task performance are weak evidence of experience. Better ways to distinguish apparent consciousness from actual experience will matter as systems become more capable of inviting human attachment.