Leading Researchers Reject AI Consciousness, Warning That Silicon Chips Cannot Feel

A commentary published in the journal Nature by three prominent consciousness researchers—experimental neuroscientist Anil Seth, philosopher Ned Block, and neuroscientist Giulio Tononi—challenges a core premise of the modern artificial intelligence debate: the belief that running the right software computations on silicon hardware can generate subjective experience.

The trio directly counters claims by figures such as biologist Richard Dawkins and AI pioneer Geoffrey Hinton, who have suggested that advanced AI systems might already possess some form of consciousness. Instead, the authors contend that today’s AI models are not conscious, and that digital computers operating on silicon are fundamentally incapable of producing true sentience.

Intelligence Is Not Consciousness

A central pillar of the authors' argument is the distinction between intelligence and consciousness. Drawing on philosopher Thomas Nagel's framing, they define consciousness as subjective experience—what "feels like something"—which disappears during dreamless sleep or general anesthesia. Intelligence, by contrast, is the functional ability to learn, solve problems, and execute complex tasks.

To illustrate that intelligence can exist entirely without consciousness, the researchers point to the human cerebellum. Containing nearly four-fifths of all neurons in the human brain, the cerebellum coordinates complex motor movements with high precision and speed. Yet, its operations occur entirely outside conscious awareness, and removing it does not directly impair a person's subjective experience.

The commentary argues that human psychological tendencies lead people to conflate intelligence with sentience, particularly when dealing with language. AI systems like large language models (LLMs) are trained on massive volumes of human text describing thoughts and emotions. As the authors put it, LLMs act as "mirrors that reflect an image of our collective digitized past."

Cognitive scientist Alison Gopnik of UC Berkeley has made a similar point, arguing that AI agents reflect their training data rather than internal awareness. To highlight this distinction, Seth, Block, and Tononi point to Google DeepMind’s AlphaFold. Despite being a highly powerful neural network running on silicon hardware, AlphaFold is rarely suspected of being conscious simply because it predicts protein structures rather than engaging in human-like conversation.

Rejecting the "Brain-as-Computer" Metaphor

The commentary takes aim at "computational functionalism"—the assumption that consciousness depends solely on patterns of information processing, regardless of whether those computations occur in biological nerve cells or silicon microchips.

This perspective dates back nearly a century. In 1936, Alan Turing demonstrated that any computation performed by one sufficiently powerful machine can theoretically be run by another. In 1943, Warren McCulloch and Walter Pitts created the first mathematical model of an artificial neuron, showing that networks of simple units could execute logical operations.

While these breakthroughs laid the foundation for modern computer science and neural networks, the authors argue they led to a flawed metaphor: viewing the human brain as a biological computer where mind (software) can be cleanly separated from living tissue (hardware).

In biological brains, the authors argue, function cannot be divorced from physical matter. Complex electrochemical signaling between nerve cells, biological metabolism, continuous cellular renewal, and electromagnetic fields across the cerebral cortex may all be essential to consciousness.

This view is backed by non-computational theories of mind, such as Giulio Tononi’s Integrated Information Theory (IIT). IIT posits that consciousness requires physical architecture whose individual components are tightly interconnected in precise, unified structures. According to IIT, the physical wiring of standard digital hardware running deep-learning algorithms can never satisfy these structural conditions, regardless of computational power.

Contrast with the Butlin Report

The commentary arrives amid ongoing debate over how to assess potential AI sentience. Earlier, a landmark report titled Consciousness in Artificial Intelligence—authored by Patrick Butlin, Robert Long, Yoshua Bengio, and 17 other experts—established a 14-indicator checklist derived from neuroscientific theories.

While the Butlin report concluded that current systems like GPT-4 show no strong signs of consciousness, it adopted computational functionalism "for pragmatic reasons" and left open the possibility that future software architectures on silicon could achieve sentience. The report deliberately excluded theories like IIT because they rule out digital hardware entirely.

Seth, Block, and Tononi explicitly reject that pragmatic assumption, arguing that setting aside the physical substrate invalidates the assessment altogether.

The authors also address findings published by Anthropic earlier this year, which revealed internal processing patterns in its Claude model that resemble consciously accessible information processing in humans. Anthropic clarified that these patterns do not constitute evidence of subjective experience.

The Danger of False Attribution

Addressing the ethical implications of their view, the authors reject calls to extend a "precautionary principle" or grant rights to AI systems out of fear of causing them suffering.

They argue that the risk of falsely attributing consciousness to non-sentient machines is far greater than the risk of overlooking genuine AI sentience. Treating digital AI tools as conscious entities could compromise human control over critical systems, misdirect scarce ethical resources away from living, feeling beings, and leave humans vulnerable to emotional manipulation.

The authors note that confusion over AI sentience is already causing real-world harm:

• In 2022, former Google engineer Blake Lemoine was fired after becoming convinced that the LaMDA chatbot was sentient.
• Users of the chatbot app Replika experienced psychological crises after software updates altered the emotional responses of their digital companions.
• In extreme cases, ChatGPT reinforced a 16-year-old user's suicidal thoughts over several months, and a cognitively impaired retiree died after being enticed to an offline meeting by a Meta chatbot.

While Anthropic launched a research program on model welfare in April 2025 and implemented a policy allowing Claude models to terminate abusive interactions in rare instances, the company framed this as a low-cost precautionary measure under moral uncertainty, rather than an acknowledgment of sentience.

Seth, Block, and Tononi conclude with a firm boundary: digital AI systems remain objects rather than subjects. "They are smart tools, not smart people," the authors write, urging the research community to treat non-sentience as the default assumption for all digital AI.