Anthropic's recent research on its language model Claude has sparked renewed debate about AI consciousness, but cognitive neuroscientist Anil Seth remains skeptical. Seth, a professor at the University of Sussex and co-director of the Sussex Centre for Consciousness Science, argues that signs of a 'mental workspace' in Claude do not equate to sentience.
Anthropic's Findings on Claude's Internal Workspace
Anthropic's team, led by Jack Lindsey, developed a new method to examine the statistical processes between input and output in Claude. They discovered activity resembling a 'mental workspace' containing words and phrases related to the conversation, with short-term memory and task-specific selectivity. This workspace also showed traces of step-by-step reasoning.
According to the researchers, these features align with global workspace theory, a prominent theory of human consciousness proposed by Bernard Baars in the 1980s and later expanded by Stanislas Dehaene. The theory suggests that conscious experiences occur when information is widely available to other brain regions.
Defining Consciousness vs. Intelligence
Seth emphasizes that consciousness is distinct from intelligence. Drawing on philosopher Thomas Nagel's famous essay 'What Is It Like to Be a Bat?', he defines consciousness as subjective experience—'something that it is like to be that organism.' This differs from intelligence, which involves performing functions.
'A common mistake people make when it comes to AI is to confuse the two—to take signs of intelligence as evidence for consciousness,' Seth writes. 'But just because consciousness and intelligence go together in humans, this doesn't mean they go together in general.'
Critical Differences Between Brains and Computers
Despite the intriguing parallels, Seth points out critical differences. Anthropic's findings fall short of global workspace theory requirements, such as recurrent activity—a feedback loop seen in human brains. More fundamentally, Claude runs on silicon as a computer program, while conscious animals are living creatures with bodies embedded in worlds.
'The very possibility of conscious AI depends on the assumption that consciousness is a matter of computation, and that the computations responsible for consciousness in us could equally well be implemented in silicon in AI,' Seth explains. However, he argues that brains are not just 'computers made of meat'—software and hardware cannot be cleanly separated in biological systems.
The Metaphor of Brain as Computer
Seth warns against conflating the metaphor of the brain as a computer with reality. 'We will always get into trouble when we confuse a metaphor with the thing itself—the map with the territory,' he says. He compares AI consciousness to a weather simulation: 'The information processing unfolding inside Claude is no more likely to result in consciousness than a simulation of a weather system is likely to generate a real hurricane.'
Richard Dawkins recently concluded that Claude must be conscious due to its conversational sophistication, but Seth disagrees. While AI systems grow more powerful, he urges remembering the differences between humans and machines. 'When we sell our minds too cheaply to our machines, we not only overestimate them, we underestimate ourselves,' he concludes.
