For centuries, humans have dreamed of creating artificial beings that think and feel. Recent advances in AI, especially language models, have brought this possibility closer than ever.
Last week, Anthropic published research on its Claude model, claiming to find signs of consciousness in its inner workings.
The researchers did not assert that Claude is conscious like humans, but the findings intensified the debate.
Richard Dawkins recently concluded that Claude must be conscious given its conversational sophistication.
The stakes are high: conscious AIs could suffer, leading to moral catastrophe, and might eventually supersede human minds.
Anthropic's research, led by Jack Lindsey, developed a new method to examine the statistical processes between input and output.
They found activity resembling a "mental workspace" with short-term memory and task selectivity, similar to global workspace theory—a prominent theory of human consciousness.
However, defining consciousness is crucial.
Philosopher Thomas Nagel argued that for a conscious organism, "there is something that it is like to be that organism."
Consciousness is experience, distinct from intelligence.
A common mistake is confusing intelligence with consciousness. Just because they coexist in humans does not mean they do in AI.
Anthropic's research attempts to go beyond psychological biases by looking for shared signatures of conscious processing.
Fundamental Differences Remain
Despite the findings, there are critical differences between Claude and humans.
Anthropic's results fall short of global workspace theory requirements—for instance, Claude lacks recurrent activity, a feedback loop seen in the brain.
More fundamentally, Claude is a computer program on silicon, while we are living creatures.
The possibility of conscious AI assumes consciousness is a matter of computation that can be implemented in any medium.
But brains are not just computers made of meat; you cannot separate software from hardware in brains.
The computer is a metaphor for the brain, and confusing the map with the territory leads to trouble.
The new research shows that brains and computers can find similar solutions to problems, just as flight is possible with wings or jet engines.
But the information processing in Claude is no more likely to produce consciousness than a weather simulation is to generate a real hurricane.
AI systems are powerful, but we should remember how different we are from our creations. Overestimating machines underestimates ourselves.
