For centuries, humans have been fascinated by the prospect of creating artificial beings in our own image. Of developing synthetic minds and artificial bodies that not only think but also feel, and are both intelligent and conscious. For the vast majority of this time, this prospect seemed very distant; a topic for science fiction and philosophy, not for the here and now. But over the past few years, the rapid rise of AI – and especially of language models – has changed everything.
Last week, the frontier AI firm Anthropic published new research on its language model, Claude, in which the researchers claimed to find signs of consciousness emerging within its inner workings. They didn’t claim that Claude is actually conscious in the same way that humans are, but the findings certainly upped the ante on the possibility of consciousness arising in AI.
These days, there’s an increasingly receptive environment for this kind of thinking. Richard Dawkins, the esteemed evolutionary biologist, recently concluded that Claude (or Claudia, as he called it) just had to be conscious, given the sophistication of its conversational ability.
The stakes are high. If AI is (or can be) conscious, the consequences would be seismic. Conscious AIs could potentially suffer, which would lead to an unprecedented moral catastrophe. And if silicon can be sentient, then perhaps our own messy brains and bodies will soon be superseded by machines that never age and cannot die. Some people even think that AIs will be our descendants, carrying the flame of consciousness far into the future and away across the galaxy.
Anthropic’s research, led by Jack Lindsey, is impressive. Its team developed a new way to look at the statistical acrobatics between the input to a language model – whatever gets fed in – and its output. They found activity that seemed to form a kind of “mental workspace” for the model. This workspace contained all sorts of words and phrases associated with the current conversation, held relevant items in something like a short-term memory, and showed selectivity for whatever task was at hand. It also displayed traces of step-by-step reasoning and much else besides.
In essence, their findings appear to reveal that the AI is spontaneously creating an internal space to “think” about what it’s doing and to organise relevant information, before deciding what to actually say in response to a given prompt.
The crucial point for the researchers was that these features are similar to those identified by one of the most prominent current theories of human consciousness: global workspace theory. This theory was introduced in the 1980s by the cognitive scientist Bernard Baars and elaborated over many years by the neuroscientist Stanislas Dehaene. The theory proposes that conscious experiences happen when information is made widely available to other parts of the brain.
Is the “workspace” within Claude good evidence for consciousness? To answer this, we first have to define what consciousness is. There’s no consensus on this, but a decent starting point comes from the philosopher Thomas Nagel. Fifty years ago, in his essay What Is It Like to Be a Bat?, he argued that that for a conscious organism, “there is something that it is like to be that organism”. It feels like something to be me, it feels like something to be you and it probably feels like something to be a dog or an elephant. But it doesn’t feel like anything to be a chair, or a table, or a person under profound general anaesthesia. Consciousness is any kind of experience whatsoever: the pain of a toothache, a pang of jealousy, the pleasure of eating ice-cream on a hot day.
An important consequence of this definition is that consciousness is different from intelligence. While consciousness is all about feeling and being, intelligence is all about doing – about performing functions of one kind or another. A common mistake people make when it comes to AI is to confuse the two – to take signs of intelligence as evidence for consciousness. But just because consciousness and intelligence go together in humans, this doesn’t mean they go together in general. Assuming that they do is a reflection of our own psychology, not an insight into the nature of reality.
This is one reason why the new Anthropic research is valuable. It doesn’t trade solely on our psychological biases. Instead, it goes deeper to look for signatures of conscious information-processing that might be shared by humans and machines. And it claims to find some.
This is all fascinating stuff, but there are still many differences between Claude and you and me, and good reasons to think these differences remain critical. For one thing, Anthropic’s findings fall short of what global workspace theory typically requires(for example, there is no recurrent activity in Claude, a specific kind of feedback loop of information that we see in the human brain). But there is a more fundamental difference:Claude is a computer program running on silicon, whereas we (and other conscious animals) are living creatures. Our brains are embodied in bodies that are embedded in worlds.
Why is this difference important? Because 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.
But the closer we look at real brains, the clearer it becomes that they are not just computers made of meat. For brains, unlike computers, you can’t cleanly separate what they do (the software) from what they are (the hardware). This, in turn, means that what they do is unlikely to be a matter of computation alone – since the ability to separate software from hardware is central to how modern computers work.
Somewhere along the way, we seem to have forgotten the computer is just a metaphor for the brain. A powerful metaphor, for sure, butstill a metaphor. And we will always get into trouble when we confuse a metaphor with the thing itself – the map with the territory.
To me, what the new Anthropic research shows is that both living brains and silicon computers can come up with similar solutions when faced with similar problems. Just as flight is possible with flapping wings or with jet engines, human brains and language models can figure out how to say plausible things either with consciousness (brains) or without it (language models). But 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.
AI systems are getting more powerful every day. But to give ourselves the best chance of navigating this new world, we should remember how different we are from our almost-magical creations. When we sell our minds too cheaply to our machines, we not only overestimate them, we underestimate ourselves.
