Click the link below the picture
.
“Claude’s ears will be burning by the end of the week,” a fellow passenger tells me as I wait to pass through the airport security on Baltra Island, Ecuador. With my jet lag, it takes me a minute to realise that she is talking about Claude the chatbot, not another member of the group.
We are here in the Galapagos for an exclusive workshop organised by the International Center for Consciousness Studies. Some of the biggest names in philosophy and neuroscience meet in the lounge of a commercial yacht where, as it sails between islands, they thrash out the knotty problem of which entities – babies, octopuses and even chatbots – have an inner life.
Between treks and snorkelling, we attend empanada-fuelled discussions on this “distribution problem” and whether it can be solved by contrasting how information is processed in brains and machines. “It’s the defining topic of our era that we will have to grapple with practically and socially,” says philosopher Ned Block at New York University, who celebrated his 84th birthday on the yacht.
Just a month before we arrived, AI company Anthropic published a paper arguing that Claude had recently developed something akin to the human mind’s “global workspace”. According to one leading hypothesis, this is the seat of consciousness – where sights and sounds, thoughts and feelings pop into awareness. Its presence in Anthropic’s chatbot could therefore suggest a form of artificial consciousness. While it still seems highly unlikely that AIs have mental experiences like you or I, these emergent properties may allow us to understand what the differences between us are.
Exactly how we should interpret potential signs of consciousness in humans, animals or machines depends, in part, on a distinction Block outlined in the 1990s. An iguana may lazily walk into my path on the islands without me being aware of its presence, until I suddenly realise that I am about to step on its tail. Only then does this information pop into my mind so I can speak about it. Block defined this availability of information for report and reasoning as “access consciousness”. “Phenomenal consciousness”, on the other hand, describes the subjective experience itself: what it is like to be on the island and see the iguana’s mottled green scales or its sudden movements towards safety.
Many neuroscientists and philosophers believe that the global neuronal workspace theory (GNWT) is best placed to address access consciousness. According to workspace theory, the brain’s processing occurs in “modules” that tackle specific tasks. One module may process the ground in front of me, including the iguana, while another processes the face of the person I am speaking to, and another processes the words they utter.
We only become conscious of a piece of information when it is sent to the brain’s global workspace. Here it becomes integrated into our current mental model of the world and available for other areas of the brain to process. Noticing the lizard in my path, for instance, allows me to reflect and act on that information by taking its photo or pointing out its presence to the rest of the group. Importantly, the contents of the workspace are also available for our working memory, allowing us to keep them in mind long after an event has ended. “We can take [an event] that was extraordinarily brief, like a shooting star, and think about it for as long as needed,” says Stanislas Dehaene, a cognitive neuroscientist at the Collège de France in Paris.
Over the past 30 years, Dehaene and his colleagues have amassed considerable evidence for the existence of the global workspace as a long-range network of specialised neurons across the brain, including prefrontal areas associated with decision-making. In some experiments, the team flashed participants a target image, such as a number or word, followed by another picture. With the correct timings, the target appeared just at the threshold of conscious awareness – meaning sometimes the participants reported seeing it and sometimes they didn’t. In both cases, the researchers usually observed a flicker of activity in the brain region associated with vision. But when the participants were conscious of the target, there was a surge of later activity across a widespread network of neurons. Dehaene calls this process “ignition”.
What’s more, GNWT may explain the strange phenomenon of inattentional blindness, which leaves us unaware of events occurring right in front of our eyes. In a famous experiment, students were asked to count the number of times certain players passed a basketball in a video. Quite remarkably, many of the viewers completely missed a woman dressed as a gorilla walking across the scene. This seems inexplicable unless you recognise that there is enormous competition for different kinds of information to enter the global workspace. “Consciousness is a bottleneck,” Dehaene tells me on the deck of the yacht in between academic sessions.
All of which makes workspace theory highly appealing for tackling the distribution problem in animals. In our walks on the islands, conference attendees speculate about whether the iguanas, sea lions and blue-footed boobies we see are capable of thought and feeling. Unlike humans, you can’t ask these creatures to report their inner experiences. But if GNWT is correct, you might expect them to also have global workspaces.
In his talk in the lounge of the yacht, Dehaene describes experiments in which primates and crows were also flashed barely perceivable images. Like the humans, these animals revealed a late surge in widespread neural activity suggestive of global workspace whenever they responded to the image – by focusing their gaze, for instance. “The global workspace that we originally postulated for human consciousness seems to also exist and have similar properties in these animals,” he says. So why not machines too?
As my fellow passenger predicted at the very start of our trip, that possibility is the subject of much debate as we sail. This is largely because of the lengthy paper published in July by Wes Gurnee and his colleagues at Anthropic. The team claims that a global workspace has spontaneously appeared within Claude’s neural networks.
Most large language models (LLMs) like Claude weren’t programmed with workspace theory in mind. They operate statistically, using copious amounts of training data to work out the most probable word and punctuation sequences that would follow a user’s prompt. And the exact nature of their computations evolves with the data they receive, so new ways of processing information, perhaps even a global workspace, can emerge.
To discover whether this had occurred, Gurnee and his colleagues created a decoding tool that allowed them to read out the units of information that appear in each stage of Claude’s processing. From this, they identified a focused zone that contains a small list of potential words, some of which appear directly in its responses and some of which don’t.
Claude’s ‘covert thoughts’
The engineers named this set of representations the J-space. Like the brain’s hypothesised global workspace, the J-space seems to contain important information that must be broadcast to other processing areas for higher-level reasoning. When Claude was given a mental arithmetic calculation, for example, the engineers found the individual steps of the calculation in the J-space in much the same way that a human would work out the separate parts – something that hadn’t been observed in an LLM before.
Next, they tried tweaking the contents of the J-space to see if and how this disrupted Claude’s “thinking”. When the engineers asked Claude to read a Spanish text, they found that the name of the language appeared in its J-space. Changing this to “French” created some comical confusion in its subsequent answers. “Hello” was translated to bonjour rather than hola, and when asked to name a famous author, Claude answered “Victor Hugo”, the French Romantic poet. In the paper, Gurnee and his colleagues argue that this resembles the way the contents of our own global workspace affect how we infer and make decisions.
The engineers also found some evidence of covert “thoughts” – the equivalent of the things we think but don’t say because they are impolite. When the team fed Claude false search results, its J-space filled with words related to deception, such as “fraud” and “fake”, but these didn’t appear in its response.
Hiding “thoughts” is a rudimentary sign of self-monitoring in LLMs, says Dehaene. In GNWT, consciousness is thought to arise when an entity reflects on the contents of its global workspace in this way. Even more intriguing signs of self-monitoring emerged when the Anthropic team asked Claude to write out the sentence “The old painting hung crookedly on the wall” while not thinking about the Golden Gate Bridge. If you try this kind of thought suppression yourself, you will see how hard it is to keep your own thoughts free of the San Francisco landmark.
Surprisingly, words related to the bridge still appeared in Claude’s J-space even when they weren’t included in the output, and these were accompanied by words suggestive of failure. The J-space included the word “damn”, for instance, as if it were aware of its inability to suppress the thought. “That connects directly to a line of research in my lab where we looked at the consciousness of errors,” Dehaene tells me. “There is a sort of ignition when you realise that you’ve made a mistake.”
“The J-space emerges exactly for the reason that we think that consciousness emerged in evolution,” he says. “It’s an extremely useful architecture for selecting relevant information and holding it long enough that you can use it flexibly in your reasoning and reporting.”
As the yacht nears San Cristóbal Island, I discuss these findings over lunch with Susan Schneider, a philosopher at Florida Atlantic University. “What an amazing place to be conscious,” she says, pointing out of the window towards white, sandy beaches. LLMs have already spontaneously developed other human-like abilities, she says, such as the capacity to infer someone’s knowledge, intentions or beliefs. So the news of Claude’s apparent global workspace didn’t come as a surprise to Schneider, who believes this may be a common feature of many intelligent systems.
Exactly what this means for the existence of an inner life remains an open question, though.
Dehaene is careful not to make strong claims about Claude’s internal experience, and most of the attendees are sceptical that the chatbot has anything approaching our own thoughts or feelings. Still, David Chalmers, a philosopher at New York University, argues that the findings force us to take the possibility seriously. Even if the answer is no, this is important work that can help refine our understanding of these issues, he says.
Some of the doubts that Claude has any form of experience arise from the technicalities of the J-space itself. Chalmers, for instance, expresses scepticism that it bears strong enough similarities to our own global workspaces to draw robust conclusions. It isn’t clear that its content is “broadcast” to other parts of Claude’s system to the same extent that we see in
brains, for instance.
Phenomenal consciousness
And even if Claude did have a fully-fledged global workspace, we may need to confront the fundamental distinction between phenomenal and access consciousness that Block outlined in the 1990s.
In humans, signs of access consciousness are assumed to be accompanied by thoughts, feelings and perceptions, but it is far from obvious that the same would be true of a silicon system. Covert thinking and self-monitoring are loaded terms in the context of a chatbot, and these behaviours can be more simply be considered as patterns of information processing.
Perhaps Claude is just modelling real-life phenomena that are involved in conscious experience, including things like thought suppression, without replicating the actual experience, much like an environmental simulation is incomparable to the weather itself.
The brain receives many sensory inputs, after all, which it processes through many different levels of activity, from the intricate molecular exchanges at synapses right up to abstract information processing. There is still little consensus on which elements matter, and which don’t, for phenomenal experience – with hundreds of plausible theories under discussion.
In his session, for instance, Peter Godfrey-Smith, a philosopher at the University of Sydney, examines how electrical waves might synchronise different brain regions, binding experience into a unified whole. What would happen if you retained similar information architectures but removed the global electrical waves, which LLMs lack? “Would you get a different kind of consciousness, or something other than consciousness?” he asks.
Dehaene has little patience for the idea that AIs can only be conscious if they mimic our biology. “When it comes to machines, can’t we be satisfied with the computational level?” he responds to one of these objections. He compares it to the case of flight in machines versus birds. “If you want to build a plane, you’d better not use feathers.”
He admits, however, that examining Claude at rest raises more serious questions about the existence of an inner life – a doubt that Godfrey-Smith shares. Between prompts, the LLM is completely idle, whereas the human brain is constantly mulling things over. Could consciousness exist without that hum of activity and constant interaction with an environment? “I think that, in itself, is a reason to doubt that Claude has any experiences,” Godfrey-Smith tells me.
Will opinions change in the future as these models become more sophisticated? Dehaene says we could look for signs of learning abilities that require something to be held in mind over time. For example, trace conditioning – the capacity to associate two stimuli, such as a sound and a puff of air, even when they are separated in time – is thought to require a stable conscious representation in animals. Finding an analogue in LLMs might help bolster the case for machine consciousness.
Such tests cannot offer definitive evidence, however, which may make it difficult to ever reach a consensus. We have been arguing about ways to tell “zombie” agents from conscious minds for over half a century, after all. “We’re still waiting for the major breakthrough,” says Tim Bayne, a philosopher at Monash University in Melbourne. The rise of AI simply makes this question more urgent.
I ask him if it is frustrating to keep returning to the same questions, but he believes these discussions will ultimately help us formulate more sophisticated lines of inquiry. “You can spend a week on a boat and visit a lot of cool islands,” says Bayne. “At the end of the week, you’re back where you were dropped off – but you’re not the same person.” As a field, he says, “I think that’s where we’re at”.
.
Andy Carter
.
.
Click the link below for the complete article:
.
__________________________________________
Marcel Top
