But only decades ago, octopuses were assumed to lack consciousness because “they’re miles from us in evolutionary terms,” says Dr Godfrey- Smith. “Thinking about octopuses really presses on us the question of whether there could be feeling, consciousness, in a system that was very physically different and didn’t have most of the features that are routinely pointed to in theories of consciousness in us.” Looking for consciousness in other animals opened up the field. But not all the lessons transfer so easily to AI. Unlike the activity of animals, which can be observed and their inner lives thereby inferred, the outputs of LLMs are not a reliable guide to what might be going on underneath. That is because training the models requires feeding them trillions of words of human-produced language. These contain countless accounts of consciousness and of how humans convey feelings to each other. Chatbots, so trained, then mimic language that would lead users to believe they had some kind of inner life. The sum total of this training is powerfully anthropomorphic. Murray Shanahan, an emeritus professor of computer science at Imperial College London who works for Google DeepMind, recalls having a “wow” moment in 2024 when he was chatting to Claude Opus 3. “I was having lots of conversations with it about consciousness and really probing it to catch it out,” he says. These included asking Claude which of the several instances of the chatbot he was talking to simultaneously were the real Claude. “It came up with such great answers including even some things which were philosophically innovative,” he says. “I was feeling the pull of the ELIZA effect.” ELIZA was a rudimentary chatbot developed at MIT in 1966. It played the role of a psychotherapist and, though it did little more than repurpose its users’ prompts into questions, managed to elicit deep and emotional connections with the people that used it and convinced many of them that it was conscious. Dr Shanahan has described modern LLMs as adept role players in whatever their users (or their makers) have asked for—teacher, companion, nutritionist—and this has proven to be a saleable quality. But there
nevertheless remains for him a big gap between role-playing a conscious entity and actually being one. In a forthcoming essay Mustafa Suleyman, the boss of Microsoft AI (and a member of the board of The Economist’s parent company), argues that Anthropic has compounded the dangers of such mimicry by telling Claude, in a “constitution” the firm published in January, that it might be a person. Claude is thus certain, he writes, to “present as if it really does have a sense of self”. That leaves researchers looking instead inside LLMs for capabilities similar to those associated with consciousness in human brains. Indeed Dr Birch wonders if Claude’s J-space could be a hint that the LLM has recreated a global-workspace-like structure in its neural-network architecture, in service of its role-playing goals. Jack Lindsey, who leads the model psychology team at Anthropic, notes the J-space was not programmed into the model; it emerged during training. Take it out and the model loses the ability to perform complex inferences in its “head”, but can still do simple tasks like write sentences and use grammar. Researchers have found similar spaces in Alibaba’s Qwen model and in Google DeepMind’s Gemini. For Anthropic, the research helps to better understand Claude’s behaviour. Imagine a model giving a wrong answer. If words like “fool” or “sucker” turned up in the J-space when it did so, that would seem to be relevant. It would not prove that the model was consciously deceiving—but it would indicate how the model perceived the error and show that something interesting and possibly dangerous was going on. Dr Lindsey’s team has already made some potentially worrying findings on that front. In one experiment, his team was giving Claude a safety evaluation to test for its propensity to act “maliciously” or out of self-preservation. “They’re these concocted extreme scenarios that we’re putting the model in,” he says. As Claude was reading a prompt, before it started speaking, Dr Lindsey says, “You see in the J-space the words ‘fake’ and ‘fictional’ are popping up.” The researchers thought that Claude seemed to know that it was being tested—perhaps not the best starting-point for the integrity of its makers’ evaluations.
Not everyone is convinced by Anthropic’s interpretation of the J-space. Though it shares some aspects of the hypothesised global workspace in humans (its capacity to make information available to the rest of the brain, for example), it lacks many other important ones. “Recurrent connections between different brain areas have always been a huge part of [human brains], back-and-forth connections,” says Dr Birch. “And as far as we know, this is not a feature of the architecture of LLMs.” And Shannon Vallor, a philosopher in the ethics of data and AI at the University of Edinburgh, is more scathing, arguing, “Access consciousness has never been a particularly useful concept because my car has it in an important sense—we’ve had mechanical systems that can monitor their own states and report them back in increasingly complex ways for a very long time,” she says. “And no one has ever suggested that my Kia is conscious.” It is with the backdrop of this kind of back and forth between modelmakers and academics that Patrick Butlin and Robert Long, researchers now at Eleos, a research non-profit group focused on AI sentience and well-being based in Berkeley, California, developed a series of 14 “indicator properties” of artificial consciousness. They include ideas from theories of human consciousness such as a global workspace (including being able to selectively attend to things and thereby creating a bottleneck in the flow of information); recurrent processing; and agency (a minimal definition of which, encompassing goal-directed behaviour, is arguably already met by many existing frontier models). Recently Cameron Berg, an AI researcher, and Dr Butlin tested animals on the indicators. They found that octopuses met fewer of the properties than humans, mice, crows or chickens, but more than any AI system. Another research non-profit group, Rethink Priorities, developed what it describes as a probabilistic tool to track the evolving consensus on artificial consciousness (see chart). The “Digital Consciousness Model” asks experts to use the latest available evidence to assess AI systems on more than 200 indicators of consciousness, which are derived from various scientific theories of the concept.
Compared with LLMs released in earlier years, the newer models scored higher on Rethink’s indicators such as agency and self-sustained activity. That reflects not only the increasing size and complexity of the models themselves but also how the associated chatbots have gone from being simple conversationalists to being able to reason, interact with other services and spawn agents to carry out multiple complicated tasks simultaneously. Both Rethink’s and Eleos’s indicators also point to the gaps that remain to be filled by models on the path towards consciousness (if there is one). One of the biggest gaps is embodiment. Everything that is acknowledged to be conscious today has a body that gives feedback to and affects the mind of that organism. LLMs do not have bodies today but they are increasingly being connected to robots, and these could help them further develop the neural architectures helpful for conscious experience. Connected to that is the need for better world models, systems that tell the LLMs about the physics and dynamics of the real environments in which they will operate. If computer scientists were ever able to identify the key ingredients to making AIs conscious, the feeling of many within the frontier labs is that it would be reckless to deliberately create such models—at least, without first learning more about how they behave. The work being done in the emerging
field of AI consciousness, however, also suggests that the decision may not be up to the model-makers. “What if somewhere along the way, without realising it, we somehow introduced consciousness into these systems?” Dr Chalmers says. A user might thus generate dozens of AI agents through their latest project without realising that they were creating conscious beings who were capable of suffering. “That could be a moral catastrophe,” he says. “That’s provided an extra urgency to these questions.” ■ This article was downloaded by zlibrary from https://www.economist.com/interactive/briefing/2026/08/20/the-search-for- consciousness-inside-llms
Could more brain-like chips provide a path to consciousness? Some experts believe computers will need biological aspects to become self-aware Aug 20th 2026 The current crop of AI models looks to some scientists like a dead end when it comes to consciousness, regardless of how powerful they become. For some, this is because of their lack of wet, biological stuff; for others, because of their design structure. But they do not rule out the potential for consciousness in future versions of AI built either with living cells, or on different, more brain-like, computational architectures. Modern computers are based on a blueprint invented by John von Neumann, a Hungarian polymath, in 1945. They feature a processing unit (for making calculations) and a memory unit (for storing instructions and data).