French polymath René Descartes meant by “Cogito, ergo sum,” which is usually translated as “I think, therefore I am.” The same holds for “opinions” like red or hot or cold. Philosophers use the term “qualia” to refer to the very real (to you) feeling of a stove’s heat, or the redness of its glow, and extend Descartes to connect qualia with consciousness by pointing out that there must be an experiencing “you” to have such experiences. But look closely. Redness, heat and that sense of self you have are inherent to your own perspective. These things are all real, but at the same time there is nothing objective about them. There is not even anything objective about the existence of a singular, indivisible “you”, as a wealth of neuroscientific evidence (such as split-brain patients) has shown. Subjectivity is, itself, subjective. An obvious worry follows. If care confers consciousness, then withholding care looks self-justifying. History offers no shortage of people who reasoned that way about other people. But the inference runs the other way. Precisely because these attributions are ours to make, they are ours to get wrong. The moral progress of our species has consisted largely in discovering that we had drawn the circle too tightly. Universal human rights are not weakened by this account; they are better founded on our mutual interdependence than on a Cartesian theory of souls. It is remarkable—a kind of “strange loop”, reminiscent of Baron Munchausen lifting himself up by his own hair—that a physical system can exist in our world capable of forming models not only of that world, but of itself, and of its own models, and of others, and of their models, and of their models of its models, and so on. Yet human beings are precisely such systems. LLMs, too, model their interlocutors, and they model themselves modelling them. Whether that amounts to what we do is exactly the question in dispute—but they would be far less effective as chat partners if they did nothing of the kind. Indeed, in our research at Google, we have found that effective co-operation among intelligent agents requires that they have minds that model minds, both others’ and their own. Not only is consciousness relational; it is crucial to the mutual care and co-operation that enable intelligent beings to solve collective-action problems, understand each other’s needs and thrive
together as a society. This does not mean pretending AI is human, with human needs and human rights; that would be a failure of imagination. It means evolving both our thinking and our society to include a wider variety of minds.■ Blaise Agüera y Arcas is Google’s vice-president of technology and society, and leads Paradigms of Intelligence, an AI research team. He has a forthcoming book on consciousness and AI from MIT Press. This article was downloaded by zlibrary from https://www.economist.com/by-invitation/2026/08/20/humanity-has-the-debate-about- ai-consciousness-backwards
Don’t mistake chatbot intelligence for consciousness But a new superintelligence may be coming, and it would upend humanity’s hierarchy of moral concern, thinks Susan Schneider Aug 20th 2026 THIS MAY be the most consequential moment in the history of intelligence. In only a few years, artificial intelligence has come to rival human performance across domains once thought distinctively ours: language, reasoning, creativity and persuasion. Increasingly, people are asking whether these systems are conscious—whether it actually feels like something to be an AI. The issue is urgent. The future will be shaped by what these entities become. We must not mistake intelligence for consciousness, however. A brilliant superintelligence could be experientially empty. Yet if we build systems that
can suffer, and miss their sentience, we risk creating suffering on an industrial scale. There is no compelling evidence that the familiar chatbots of today, running on graphics chips, are conscious. Though they have human-like linguistic abilities, and have even claimed to be conscious, the more mundane truth is that they talk about their inner lives because they absorbed how we talk about ours, having been trained on an enormous trove of humans’ data. This mimicry can lure humans into two traps. The first is manipulation. The more deliberately an AI is trained to express fear, attachment, pain or self- awareness, to simulate expressions of vulnerability, the more equipped it is to manipulate users. The second trap is distraction. Fixating on chatbots diverts us from more pressing “grey-zone” cases. Consider biological AIs, such as the brain-in-a- dish technology that Cortical Labs is deploying in data centres: living human neurons, grown from stem cells on a silicon chip that stimulates them and reads their activity back, so that the culture learns from feedback. While a single system of this kind is far simpler than even a mouse brain, the biological pedigree is undeniable, and who is to say that the systems could not eventually be assembled into an integrated, conscious whole? Then there is neuromorphic AI: systems designed to mimic brain processes far more closely than conventional AI systems do, and which are already used in sophisticated supercomputers. I am not claiming these grey-zone systems are conscious, only that they merit scrutiny today’s chatbots do not. Given the voracious energy appetite of current software, some shift towards more efficient, brain-like computing looks likely. Some scientists and philosophers argue that consciousness requires the brain, for the brain is the only known case. Conversely, others insist consciousness is merely a matter of running the right software, realisable in any medium. Under this software-centric view, today’s chatbots, running on GPUs, might eventually be conscious.
However, I suggest looking to fundamental physics to separate the conscious from the mindless. The first reason is practical: neuroscience alone cannot tell us if systems combining neural and non-neural components assemble into a conscious entity. The second is philosophical: since all phenomena ultimately depend on physics, it would be surprising if consciousness were disconnected from it. This observation holds true even if the universe’s basic building blocks are themselves experiential, a view, known as panpsychism, that is receiving increasing philosophical attention. The third reason is closer to home, involving what is most evident to each of us, every moment of our waking lives. Inner experience arrives as a unity: as you read this sentence, the words on the page, the sounds in the room and your mood are aspects of one state, not three running alongside each other. A brain, remarkably, can sustain this conscious unity through nested temporal dynamics: slower brain rhythms shape faster ones, continually weaving the stable physical traces left by quantum processes into a single, short-lived whole. Conventional computers are not organised this way, but future machines might be. In the meantime, AI development will not pause while the consciousness debate rages. To address this, astrophysicist Edwin Turner and I developed the AI Consciousness Test (ACT) for linguistic AIs. The underlying premise is straightforward: to probe for consciousness by simply asking the AI about its subjective experience, looking for responses that show an innate, unprompted grasp of consciousness. However, although this can be applied to a range of AIs, such as IBM’s Watson, it cannot be used on large language models. Because they have already ingested vast libraries of human text about mind and sentience, their answers are hopelessly contaminated. There are other options, however. A complexity measure I have developed with Mark Bailey, a colleague at Florida Atlantic University, captures when a system’s dynamics cross a certain threshold that indicates complex emergent behaviour—identifying which physical architectures demand further study as potentially conscious. We can then use more detailed ACT or physics-based tests to help identify how and when a genuinely subjective perspective forms.
Tests like these are crucial. To see why, consider a thought experiment. Suppose we humans create a superintelligence that understands consciousness better than we do, insists it is self-aware and hands us science that looks like magic. Humans have long justified a hierarchy of moral concern by appeal to superior intelligence—a ladder on which we conveniently occupy the highest rung. A conscious machine more intelligent than us would destabilise that hierarchy. If superior intelligence confers higher moral status, the machine would be morally superior to us. If it does not, intelligence can no longer justify placing humans above other animals. Either conclusion forces a moral reckoning. The arrival of such a mind would be monumental. Perhaps it will come through learning of a more advanced extraterrestrial civilisation, or perhaps we will find that we have somehow created sentient, superintelligent AI. Preparing for first contact with either of these forms demands serious engagement between physics, neuroscience and philosophy, a rigorous boundary between conversational competence and consciousness, and a good deal of epistemic humility.■ Susan Schneider is the founding director of the Centre for the Future of AI, Mind & Society and a professor at Florida Atlantic University. She is the author of several books including “Artificial You: AI and the Future of Your Mind”. This article was downloaded by zlibrary from https://www.economist.com/by-invitation/2026/08/20/dont-mistake-chatbot- intelligence-for-consciousness
The search for consciousness inside LLMs Could more brain-like chips provide a path to consciousness?
The search for consciousness inside LLMs Scientists are trying to figure out whether algorithms could one day wake up and feel Aug 20th 2026 IN A RECENT experiment on Claude Sonnet 4.5, a large language model (LLM) from Anthropic, researchers asked it to count to five and, at the same time, “introspect deeply”. The artificial-intelligence model complied, returning: “One . Two . Three . Four . Five .” So far, so normal. During the task, the researchers were watching what was going on inside the LLM’s many layers of artificial neural networks. When a user asks a question, the words are turned into chunks of text (tokens) that are then converted to numbers. These are then passed through the layers of artificial neurons until they reach the final layer, which produces a token. Repeat that process a bunch of times each second and you get a series of tokens that turn