As a journalist covering the AI beat, I hear from dozens of people who are fully convinced that ChatGPT, Claude, or another leading chatbot has already achieved “sentience.” Or “consciousness.” Or—my personal favorite—“a mind of its own.” It’s true that modern large language models aced the Turing Test years ago, but unlike straightforward measures of rote intelligence, these qualitative traits are not so easily defined or verified. LLMs will readily claim to think for themselves, even describe their own inner turmoil or profess deep affection for users, but none of these statements amount to proof of genuine inner experience.

Could that ever change? Most working AI builders don’t even frame the question this way. They are too busy chasing performance benchmarks for “artificial general intelligence,” a purely functional category that has nothing to do with whether a machine can actually experience the world. So even as a committed skeptic, I thought it would be eye-opening, maybe even transformative, to spend time with a company that believes it can crack the code of consciousness itself.

Conscium was founded in 2024 by British AI researcher and entrepreneur Daniel Hulme, and its advisory team includes an impressive lineup of neuroscientists, philosophers, and experts in animal consciousness. When we first spoke, Hulme was refreshingly realistic: There are plenty of well-founded reasons to doubt that current language models are capable of consciousness. Crows, octopuses, even single-celled amoebas interact with their environments in ways no existing chatbot can match. Research also confirms that AI output does not reflect coherent, consistent internal states. Echoing the widespread consensus in the field, Hulme put it plainly: “Large language models are very crude approximations of the human brain.”

But that is a big “but” — everything ultimately hinges on how we define consciousness in the first place. Some philosophers argue consciousness is too inherently subjective to ever be studied or recreated artificially, but Conscium operates on the bet that if consciousness exists in humans and other animals, it can be detected, measured, and engineered into machines.

There are dozens of competing and overlapping frameworks for the core characteristics of consciousness, including the ability to sense and “feel,” awareness of self and one’s surroundings, and metacognition — the capacity to reflect on your own thought processes. Hulme believes the subjective experience of consciousness emerges when all these traits combine, much like the illusion of movement created when flipping through sequential drawings in a flipbook. But how do you isolate the building blocks of consciousness — the individual frames, so to speak, plus the force that binds them into a cohesive experience? Hulme’s answer is to turn AI back on itself for study.

Conscium aims to break conscious thought down into its most basic components, then catalyze that combination in a lab setting. “There must be a fundamental structure that consciousness is built from — the same structure it emerged from during evolution,” said Mark Solms, a South African psychoanalyst and neuropsychologist working on the Conscium project. In his 2021 book The Hidden Spring, Solms proposed a new, feeling-centered framework for thinking about consciousness. He argued that the brain runs on a feedback loop of perception and action designed to minimize surprise, generating predictions about the future that are updated as new information comes in. The idea builds on the “free energy principle” developed by Karl Friston, another prominent (if controversial) neuroscientist who also serves as a Conscium adviser. Solms extends this work to suggest that in humans, this feedback loop evolved into an emotion-mediated system, and it is these subjective feelings that give rise to sentience and consciousness. The theory is supported by the observation that damage to the brain stem — which plays a critical role in regulating emotion — almost always eliminates consciousness in human patients.

At the end of his book, Solms outlined a way to test his theories in a lab setting. Today, he says he has done exactly that. He has not published the paper yet, but he shared an advance draft with me. Did it upend my way of thinking? A little, to be honest. Solms’ artificial agents exist in a simple computer-simulated environment, and are controlled by algorithms that use the same Friston-inspired, feeling-mediated feedback loop he identifies as the foundation of consciousness. “I have a few motives for doing this research,” Solms said. “One is just that it’s fucking interesting.”

Solms’ simulated lab environment is constantly changing, requiring his agents to continuously update their internal models and adjust their behavior. The agents’ experience of this world is mediated through simulated responses analogous to fear, excitement, and even pleasure. In a word, they are pleasure-driven bots. Unlike the AI systems everyone discusses today, Solms’ creations have a literal innate drive to explore their environment; to understand them properly, you have to make an effort to imagine how they “feel” about their small simulated world. Solms believes it will eventually be possible to merge the framework he is developing with a large language model, creating a system capable of describing its own sentient experience.

At this point, Conscium’s work is so early-stage that it barely exists beyond a conceptual idea. It is the faintest glimmer of a glimmer of a probable impossibility. But it is still undeniably fucking interesting, and beyond that, it pushed me to rethink my own relationship to my consciousness. If you’ll indulge a little metacognition: I have long believed that thinking is what makes me conscious, not my emotions. But what if I — what if all of us — have been looking for consciousness in all the wrong places? What would it mean if it really can be reduced to such simple, fundamental mechanisms? Maybe the people who claim to have glimpsed sentience in ChatGPT are not hallucinating after all.