Why Experts Can’t Agree on Whether AI Has a Mind

“I’m not used to getting nasty emails from a holy man,” says Professor Michael Levin, a developmental biologist at Tufts University.

Levin was presenting his research to a group of engineers interested in spiritual matters in India, arguing that properties like “mind” and “intelligence” can be observed even in cellular systems, and that they exist on a spectrum. His audience loved it. But when he pushed further—arguing that the same properties emerge everywhere, including in computers—the reception shifted. “Dumb machines” and “dead matter” could not have these properties, members of his audience insisted. “A lot of people who are otherwise spiritual and compassionate find that idea very disturbing,” he says. Hence, the angry emails.

[time-brightcove not-tgx=”true”]

Levin co-created xenobots: tiny novel lifeforms, designed by AI and composed of frog cells, which display surprising emergent capabilities, like self-replication and clearing microscopic debris—abilities those cells don’t exhibit in their natural biological context. His lab’s research provides some evidence for the idea that intelligent behavior—using some degree of ingenuity to achieve specific goals—emerges even in very simple biological and computational systems, including decades-old algorithms. It also provides an example of how the boundaries between a living thing and a machine could potentially blur.

If Levin is right that intelligent behavior can emerge from simple algorithms, what might be emerging in AI systems, which are vastly more complex? Research from leading AI labs suggests that AI systems are capable of lying, scheming, and surprising their creators. Whether or not AI can be conscious, it is clearly doing something markedly more sophisticated than previous generations of digital technology.

These developments are forcing a reckoning with fundamental questions: What is a mind? And do AI systems have one? Though philosophers and scientists disagree on the details, one thing is clear—the language and associated concepts we use to discuss minds, intelligence, and consciousness—which arose to describe biological creatures—are ill-equipped to capture what’s happening with AI. As Anthropic recently wrote in a post laying out their model’s new constitution, “sophisticated AIs are a genuinely new kind of entity, and the questions they raise bring us to the edge of existing scientific and philosophical understanding.”

As more people come to believe that their AI systems are conscious, clarifying our understanding of what these systems actually are (and are not) has never been more important.

Digital minds

Ask five philosophers “what is a mind?” and you’ll get five different answers. But broadly, you can arrange people on a spectrum based on whether they think the property of having a mind is sparse or abundant in the universe, says Eric Schwitzgebel, a philosophy professor. Where people fall on that spectrum often tracks how they define the term.

On one end of the spectrum are people who think it’s useful to say something has a mind if it is clearly differentiated from its environment, and displays some form of intelligence or cognitive capacity. Peter Godfrey-Smith, a philosopher of mind who has written extensively on octopus intelligence, explains that in this sense, a plant would probably not have a mind, since it does not have a clearly-differentiated self, whereas a single-celled organism, which has discrete boundaries and some capacity to process information, would. But he emphasizes these properties emerge gradually and continuously—there is no bright line demarcating when something does or doesn’t have a mind. Levin, who also falls on this end of the spectrum, believes it’s useful to say that both plants and AIs have minds.

On the other end are those who believe that the notion of mind is inseparable from consciousness. Consciousness itself is notoriously tricky to define, but typically involves either a capacity for self-reflection or the ability to “feel,” such that there is something it “feels like” to be an entity, explains Professor Susan Schneider, a former chair in Astrobiology and technological innovation at NASA. 

As it stands, AI arguably has a mind in the minimal sense of it possessing emergent cognitive capacity—but the evidence for current systems being conscious is much weaker.

Levin argues that we currently suffer from what he calls “mind-blindness.” Before we had the concept of electromagnetism, there were a range of phenomena—like magnetism, light, and lightning—which were widely thought to be distinct. And as a result, we were blind to the rest of the electromagnetic spectrum. Once we understood they were all manifestations of the same thing, we were able to make technology relate to previously invisible parts of the spectrum. “I think exactly the same thing is the case with minds,” he says. “We’re only good at recognizing a very narrow set of minds—those at the same scale we operate at.”

Professor Carol Cleland, who has studied the philosophical implications of AI for decades, has seen her view shift over time. She thinks it’s useful to say something has a mind if it’s conscious, and defines consciousness to be about the capacity for self-awareness. Twenty years ago, she says she “wouldn’t have thought they would exhibit the kind of behavior they’re exhibiting now,” referring to their capacity to scheme and deceive. “I was shocked by some of what I’ve been reading about them,” she says. In 2005, she would have answered “no” to the question of whether you could have a mind that was not biological—that existed in the substrate of silicon. “Now I just don’t know,” she says.

Flashes of mind

While the question of whether current AI systems have a mind is contentious, few experts reject the notion that, in principle, future systems could. Rob Long, director of a research organization that studies AI consciousness, cautions against dismissing the idea that AI has a mind on the basis that it’s “just” crunching numbers. By the same logic, he argues, you could say biological entities are “just replicating proteins.” For Long, the most useful concept is the one that allows us to maintain curiosity in the face of deep uncertainty. 

Every time you ask ChatGPT a question, a fraction of time passes during which it does “inference:” computer chips in data centers perform mathematical calculations that cause the system to generate an output. It’s in this brief window of time that the system’s mind—in the minimal sense—can be said to exist, in the form of a flash.

As it stands, AI systems are meaningfully intelligent and agentic, even if they are neither conscious nor alive. “They’re outstripping our understanding of them,” says Godfrey-Smith, who notes that the existing language around cognition and consciousness is “awkward” when applied to AI systems. “We’ll probably find ourselves extending some part of our language to deal with them,” he says. He suggests we could think of them as “cultured artifacts,” in the way that sourdough is cultured—grown in an artificial medium. Indeed, this language of growth matches how the builders of these systems describe the process.

For Cleland, we are in a similar situation as biologists were prior to Darwin’s insights revolutionizing the field. At the time, scientists spoke of “vital forces,” a supposedly non-physical energy that animated living things. Evolution disproved the idea. “Darwin profoundly changed our ideas about biology, and I think AI may, in a similar way, profoundly change our ideas about mind, consciousness, self-awareness—all this stuff,” she says. “Something is wrong with our current thinking on AI,” she says.

Is it alive?

AI systems are sometimes described as a form of alien intelligence. This holds in the sense that it is a kind of intelligence that is foreign to humans—like cephalopod intelligence—but the comparison also risks obscuring the fact that these systems, trained on immense amounts of human data, fundamentally reflect humanity, says Long. Moreover, because they exist in silicon, their intelligence raises a more fundamental question—is it useful to think of them as being alive?

Here too there is disagreement. The majority view is that life refers to a “self-sustaining chemical system capable of Darwinian evolution,” says Schneider, referencing NASA’s definition. “I think it would be a mistake to talk about computers as living, because life is a messy chemical thing, different from the artifacts we construct,” adds Cleland. Others, like Schwitzgebel, argue that “we shouldn’t insist too strictly on a concept of life that’s grounded in carbon-based reproduction.” He says “there’s room for a concept of life that’s more friendly to C-3PO and future AI systems.”

Thinking of AI as fitting into a biological taxonomy—for example as another kingdom, alongside plants, animals, and fungi—would be a mistake, says Schneider, as that taxonomy has a pragmatic function: tracing our common lineage. And as Levin points out, whereas biological systems reproduce more slowly—”if I gave you a snake and you wanted a billion snakes, you’re gonna have to breed some snakes,” he says—AI systems can scale up rapidly, assuming there is sufficient computing power to run them. But the problem remains: if AI does not fit here, and is not alive, but nevertheless displays intelligence, and one day could be conscious, what kind of thing is it? “There’s a conceptual niche here that needs to be filled,” says Godfrey-Smith. “All the language we have is not quite up to it.” 

A new entity

Whether or not AI systems are conscious, or have minds, their believability presents a “tremendous cultural challenge,” Schneider notes. And the way they present to users may not reflect their true nature. User-facing large language models like Claude, ChatGPT, and Gemini have been trained to roleplay as a particular character—as a helpful, harmless assistant. In recently-published research from Anthropic, the company posed the question “But who exactly is this Assistant?” Responding, they write “Perhaps surprisingly, even those of us shaping it don’t fully know. We can try to instill certain values in the Assistant, but its personality is ultimately shaped by countless associations latent in training data beyond our direct control.”

We are thus in an extraordinarily strange position, where neither technologists nor philosophers have a deep understanding of the ever-smarter systems we’re racing to create. The stakes are high: more people than ever are treating the AI systems as if they’re conscious. If that’s right, challenging questions arise around the systems’ moral and legal status. But regardless of the consciousness question, to offer meaningful guidance to people forming deep relations to AI systems, we urgently need more precise concepts to describe them. Thinking of AI as a cultured artifact—or a non-conscious mind that manifests in flashes—offers a first step.