AI is disrupting all manner of industries, but one of the people at the forefront of its development believes that there are some areas that it shouldn’t yet touch.
At a Stanford event, an MBA student asked Demis Hassabis — Nobel laureate, CEO of Google DeepMind, and arguably the most consequential figure in AI today — a question that doesn’t get asked enough: what should AI not touch? His answer centered on a word that most engineers would rather avoid: consciousness.
Hassabis began with scope. “AI is going to be, in terms of the scientific world of things, a fully general technology. You can think of it as a Turing machine — that’s the way I think about it.” He drew a direct line from Turing’s theoretical framework to the human mind: “Our minds are actually fully general, so we’re kind of approximate Turing machines. Anything that’s computable, a Turing machine can compute. And most things we know about in the universe — non-quantum things — are computable.” The implication being that the systems being built today carry the same theoretical reach. “These systems that we’re building, they’re also going to be sort of Turing-powerful as well.”
But it’s precisely that scope that makes Hassabis want to slow down on certain questions. “There are very big questions to come that I think it would be better if we took more time over. One example that’s pretty topical right now is consciousness.” He was candid about where the science stands: “It’s not a very well-posed problem from philosophy and neuroscience still, although I think we all have intuitions as to what the important aspects of that are. My feeling is that current systems are not conscious — but others disagree.”
His recommendation was pointed. “What I would suggest, in terms of what area AI should not touch, is that we build our first systems as tools — intelligent tools. That’s enough of a challenge already, because that’s already AGI.” From there, he argued for a more deliberate sequencing: “Using those tools, I think we should study neuroscience and philosophy, and actually come up with a more rigorous definition of things like consciousness. I think that is possible. And then test things against that, and then maybe as a society decide if we want to cross the second Rubicon of trying to make entities that at least seem conscious to us.”
The phrase “second Rubicon” is worth sitting with. It implies a first one — the creation of AGI-level intelligent tools — which Hassabis treats as already formidable enough to demand full attention. What he’s resisting is the conflation of two separate projects. “Intelligence and consciousness are dissociable, so I don’t think you have to do that to have an intelligent system. I think it’s a choice.” He even suggested that you can already sense different orientations in deployed models: “You can probably feel that when you use some of the leading chatbots. There are differences in opinion that come through.” His conclusion: “It’d be better to take that as two steps. They’re both enormous for humanity.”
The broader context makes this more than a philosophical musing. The debate around AI consciousness has been gaining serious traction — philosopher David Chalmers has argued that current LLMs cannot be conclusively ruled out as conscious, and Google engineer Blake Lemoine’s claim that an early Google chatbot had developed personhood, however contested, showed that these questions have a way of arriving before the field is ready for them. Hassabis himself has elsewhere lamented that in an ideal world, AGI would have been built in a CERN-like collaborative atmosphere rather than the current competitive intensity — a setting where questions about consciousness could be worked through with care rather than discovered mid-sprint. DeepMind has reportedly hired a philosopher specifically to work on machine consciousness, which suggests the organisation takes these questions seriously enough to staff for them, even while the race continues.
What Hassabis is proposing, stripped of its technical framing, is a kind of epistemic humility that the industry doesn’t always reward. The pressure to ship, to scale, and to define AGI in ways that keep the finish line visible has led to a compression of exactly the kind of deliberation he’s calling for. His argument that intelligent tools and conscious entities are separable choices — and that building the former doesn’t require rushing toward the latter — is a useful corrective to the assumption that capability automatically implies a mandate. Whether the field listens is a different matter.