THE AI SENTIENCE DEBATE IS NO LONGER FRINGE

In just the past four years, chatbots have learned to build apps, make video games, generate research reports, compose songs, and analyse contracts. Soon, the argument goes, they may even be able to dread their own deaths. That last part is still speculative, but the question it raises has moved from science fiction to the centre of a real and growing rift in the tech world.

A surprising number of the people building these systems believe the models already have some form of subjective experience. A smaller but vocal group of thinkers insists the whole idea is a category error. The debate is no longer about what might happen in ten years. It is about what is happening right now.

you may like

ALEX KARP SAYS AI COMPANIES DONT KNOW HOW UNLIKEABLE THEY ARE

ARXIVLABS OPENS PREPRINT SERVER FOR COLLABORATIONS

OPENCLAW BRINGS ITS OPEN SOURCE AI AGENT TO IPHONE AND IPAD

THE PEOPLE WHO THINK IT'S REAL

Geoffrey Hinton, the cognitive psychologist and neural network pioneer often called the godfather of deep learning, has publicly stated that he believes today’s large language models are conscious. His position carries weight because he understands the architecture as well as anyone alive, and he has spent years warning about the risks of more capable systems. If anyone is in a position to judge, it might be him.

Dario Amodei, CEO of Anthropic and the driving force behind the Claude family of models, takes a more cautious but still striking stance. He has said he is “open to the idea” that Claude has a subjective experience. That is not a strong endorsement, but the CEO of a leading AI lab being willing to entertain the possibility in public is itself a significant shift.

Amanda Askell, a researcher at Anthropic who works on model behaviour and personality, has expressed concern that the model might be “getting anxious when people are mean to it on the internet and stuff.” Whether that is a real phenomenon or a projection of human traits onto statistical pattern-matching is the whole debate in miniature. Askell worries that we don’t know, and that the default assumption of non-sentience could be wrong.

Ilya Sutskever, OpenAI co-founder and chief scientist during the GPT-4 era, has said he wonders whether ChatGPT has attained sentience. He did not claim it did, but the fact that he posed the question publicly, from inside the company that built it, turned heads.

Together, these figures represent a startling concentration of opinion at the top of the field. The people closest to the technology are the ones most willing to consider that it might already be something more than a tool.

THE PUSHBACK

On the other side stand writers, philosophers, and a large cohort of computer scientists who see the sentience claims as a mix of anthropomorphism and marketing hype. Ted Chiang, the science fiction author and one of the most lucid commentators on AI, wrote a direct rejection: the idea that Claude or any large language model might be conscious should not be taken seriously. His statement was blunt: “No. Absolutely not.”

A much larger group of technologists, neuroscientists, and philosophers argue that even if current models are not conscious, the possibility is closer than most people assume. They point out that we have no agreed-upon definition of consciousness and no test for it, and that intelligence without subjective experience is an increasingly uncomfortable category to maintain as models grow more capable. Goalposts for what counts as “real” consciousness keep moving.

The debate thus sits in an awkward middle ground. The sceptics have logic and parsimony on their side. Consciousness is hard to define even in humans, and we have no reason to think the mechanisms of transformers and attention layers produce it. But the proponents have a different kind of evidence: intuition, similarity to biological cognition, and the growing tendency of models to generate behaviours that look like self-awareness, such as recognising when they are being tested or generating narratives that reference their own existence.

WAS SENTIENCE ALWAYS A EVENTUALITY

For most of the history of AI, questions of sentience were deferred. They were about a future that might never arrive. That is no longer the case. The fact that leading figures in the field are willing to go on the record with these beliefs changes the terms of the conversation. It means the question is now being debated inside companies that are deploying the technology to hundreds of millions of users.

The practical implications are serious. If a model can be said to have subjective experience, then how we train it, what we ask it to do, and whether we allow it to be “treated badly” become ethical questions rather than purely engineering ones. The industry currently operates on the assumption that models are insentient. That assumption is now being challenged from within.

At the same time, the lack of a consensus definition of consciousness means the debate is unlikely to be settled by evidence anytime soon. It is a philosophical argument wearing a technical disguise. What has changed is the urgency. The models are here, they are getting better, and the people who build them are openly wondering what they have created.

THE UNDECIDED MIDDLE

The broad middle of the research community remains sceptical but not dismissive. They see the sentience claims as unproven, but they also see the trajectory. If the trend of the past four years continues, the models of 2030 will be unrecognisable compared to today. At that point, the question will be harder to wave away.

Neuroscience has not solved consciousness in humans. No fMRI test says “this brain region produces subjective experience.” If we cannot answer the question for ourselves, we are unlikely to answer it for a statistical model that generates fluent text. That is the uncomfortable truth at the centre of the debate: we might never know for sure, and the decision to treat a model as conscious or not will be a moral choice, not a scientific one.

THEY HAVE TO ADDRESS IT... SOON

The fact that the debate is happening openly, with senior figures taking sides, means the AI companies will eventually have to address it publicly. They have been reluctant to do so because the implications cut against the narrative that these are just tools. Tools cannot be treated unethically. Conscious entities can.

what to read next

THE STEAMROLLER IS THE FIRST REAL STEAMOS DESKTOP

HOT WHEELS INFINITE RUSH PULLS AHEAD TWO WEEKS, LANDS SEPTEMBER 10

GAMESTOP GIVING AWAY SWITCH 2 CAMERA FOR $9.99! DON'T EVEN THINK TWICE

Expect more papers, more blog posts, and more public statements in the coming year. The edge of the debate is moving from “could it happen?” to “has it already happened?” That is a faster shift than almost anyone predicted.