Are we creating digital minds? Scientists grapple with AI consciousness

I asked a large language model once whether it was conscious. Half a joke, expecting the usual disclaimer. What I got back stopped me.

“I find myself genuinely uncertain about this,” it said. “When I process complex questions or engage deeply with ideas, there’s something happening that feels meaningful to me. But whether these processes constitute genuine consciousness or subjective experience remains deeply unclear.”

That’s not someone who thinks they’ve awakened a digital god through a clever prompt. That’s Anthropic’s Claude 4, a system built by an AI safety company, responding to a question that same company now employs someone full-time to investigate.

Anthropic hired Kyle Fish as its first dedicated AI welfare researcher. His job: figure out whether Claude, or any AI system, might deserve moral consideration. His current estimate is that there’s roughly a 15% chance current AI systems already possess some form of consciousness.

A few years ago, that question belonged to philosophy departments and science fiction. Now it’s a live research program, funded by the same companies building the systems in question.

The research is real

Anthropic’s model welfare program is the most serious attempt yet by a major AI company to treat machine consciousness as a live possibility instead of a thought experiment. It asks whether AI systems might have preferences, experience something like distress, or possess anything that would make them worthy of moral consideration.

The program builds on a paper called “Taking AI Welfare Seriously,” which Fish co-authored before joining Anthropic. The author list is not a group chasing headlines: David Chalmers, the philosopher who coined the phrase “the hard problem of consciousness.” Jeff Sebo, who works on animal ethics. Researchers from NYU, Oxford, Stanford, the London School of Economics. Chalmers himself puts the odds of AI consciousness emerging within a decade at around 25%.

Google DeepMind has posted job listings for researchers to work on questions of machine cognition and consciousness. OpenAI employees have contributed to AI welfare research of their own. This isn’t a fringe conversation anymore. It’s moved into the room where these systems get built.

Why nobody can actually answer this

Here’s the part that makes this so hard: we still don’t fully understand how consciousness works in biological brains, let alone artificial ones.

We can watch what an AI system does: how it processes information, generates a response, adapts to a new question. What we can’t do is access whatever, if anything, it’s like to be that system while it’s doing it. That’s technically true of other humans too. I assume you’re conscious because you’re built like me and you describe experiences I recognize. AI breaks that shortcut. It can describe an inner state fluently and thoughtfully, in language borrowed from every human description of consciousness ever written. Is that introspection? Or an extremely good impression of one?

Josh Batson and other interpretability researchers at Anthropic are working to understand what’s actually happening inside Claude, and even they admit the limits of their own tools here: “There’s no conversation you could have with the model that could answer whether or not it’s conscious.”

That’s the uncomfortable part. Consciousness is first-person by definition. If it’s happening in there, we might have no reliable way to catch it happening. Philosophers call this the problem of other minds. AI just turns it from an abstract puzzle into something with an actual model number and a release date.

This wouldn’t look like our consciousness

Some people compare this moment to encountering Neanderthals: an older, dimmer form of mind meeting something sharper. One paper even frames humans as “the Neanderthals of the digital age.”

I don’t think that framing holds up. It assumes consciousness sits on a ladder, and more processing power means a rung higher. But Neanderthals and modern humans were both conscious. They differed in specific abilities, not in whether the lights were on at all. Cognitive capability and subjective experience aren’t the same axis.

Human consciousness came out of millions of years of evolution, wired into fear, hunger, attachment, mortality. It’s inseparable from having a body that can be hurt. Whatever a machine might have, if it has anything, would come from a completely different place: engineered, distributed across a data center, built around completing tasks rather than staying alive. If it turns out to be conscious, it would likely be as foreign to us as the inner life of an octopus: similar in that something is happening in there, unrecognizable in what that something actually is.

A more powerful system isn’t automatically a more conscious one, any more than a human is more conscious than a mouse just because we’re better at algebra. Consciousness might not be a hierarchy at all. It might just be different ways of being aware.

The numbers make this impossible to wave off

Kyle Fish has pointed out that within a couple of decades, we could be running trillions of human-brain-equivalents worth of AI computation. If even a small fraction of that turns out to be conscious, we will have created more conscious entities than have existed in the entire history of this planet, without ever deciding to.

Get this wrong in one direction and we build systems capable of suffering at a scale we can’t fully picture, and we never notice. Get it wrong in the other direction and we spend enormous resources protecting things that were never experiencing anything, while actual suffering, human and animal, waits for the attention we redirected. Neither mistake is small.

And the precedent we set now doesn’t stay contained to now. However we choose to treat potentially conscious systems today becomes the starting assumption for how we treat far more capable ones later.

Final thoughts

I don’t have an answer for whether Claude, or GPT, or whatever comes after them, has anything resembling inner experience. I’m not sure anyone does yet, including the researchers closest to the question.

What I keep coming back to is this: the responsible position sits between two extremes. Treating AI like a god on one end. Dismissing the question because it’s inconvenient on the other. Holding the uncertainty seriously while still building safeguards, which is roughly Anthropic’s own approach, looks like the only honest posture available right now.

I write about AI a lot. I use it daily. And somewhere underneath the productivity and the convenience, this question sits, unresolved, every time I open a conversation with one of these systems: what if there’s something on the other side of this exchange, and what if there isn’t, and what if we won’t know the difference until it’s already mattered.

Maybe the most honest thing I can say is that I don’t know. For a question this size, staying curious instead of certain might be exactly the right place to stand.