What first drew you to the puzzle of whether machines could ever be conscious? 

As a child my imagination was shaped by exposure to a lot of sci-fi. The Star Trek character of Commander Data – an android who wants to be human – really fascinated me. The question of whether this robot could be conscious was such an interesting one. Although it was fun to think about, it seemed like a really distant possibility. Now though, it's a live possibility that we need to take seriously. That's why I've started to direct my research to the question.

Which breakthrough or idea in AI has most surprised you about the possibility of consciousness?

For me the breakthrough that's most surprising is how convincingly AI can appear consciousness. We can discuss what constitutes a good test for consciousness, but the thing that's really going to fix your opinion is the gut feeling of something conscious. Large Language Models can, in some cases, be very good at provoking that gut feeling. I thought it would take longer for AI to get that stage.

How do we tell the difference between a machine that seems conscious and one that really is?

Some of the things current AI does can really make it seem conscious, but I'm pretty confident that appearance is an illusion. What we need to do is think about why the AI is doing what it does. If a human says 'I'm a conscious being with my own experiences of the world' then, other things being equal, their utterance will probably be caused by them being a conscious being with their own experiences of the world.

Things aren't quite so simple with AI. Its outputs are caused by what it has learned from its training data, and that training data includes lots of humans talking about their conscious. Borrowing an idea from my colleague Murray Shanahan, you can think of large language models as advanced role-players. They are really good at adopting roles based on what they've learned from their training data. That can include playing the role of a conscious robot. But none of what it says is actually indicative of consciousness. 

Now, when we get to more advanced future AI it might not be so easy to debunk the appearance of consciousness. That's when we'll need a good test for consciousness. But in my view, no good test of AI consciousness is available.

What ethical dilemmas about AI consciousness are you most excited to explore with the public?

If AI is conscious then we ought to treat it like it's conscious, and if it's not conscious then we ought not to treat it like it's conscious. The dilemma is what to do when we're uncertain about AI consciousness. Should we treat it like any other object? Should we treat it as conscious just in case it is? Either policy comes with big risks. I think this is where some serious public debate is needed. This isn't the kind of question that can be outsourced to some experts!

Why does thinking seriously about artificial consciousness matter for our society today?

One reason it matters is because of the huge ethical implications I mention above. Another reason is that we risk drifting into a society that becomes polarised on the issue of AI consciousness. In my experience, people either think it's just obvious that AI can't be conscious or just obvious that it can be. Talking across that division is important so we can make sensible balanced decisions.

If people leave your talk with one big insight about conscious machines, what should it be?

Well, it's that we should be agnostic about whether AI consciousness is possible. If you come into the talk with a confident opinion one way or the other, I hope you'll leave with less confidence. I think facing up to this uncertainty is very important.