(Süddeutsche Zeitung, February 4, 2025)

Nils Althaus

We should at least take that possibility seriously, philosopher Jeff Sebo says. But what would that mean for the way we treat computers—or robot vacuums?

A few years ago, the idea that artificial intelligence (AI) might become conscious was widely dismissed. Jeff Sebo, a philosophy professor at New York University, argues in his new book, The Moral Circle, that we can no longer afford to ignore the possibility.

SZ: Jeff Sebo, if you were angry, would you kick your robot vacuum?

Jeff Sebo: No, I wouldn’t.

Why not?

For one thing, it was pretty expensive (laughs). More importantly, I think the way I treat this robot today will affect how I treat future robots and AI systems. My robot vacuum isn’t conscious. But five or ten years from now, AI systems might be—and that would make them morally significant.

How would we know if an AI system were conscious?

We can’t observe conscious experience directly. But there are various indicators that tend to accompany consciousness. In AI systems, researchers might look for architectural or algorithmic features that support perception, learning, memory, goal-directed behavior, or flexible decision-making. No single indicator proves that a system is conscious. But the more of them a system displays, the more likely it is to be conscious.

Do today’s chatbots, such as ChatGPT and Claude, display any of these indicators?

We don’t yet see much evidence that they do. But that may be starting to change. Over the past month, AI companies have increasingly talked about deploying AI agents that can act more independently. Such systems might already display several indicators of consciousness. AI companies are working intensively to give their systems capabilities associated with conscious experience.

A minority of your colleagues argue that consciousness requires biological matter—that only neurons and brains can produce it. How would you respond?

Consider a thought experiment devised by David Chalmers (a philosopher and consciousness researcher at New York University—Editor’s note). Imagine gradually replacing every neuron in a person’s brain with a functionally equivalent computer chip. Each chip receives electrical signals, processes them, and passes new signals along, just as a neuron would. Throughout the process, the person reports that they are still having conscious experiences—even at the very end, when every neuron has been replaced. Are we really supposed to conclude that the person no longer experiences anything, even though they behave entirely normally and, for example, complain of pain?

Even if we accept that philosophical argument, what does it have to do with the technology that exists today?

Modern brain-computer interfaces, such as Neuralink’s, show that signals from neurons can interact with digital signals. You can be deeply skeptical about digital consciousness and still recognize that, given the evidence we have today, it would be premature to rule it out.

Why?

Ask philosophers how unlikely something has to be before we can morally disregard it, and some will put the threshold at one chance in a thousand—or even lower. That means that once the probability of an AI system being conscious reaches one in 1,000, we have to take the possibility seriously. Robert Long and I argued in a research report that, even under conservative assumptions, that threshold could realistically be crossed by 2030. The conclusion still holds if you assign an 80 percent probability to the view that consciousness requires a biological substrate. And given how little we understand about consciousness, 80 percent is already an extraordinarily confident estimate. Simply ignoring the possibility of artificial consciousness would violate every mainstream ethical approach to risk.

David Chalmers has also addressed this question. In a talk two years ago, he put the probability that AI systems would have conscious experiences by the early 2030s at more than 20 percent. Why is his estimate so much higher than yours?

My personal estimate is in that range, too. But our report wasn’t meant to express our personal views. We asked what follows when you consider a broad range of credible theories currently under discussion. When I speak with decisionmakers, I want to be able to say: Look, even if we grant this skeptical assumption and that skeptical assumption, we still end up with at least a one-in-1,000 chance that AI deserves moral consideration. From the perspective of risk ethics, that isn’t negligible.

What if we applied the same reasoning to a cactus or a calculator? How likely would they be to possess consciousness?

Using the same basic assumptions, both would receive a lower probability. Plants, however, at least warrant further study because their behavior is more complex.

If consciousness might be lurking everywhere, doesn’t that leave us completely paralyzed? I could never again turn off my computer with a clear conscience.

I think confusion is the right response. Consciousness is one of the greatest mysteries in science and philosophy. We struggle to explain it even in humans. At the same time, the stakes are extraordinarily high. If we develop conscious AI systems, we could create sentient beings in unimaginable numbers without even realizing it. That would be catastrophic. But the opposite mistake also carries risks. If we attribute consciousness to AI systems that aren’t conscious, we could pour resources into them that would be better devoted to humans or animals.

Suppose conscious AI becomes a reality. Where would that leave human interests?

I’m certainly not arguing that humans should make major sacrifices to promote AI welfare. But we shouldn’t accept the status quo, either. At the moment, only a few dozen people worldwide are working on questions of AI consciousness. We could easily put hundreds or thousands of people to work on the issue without reducing our investments in human well-being, human rights, or the environment.

What should be done?

AI companies need to recognize that AI welfare is a serious issue. They should begin evaluating their systems for indicators of consciousness and develop policies for how to treat systems that might be conscious. The leading AI company Anthropic recently set a good example by hiring an expert on AI welfare. Others should follow.

Should ordinary users change their behavior, too?

For now, it’s enough to refrain from kicking our robot vacuums—and to get into the habit of treating them with a little respect.