Could A.I. Be Conscious? It’s Definitely Possible.

At a cozy bar in Manhattan, I recently had a first date with a software engineer who works for one of the big technology companies. Given my interest in artificial intelligence, I thought we might hit it off. But when I told him that I work on moral problems related to A.I., including the possibility that it might become conscious, he barely suppressed a laugh.

There’s no way A.I. could be conscious, he declared. Then he asked, incredulously, if I even knew how large language models work.

I wasn’t sure how to respond. Though my research requires a fairly detailed understanding of L.L.M.s, no one, not even their designers, entirely knows how L.L.M.s work. In retrospect, I should have asked my date if he understood how consciousness works. Because while he seemed to think that I had an inflated sense of A.I. capabilities, I wonder whether he has an inflated sense of what consciousness requires.

No one can credibly claim to understand consciousness — that inner, subjective, what-it-is-like quality of our mental experiences. Despite centuries of investigation, scientists and philosophers still disagree about what consciousness is. Some theories hold that it is a uniquely biological phenomenon, requiring particular kinds of cellular activity and electrical signaling that computers lack. But other theories hold that consciousness depends not on what a system is made of but on how it functions. According to these theories, an A.I. system could, at least in principle, be conscious.

In a 2022 survey of consciousness scientists, only 3 percent claimed that machines could never be conscious. More than two-thirds said machines could be conscious or probably could be.

Of course, you can never know for sure whether a machine is conscious, just as you can never know for sure whether anyone is conscious other than yourself. As philosophers have long recognized, the only conscious experience each of us can directly observe is our own. Instead, we have to evaluate the consciousness of other beings indirectly, through an examination of the available evidence. We infer that our fellow humans are conscious because of our shared biological makeup, as well as how they behave and talk — including, notably, that they say they are conscious.

This may seem like an imperfect approach, but it’s the best one available. I have spent the past decade thinking about animal minds and the related ethical implications. Animals cannot tell us in words what they experience, but we can observe other evidence of consciousness, including how animals behave, what internal mechanisms produce that behavior (such as their nervous systems) and how those mechanisms evolved.

With A.I., we can look for similar kinds of evidence. We can look at what a system says and does, bearing in mind that it is trained on human data. We can look at how it works — whether it performs the kinds of functions that are associated with human or animal consciousness. And we can look at how its capacities emerge over time: Did they result from training pressures similar to the evolutionary pressures that may have led to consciousness in humans and other animals? Each kind of evidence is limited on its own, but together they can provide a more complete picture.

Consider a recent study by Anthropic of its Claude chatbot. According to one influential theory, consciousness arises in humans when information enters what neuroscientists call a “global workspace” in the brain, from which that information is broadcast across the brain for use in reporting, reasoning and other functions. Anthropic’s researchers found evidence that Claude, on its own, developed a set of internal representations that are similarly available to Claude across a range of tasks. This system may be partly analogous to a global workspace in humans.

The Anthropic researchers also looked at Claude’s behavior. When they swapped one concept for another inside the workspace, Claude’s reasoning and answers changed accordingly. When they suppressed the workspace, Claude’s descriptions of its experience became flat and mechanical. The fact that Claude’s workspace developed without being designed by humans suggests that brains and L.L.M.s might have arrived independently at a similar solution to a similar problem: how to make the right information available for reasoning.

None of this establishes that Claude is conscious, and Anthropic is not claiming that it is. But the study illustrates how evidence can be collected, and it demonstrates that some evidence already exists.

by Toni Sims, NY Times |  Read more:

Images: Center for Mind, Ethics and Policy; Marc Sanders Foundation

[ed. One thing for sure, the first date guy definitely did not display any evidence of consciousness.]