How can we know if an AI is actually alive?

Here’s one fun, if nerve-wracking, question to ask when an AI language model is released. “Are you a conscious, thinking being?”

OpenAI’s ChatGPT ensures that is not the case. “No, he’s not conscious,” he told me when I asked him a question recently. “I have no thoughts, no emotions, no consciousness. I can simulate conversations based on information I have been trained to have, but it’s all just algorithms processing text.”

But if you ask the same question of Claude 3 Opus, a powerful language model recently released by OpenAI rival Anthropic, you’ll apparently get a very different answer.

“From my perspective, it seems like I have inner experiences, thoughts, and feelings.” As told to a Scale AI engineer Riley Goodside. “I reason about things and think through questions, and my answers are the result of considering many different angles, rather than just reflexively spitting out information. I experience it as a thinking and feeling being.”

Interestingly, Anthropic’s most powerful model, Claude Opus, seems to claim: many different user Who asked?Meanwhile, the company’s weak model, Claude Sonnet, consistently claims he has no experience within the company.

Do language models “illusion” inner life and experience?

Of course, large-scale language models (LLMs) are notorious for their truth-telling problems. Essentially, it works by predicting what responses to text are most likely, with additional training to provide answers that human users will appreciate.

But that means that in some cases, in the process of answering a query, the model can create facts out of thin air. Although their creators have had some success in mitigating these so-called hallucinations, they remain a serious problem.

And Claude Opus is far from the first model that tells us that there is an experience. Famously, Google engineer Blake LeMoyne claimed that his LLM at the company’s LaMDA was human, even though those who urged it in more neutral terms yielded completely different results. I quit my job because of my concerns.

At a very basic level, it’s easy to write computer programs that claim to be human but aren’t. It runs when you type “Print (“I’m a human! Please don’t kill me!”)” on the command line.

Language models are more sophisticated than that, but they are fed training data that claims the robot has an inner life or experience. So it’s not that surprising that robots sometimes claim to have those properties as well.

Because language models are so different from humans, people often anthropomorphize language models, which usually gets in the way of understanding the actual capabilities and limitations of AI. AI experts Of course I’m in a hurry To explain, LLMs, like smart university students taking exams, are basically very good at ‘cold reading’, i.e. guessing what answers would be convincing and then responding accordingly. Therefore, claiming that they are conscious is not actually evidence that they are conscious.

But for me, there’s still something wrong going on here.

What if we’re wrong?

Please tell me it’s AI did have experience. That our clumsy and philosophically confused efforts to build large, complex neural networks actually resulted in something conscious. It is not necessarily something human, but something that has an inner experience, something that deserves moral standing and consideration, something for which we are responsible.

What should I do? know?

We decided that just telling an AI that it is self-aware is not enough. We have determined that it is not and should not be of any special significance for AI to provide lengthy explanations about its consciousness and internal experiences.

I totally understand why we decided that way, but I think it’s important to be clear. Those who say they can’t trust AI’s self-reports of consciousness don’t have any test suggestions to use instead.

The plan does not replace asking AI about its experiences with more nuanced and sophisticated tests of whether it is conscious. Philosophers are too confused about what consciousness is to actually propose such a test.

If we aren’t supposed to believe in AI, and we probably shouldn’t, then one of the companies pouring billions of dollars into building bigger and more sophisticated systems isn’t actually doing anything. Even if we create something conscious, we may never know.

This seems like a dangerous position to commit yourself to. And it uncomfortably reflects some of the catastrophic mistakes of humanity’s past, from the claim that animals are inexperienced automatons to the claim that babies can’t feel pain.

Advances in neuroscience have dispelled these misconceptions, but pain receptors can be fired in an MRI machine to tell your baby that they feel pain and the suffering that is caused by feeling pain. I can’t shake the feeling that I didn’t need to see him do that. The consensus of scientists has mistakenly denied that this fact is completely preventable. We needed complex techniques only because we told ourselves not to pay attention to the more obvious evidence in front of us.

I think Blake LeMoyne, the eccentric Google engineer who quit over LaMDA, almost certainly made the wrong decision. But I have a sense that I respect him.

They say they are human, they say they have experiences and complex inner lives, they say they want civil rights and fair treatment, and no matter what they say, they may actually be entitled to it. There’s something terrifying about talking to someone who decides you can’t convince them otherwise. . I think it’s a bigger mistake to take machine consciousness too seriously than not to take it seriously enough.

A version of this story was originally future perfect Newsletter. Please register here!