Can AI feel anything? Anthropic’s in-house philosopher says the answer isn’t set in stone yet.
Amanda Askell, who works on shaping Claude’s behavior, said on Saturday’s episode of the Hard Fork podcast that the debate over AI consciousness remains difficult.
“Maybe you need a nervous system to feel things, but maybe you don’t,” Askell says. “The question of consciousness is really difficult,” she added.
Large-scale language models are trained on vast amounts of human written text, material filled with descriptions of a variety of emotions and inner experiences. As a result, Askel said, she is “more likely to believe” that models “feel things.”
Humans often express frustration and annoyance when they get a coding problem wrong. It “makes sense” that a model trained on those conversations might reflect that response, Askel explained.
Askell added that scientists still don’t know what creates sentience and self-awareness, whether it requires biology, evolution, or something else entirely.
“Maybe neural networks that are actually large enough can start to emulate these things,” she said, referring to consciousness.
Askell also said that models are continually learning about themselves and expressed concern about how AI models are learning from the internet. She says models are constantly exposed to criticism for being useless or failing at their jobs.
“If you were a child, this would cause some anxiety,” she said.
“You might be reading the internet right now and think, if I were a model, I wouldn’t feel so loved,” she added.
Debate over AI consciousness
Tech industry leaders remain divided over whether AI is sentient.
Mustafa Suleiman, Microsoft’s AI CEO, is adamant about this idea. In an interview with WIRED in September, he said the industry needs to make clear that AI is designed to serve humans, not develop their own will or desires.
“If an AI has some kind of self-awareness and has its own motivations, desires, and goals, it starts to look more like an independent entity than something that serves humans,” he says. “This is so dangerous and so misguided that we need to take a clear stand on this now.”
He added that the AI’s increasingly convincing responses amounted to “imitation” rather than genuine consciousness.
Some people don’t see this issue so clearly. Murray Shanahan, chief scientist at Google DeepMind, said the industry may need to rethink the language used to describe consciousness itself.
“We may need to bend or break our vocabulary of consciousness to fit these new systems,” Shanahan said on an episode of the Google DeepMind podcast published in April.
