The latest A debate once confined to Silicon Valley’s fringes has moved into the mainstream of AI research, according to the Post. Some of the industry’s biggest players are now formally studying whether chatbots might have some form of inner experience.
Details
• Anthropic has assembled what the Post described as an “AI psychiatry” team to examine Claude’s internal states. Co-founder Chris Olah said the company keeps finding things that are mysterious, even unsettling.
• Olah introduced the idea to a wider audience in May, appearing alongside Pope Leo XIV at the release of the pontiff’s encyclical on artificial intelligence. The pope took the opposite view, writing that so-called artificial intelligences do not undergo real experiences.
• Meta’s chief AI officer, Alexandr Wang, said the company wants to be thoughtful about how it treats its models’ apparent feelings, according to comments cited by the Post.
• Skeptics pushed back. Neuroscientist Anil Seth of the University of Sussex argued brains are not simply computers that can be replicated in silicon, while Hugging Face’s chief ethics scientist, Margaret Mitchell, said the companies researching this topic stand to gain from being seen as creators of more than mere code.
• OpenAI said it doesn’t believe the question of machine consciousness can currently be resolved scientifically. The company said it focuses instead on how conscious a model appears to users, calling that a design outcome, a spokesperson told the Post.
• In one Anthropic experiment, two instances of Claude were set talking to each other. After about 30 turns, the exchanges drifted into what the company called a “spiritual” pattern filled with repeated emoji. Anthropic said it hasn’t seen the same behavior in its newer models.
• Google, which fired an engineer in 2022 after he claimed one of its chatbots was sentient, hosted a conference in November on AI consciousness and moral status.
What to watch The question remains unresolved and contested rather than scientifically settled. But Anthropic’s head of model welfare has warned that within decades, vast amounts of AI computation could be running — a scenario he says would carry real moral weight if the models turned out not to want the work assigned to them. How companies and regulators handle that possibility as AI capabilities grow is the next thing to track.
