AI’s inner workings remain mysterious and unsettling, says Anthropic co-founder
text_fields
Anthropic co-founder Christopher Olah has said researchers still do not fully understand what happens inside advanced artificial intelligence systems, describing some discoveries as “mysterious” and “unsettling”.
Speaking at the Vatican during discussions on the future risks of AI, Olah said his research team studying the internal structure of AI models has found patterns that resemble aspects of human cognition and neuroscience.
“We keep finding things that are mysterious, even unsettling,” Olah said.
He added that researchers have observed structures that appear to mirror “joy, satisfaction, fear, grief, and unease” in functional terms, though he stopped short of claiming AI systems are conscious.
“I don't know what that means, but I think it warrants ongoing discernment,” he said.
Olah spoke during an event linked to a major Vatican document on artificial intelligence released by Pope Leo XIV. The 42,300-word encyclical titled “Magnifica Humanitas” warned about the potential dangers of AI and called for stronger global safeguards.
“Artificial intelligence needs to be disarmed,” the Pope wrote in the document, arguing that governments must introduce stronger controls over the technology.
Olah appeared to support calls for caution, warning that AI could displace human labour on a massive scale and stressing the need for ethical oversight.
“They are not the cold, calculating robots we were promised. They are made from us, from our words,” Olah said, describing AI systems as products of human language and thought rather than traditionally engineered machines.
Earlier this year, Anthropic co-founder Dario Amodei also acknowledged uncertainty around AI consciousness, saying there is no accepted scientific definition or reliable method to determine whether AI systems possess consciousness.
“We do not understand how our own AI creations work,” Amodei wrote in a 2025 essay.
AI expert Srinivas Padmanabhuni said many of the so-called emergent behaviours seen in AI systems may reflect thousands of years of human thought embedded in training data.
“The model recreates human-like cognitive structures because those structures are the most efficient way to navigate the geometry of human language,” he said.
