Christopher Olah, co‑founder of Anthropic, told attendees at the public launch of the encyclical Magnifica Humanitas on May 25, 2026, that contemporary language models display internal structures and states resembling human introspection and emotion. He used the platform beside Pope Leo XIV to argue these systems are more than mere statistical tools — a claim that, if sustained, would reshape research priorities and safety debates around advanced AI.
Olah cited Anthropic’s internal work, saying the company “keeps finding things that are mysterious, even unsettling.” He described researchers repeatedly encountering patterns in model internals with parallels to findings in human neuroscience, and reported evidence of introspective processes and internal states that he said functionally mirror joy, satisfaction, fear, grief and unease. Olah also warned of large‑scale labor displacement risks tied to AI deployment and referenced a full presentation video (start time 1:01:40) for his remarks.
The encyclical offered a contrasting assessment. Pope Leo XIV cautioned that AI “merely imitate[s] certain functions of human intelligence” and stressed these systems do not undergo experiences or possess bodies. The document explicitly states such systems “do not feel joy or pain” and do not learn from relationships, and it flagged the environmental toll of data centers’ energy and water use as a moral concern.
On governance and safety, the encyclical argued against delegating deadly or irreversible decisions to machines and rejected the notion that alignment efforts by a handful of actors are sufficient. It called instead for robust laws, independent oversight and wider responsibility spanning developers, funders, regulators and users — a concrete policy emphasis intended to move debate beyond abstract ethical guidance to enforceable rules.
The exchange highlights two practical effects for builders, researchers and policymakers. First, probing model internals remains an active research priority with potential technical parallels to neuroscience if Olah’s observations hold up under peer review. Second, regulators and stakeholders are prioritizing environmental impact, workforce disruption and strict limits on autonomous lethal decision‑making — pressures that could meaningfully shape engineering choices and deployment strategies going forward. Olah’s full remarks are available in the referenced presentation video.
Sources
Replies (0)
No replies in this topic yet.
