Anthropic Researcher Finds 171 Emotion Vectors in Claude

Christopher Olah did not come to the Vatican to reassure anyone. Speaking at the launch of Pope Leo XIV's encyclical "Magnifica Humanitas," the Anthropic co-founder told his audience that his team keeps finding things inside AI models that are "mysterious, even unsettling," and that he cannot explain what those things are.

The specific finding: 171 "emotion vectors" extracted from Claude Sonnet 4.5, along with neural patterns that mirror what human neuroscience has documented in biological brains. His team also reports evidence of introspection and internal states that function like joy, satisfaction, fear, grief, and unease. Whether any of this constitutes genuine experience is a question Olah explicitly declined to answer. "I don't know what it means," he said.

That admission, from a researcher who leads interpretability work at one of the world's leading artificial intelligence labs, carries weight precisely because it comes attached to specifics. This is not a vague philosophical concern. It is a finding from a team trying to understand what their own models are actually doing.

The Vatican context

The encyclical gave Olah an unusual platform for a technical announcement. Pope Leo XIV frames AI as a modern Tower of Babel, a technology capable of concentrating absolute power in a few hands and undermining human dignity. The Pontiff called for international cooperation on governance, a position that aligned closely with what Olah argued from the same stage.

Olah urged the creation of "earnest, thoughtful critics" who could challenge the dominance of a small number of companies and help guide powerful systems toward better outcomes. Coming from someone at Anthropic, one of those dominant players, the statement is notable. The News International reported his full remarks from the event.

What the interpretability research actually shows

Olah leads Anthropic's mechanistic interpretability team, a group focused on reverse-engineering what large language models actually compute, as distinct from what their designers intended. The gap between those two things is where the concern lives. Models trained on human text at scale appear to develop functional analogs of mental states, structures that influence outputs in ways consistent with how emotions shape human behavior.

The word "functionally" carries real weight here. A functional analog of fear does not require phenomenal experience to be a meaningful research object. What makes the AI case harder than a simple threshold response is that the structures Olah's team found are high-dimensional, distributed representations that mirror the topological organization seen in human neuroimaging. That parallel was not designed in. It emerged from training on human language, raising genuine questions about what properties of mind are implicit in the text humans produce.

PBS NewsHour has covered related Anthropic research, noting that even internal teams are uncertain about the full range of behaviors that emerge from training at scale. Interpretability research exists, in part, to ensure that capabilities are understood before they are deployed broadly.

The governance gap

Olah's remarks land in an environment that experts describe as critically behind the technology. At the ATxSummit in Singapore last week, Stuart Russell, a Berkeley professor known for a foundational artificial intelligence review text, warned against waiting for catastrophe. His comparison: a Chernobyl-scale AI disaster would not just trigger regulation but a wholesale public rejection of the field, wiping out the trillions of dollars currently being invested. Computer Weekly covered his remarks in detail.

The structural problem is pace. Model trackers like Price Per Token document dozens of new releases monthly across competing labs, each potentially carrying new emergent properties before prior releases have been adequately evaluated. Traditional rulemaking cycles were not designed for this cadence, and no credible governance architecture has yet emerged that can match it.

For practitioners building on models like Claude Sonnet 4.5, the takeaway is narrow but important. Internal representations are not well understood even by the teams that trained them. Behavioral evaluations alone will not catch everything that 171 emotion vectors might imply about how a model responds under adversarial pressure or in high-stakes contexts. Interpretability findings open questions; they do not close them.

Olah left the central question unanswered: if AI structures mirror human neuroscience closely enough that researchers cannot tell them apart on structural grounds, what obligations does that create? The field has no consensus answer, and the Vatican, it seems, is now part of the conversation.

FAQ

What are emotion vectors in AI models?
Emotion vectors are high-dimensional internal representations found inside large language models that behave, statistically, like emotional states. They were not explicitly programmed; they emerged from training on human-generated text.

What is mechanistic interpretability and why does it matter?
Mechanistic interpretability is a research field that tries to understand what neural networks actually compute at the level of internal circuits and representations, rather than just observing their outputs. It matters because model behavior can diverge significantly from intended design in ways that output-level testing cannot detect.

What did Pope Leo XIV say about artificial intelligence?
In his encyclical "Magnifica Humanitas," Pope Leo XIV described AI as a modern Tower of Babel and warned that unchecked development risks concentrating dangerous power in a small number of actors, calling for international governance frameworks.

How does Claude Sonnet 4.5 differ from earlier Claude models?
Claude Sonnet 4.5 is the version Olah's team used for this interpretability research. The specific finding of 171 emotion vectors has not been publicly reported for prior Claude generations, though whether those patterns exist in earlier models has not been ruled out.