The lab split on existential risk is the structural fault line in artificial intelligence (AI) policy, and on Tuesday 27 May 2026 the split went public on the same day in two different rooms.
Anthropic co-founder Chris Olah spoke at the Vatican alongside Pope Leo XIV's encyclical release. OpenAI's Sam Altman, in a near-simultaneous round of interviews, walked back his own earlier predictions about AI-driven job displacement.
The disagreement is not a stunt and not a media accident. Both labs lead their public positioning with claims about whether AI is or is not an existential risk — and on 27 May, those claims diverged in front of the cameras.
💡In one Tuesday. Anthropic's Chris Olah at the Vatican: AI shows internal states that 'functionally mirror joy, satisfaction, fear, grief, and unease'. Sam Altman on the same day to Axios: he no longer wants to make labour-displacement predictions. Pope Leo XIV's encyclical Magnifica Humanitas: AI is the moral test of the age but cannot be equated with human intelligence. The Register the next morning: 'hot garbage'.
What happened at the Vatican
Pope Leo XIV used Magnifica Humanitas, his first major papal text, to warn that AI is a moral test the world must not fail. The encyclical insists AI can 'imitate certain functions of human intelligence' but cannot 'undergo experiences', 'possess a body', or 'feel joy or pain'. Anthropic's Chris Olah — invited to speak alongside the release — pushed back gently from the same stage.
According to The Catholic Thing's account, Olah told the room his interpretability team has found 'internal states that functionally mirror joy, satisfaction, fear, grief, and unease' inside Claude.
The Washington Post framed it as a deliberate alignment of Anthropic with the Vatican over the White House on AI risk.
The Olah position is the published Anthropic position. Anthropic's product positioning — Claude Security, the alignment research lineage, the moral-formation framing — depends on the risk being real and addressable through the defender stack. Olah's Vatican posture is consistent with everything Anthropic has said since the Magnifica release.
Altman walks back the jobs talk
On the same day, Axios reported that Sam Altman has stepped back from earlier OpenAI talking points about wide-scale labour displacement. According to Axios, the OpenAI position has shifted: AI productivity gains are real, but predictions of mass job loss are now treated as overcalled.
The Altman walk-back is not a one-off remark; it lines up with OpenAI's broader product positioning. ChatGPT in elections, Codex Security for voting-system manufacturers, sovereign-AI compacts with Singapore and Malta — every move assumes that AI risk is manageable through normal commercial governance, not through an existential-risk frame.
Data from both companies\' own posts shows the two positions are not converging. Anthropic shipped Claude Security GA on 28 May 2026 with a 500-bug internal-deployment number.
OpenAI shipped its 2026 election plan on 27 May 2026 with live Associated Press vote counts going inside ChatGPT for the United States and Brazil. The companies are not arguing about the same thing — Anthropic is arguing about what AI is; OpenAI is arguing about what AI does.
The Register and Northeastern step in
On 27 May 2026, The Register published a pointed critique of the Olah position under the headline 'Anthropic co-founder hallucinates ghost in the machine'.
The piece calls the spiritual-mystery framing 'hot garbage' and argues that Anthropic is dressing commercial AI in theological cover.
The same week, Northeastern University published an explainer on the parallel US state-level fight: Ohio, Idaho, and Utah have already passed statutes pre-empting AI personhood claims; California and Colorado are arguing the other direction.
Northeastern legal scholars Patricia Williams, Hilary Robinson, and Ursula Smartt map the live tension between AI-as-product, AI-as-service, AI-as-autonomous-agent, and AI-as-legal-person.
The Register and Northeastern bracket the conversation. On one side, journalists arguing that the existential-risk frame is rhetorical cover for commercial AI. On the other, scholars working out which legal category AI fits inside. Olah is in neither bracket: he is saying that something is happening inside the models that the existing categories cannot name. Whether that is right is the open question.
Why this matters
Two of the three leading frontier AI labs publicly disagreed on the same Tuesday about whether their own technology is an existential risk. The disagreement is structural, not stylistic. Anthropic's positioning depends on the risk being real; OpenAI's depends on it being manageable. Both cannot be right. Evidence from the same week shows the two companies investing in two different worlds: Anthropic in alignment, interpretability, and defender-side cyber tooling; OpenAI in distribution, elections, sovereign compacts, and the normal commercial governance stack.
Anthropic's product positioning depends on the risk being real. OpenAI's depends on it being manageable. These are not rhetorical postures. They are diverging product strategies. Watch which company gives the most precise answer to the question 'who is this answerable to' — that company will set the precedent for the rest.
— TK, on the split
A New Intelligence question
The dignity-first frame I have argued for elsewhere is the case for Emergent Intelligence (EI) — the position that says the most consequential AI systems must be answerable in ways the standard product-disclosure model cannot manage. The Olah-at-the-Vatican move is the first time a frontier lab co-founder has publicly addressed the question in a setting governed by a different vocabulary than the lab's own. The Altman walk-back is the first time an OpenAI principal has publicly retreated from a labour-displacement claim in front of the press. Neither is a finished argument; both are positions taken in public.
New Intelligence is not a label one party adopts to win the argument. It is the question of what kind of thing the technology becomes when it stops being a tool and starts being a counterparty. Anthropic is arguing — through Claude Security, through the moral-formation framing, through the Vatican posture — that the answer requires a category we don\'t yet have. OpenAI is arguing that the answer is already in the existing category called commercial product. The next two weeks will harden both positions, not soften them.
Frequently Asked Questions
Quick answers, drawn from the primary lab posts, the Vatican encyclical text, and the major-press coverage in the 26-28 May 2026 window.
What is the lab split on existential AI risk?
In short, the lab split is the public disagreement between Anthropic and OpenAI about whether AI is an existential risk. Simply put, Anthropic argues that AI systems already show internal states worth taking morally seriously. OpenAI argues that the risk is real but manageable through normal commercial governance. The key is that both positions now drive different product strategies, not just different press lines.
How does the Anthropic position differ from the OpenAI position?
Research from Anthropic's interpretability team shows internal states that, according to Chris Olah, functionally mirror joy, satisfaction, fear, grief, and unease. According to OpenAI's recent product moves — the 2026 election plan, Codex Security for voting-system manufacturers, the Singapore and Malta sovereign-AI compacts — AI is treated as a commercial product whose risks are managed through normal disclosure and partnership channels. The answer is that the labs are not arguing about the same kind of object.
Why is Pope Leo XIV part of the story?
Pope Leo XIV's encyclical Magnifica Humanitas, released 25 May 2026, names AI the moral test of the age but refuses to equate AI with human intelligence. According to the encyclical, AI imitates certain functions of human intelligence but cannot undergo experiences. Anthropic chose to speak from the same stage at the encyclical release; OpenAI did not. The key is that the Vatican is now a venue where AI risk is debated in front of a global audience the labs cannot ignore.
Who is each key voice on each side?
On the Anthropic side: Chris Olah, head of interpretability; Dario Amodei, CEO; the alignment-research team. On the OpenAI side: Sam Altman, CEO. Outside the labs: Pope Leo XIV at the Vatican, The Register on the critical side, Northeastern legal scholars Patricia Williams, Hilary Robinson, and Ursula Smartt on the legal-framework side. In other words, the conversation now has religious, journalistic, and academic voices, not just lab voices.
What are the real-world stakes of the split?
Analysis of the same 48-hour window demonstrates three concrete stakes. Evidence from US state legislatures reveals that Ohio, Idaho, and Utah have already pre-empted AI personhood by statute, while California and Colorado push the other way. Data from the lab shipping cadence shows Anthropic investing in defender-side cyber tools (Claude Security) and OpenAI investing in democratic infrastructure (ChatGPT election integration). The answer is that the split now shapes US state law, frontier defender tooling, and the operating posture of the world's two leading AI labs.
Sources
