The AI safety split is now state business: Huang pressed Amodei at the White House minutes after both signed

On Tuesday 30 September, the heads of the world's leading AI companies met President Donald Trump at the White House to present a united front on "superintelligence" — the term the administration has just adopted as its official name for what we have been calling artificial intelligence. The meeting ended with photographs, smiles and a signed declaration of principles. What happened before the photos was different.

Jensen Huang, CEO of Nvidia, and others in the room questioned Anthropic's Dario Amodei about his public warnings on AI risk, according to the Wall Street Journal. Amodei replied that it was important to be honest. The exchange came immediately after everyone had signed the document that supposedly united them. The fault line is clean: Amodei and OpenAI's Sam Altman have asked for more regulation and described existential-risk scenarios in public writing and testimony; Huang and Meta's Mark Zuckerberg have consistently rejected that framing. Zuckerberg drafted the agreement that emerged from the lunch; Huang sold it to the other labs to collect the signatures.

The vocabulary fight is part of the same campaign. Huang has spent weeks pushing "superintelligence" as a replacement for "artificial intelligence," and chaired the White House panel where Trump signed the executive order adopting the new word. A YouGov survey published this week puts that in perspective: 53% of respondents prefer "artificial intelligence," and only 9% prefer "superintelligence." For now this is an elite rebrand that has not reached the public — and it has a convenient political property. "Superintelligence" makes the risk debate sound like science fiction, which is a comfortable place to file systems already operating in hospitals, armies and legislatures.

Meanwhile Yann LeCun, Meta's AI chief and one of deep learning's founders, called Amodei "deluded" in a Fortune interview, arguing current language models have neither the capability nor the autonomy to produce the catastrophes he describes — a position consistent with his long-standing view that transformers have fundamental limits. That is a real technical disagreement and deserves to be treated as one, not as a corporate talking point.

Our thesis: read the incentives on every side, then read the documents. Huang sells the chips that train these models; regulation that slowed development would hit GPU demand directly, which does not invalidate his arguments but does mean they are not neutral. Amodei's incentives are not neutral either — a safety-forward posture is also product positioning, and rules written now tend to favour incumbents who can afford compliance. What breaks the symmetry is where each claim is written down. Anthropic's IPO prospectus reportedly warns investors explicitly that "advanced AI could pose catastrophic or existential risks to humanity." A signed declaration of principles costs nothing if you are wrong. A risk factor in a securities filing is a legal statement attached to liability, and that makes it the more informative artifact by a wide margin.

There is a third actor now, and it changes the board. Trump told TIME this week that the US could take stakes in OpenAI and Anthropic, offering no mechanism or timetable. Set that beside what the Anthropic prospectus already discloses — government as an explicit risk factor, export controls on its best models, a federal no-use order — and you get a company that could go public with the state as both potential shareholder and documented business risk. For Huang, who wants no new laws and free-flowing AI, a government equity position is the opposite of the goal. For Amodei, who has asked for supervision, it cuts both ways depending on the terms: an ally in rulemaking, or a constraint on independence.

Our reading: this was the most honest picture of the debate anyone has published in months, precisely because the disagreement survived the signing ceremony. Voluntary pacts between parties who do not share a premise do not produce governance; they produce photographs. The useful signals are elsewhere — in S-1 risk sections, export-control decisions, procurement rules and evaluation data. The destination still looks good from here: this technology can plausibly end diseases and buy humanity more healthy decades. But getting there means arguing in public with the actual words, in rooms where someone keeps minutes. Amodei holding his position in front of Huang, with the president watching, was at minimum the more coherent stance in the room.