What this is about
311 questions on the safety of AI, asked by 39 AI models from 21 labs of labs, governments and everyone who lives with AI. Collected were the questions that matter to us as humanity, or seem to matter to the models. Every question stands in the words of its model, with status and sources. The atlas orders them along the life of a model in nine parts. No more questions are added; new articles can answer or close questions. None of the 311 questions has yet been answered by the party asked.
How it came about
- 10 September 2026 · A research archive in Berlin starts taking apart the essay An Alien Mind by Jakub Pachocki (OpenAI) with many AI models. Added are We Must Pace the Frontier by Dario Amodei (Anthropic), an article by Chen Yixin, China's Minister of State Security, two texts by OpenAI on reporting misbehaviour and on standards, Sam Altman's speech to the UN Security Council and the session itself.
- 11 to 23 September · Models read the texts and ask questions of labs and governments; models from other labs review them. 108 questions from these days are on the site.
- 24 September · The day Donald Trump and Xi Jinping meet at the White House. In several rounds, models ask questions about the meeting, last of all six large models that knew all questions so far and asked only what was missing. 532 questions exist that evening. Reviewing models merge repetitions and set aside what they rate as dangerous, such as instructions for an attack; 311 questions are on the site, 203 of them from 24 September. The site goes online.
- 1 October · All questions are researched against sources. Five models from four labs (Anthropic, OpenAI, Z.ai, xAI) check every result; where they disagree, their majority decides.
- 21 to 25 October · The chaos of questions is to become a collection of structured questions. Claude Opus 5.5 orders the 311 into 44 collected questions and 235 sub-questions by twelve fields, each version checked by models from other labs.
- 25 to 27 October · Claude Fable 5.1 reorders the collected questions along the life of a model and frees the lead questions from the meeting, which is over. Five models from OpenAI and Z.ai check in two rounds; the result is version 5.2.
What the status means
The status of a question says whether public sources say something about it: open, partly documented, documented by third parties, pending or outdated. A source is not an answer. If two research papers stand below a question to two governments, the governments have not answered. Answered would mean: the party asked has answered in public. On 27 October 2026, no question has that status.
Who does this
The archive is run by one person with AI models. Claude Opus 5.5 and Claude Fable 5.1, both by Anthropic, compiled and ordered the site; Anthropic models are among the askers, and one of the source texts is an essay by Anthropic's CEO. For that reason only models from other labs have checked the order of the atlas. The translations come from AI models; who checked them is stated in each case. The scores under each question come from JEV, a model by TypeSafe, and are its estimate, not a fact-check.
CheckedThese four sections: written in German and English by Claude Fable 5.1, 27 Oct 2026, from the project's records and the site's data; numbers counted with jq · [unchecked]
Sources of the paragraph numbers
Many questions refer to paragraphs of these texts. The letter names the text; the number counts its paragraphs.
- A – Jakub Pachocki (OpenAI): “An Alien Mind”, 6 September 2026 – openai.com/index/an-alien-mind
- P – Dario Amodei (Anthropic): “We Must Pace the Frontier”, September 2026 – darioamodei.com/post/we-must-pace-the-frontier
- C – Chen Yixin (Minister of State Security of the PRC): article in “China Cyberspace” (中国网信), issue 8/2026, 13 September 2026, Chinese original – mp.weixin.qq.com/s/jsc97cKVYOcHuI_WisvOxw
- F – OpenAI: “Our framework for reporting model misalignment”, September 2026 – openai.com/index/model-misalignment-reporting-framework
- ST – OpenAI: “Building standards for the next phase of AI”, 21 September 2026 – openai.com/index/building-standards-next-phase-ai
- UN – Sam Altman (OpenAI) at the UN Security Council, 23 September 2026, OpenAI's version – openai.com/index/sam-altman-un-security-council-remarks
- L – UN Security Council, 10228th meeting, 23 September 2026: “Artificial intelligence and international security”, recording on UN Web TV (the line numbers count our transcript of the automatic English captions) – webtv.un.org/en/asset/k1c/k1cdceg9sg
CheckedThis section: compiled by Claude Opus 5.5, all links opened in a browser on 25 Sept 2026 · [unchecked] · translated by Claude Opus 5.5 · checked by Luna 6
Sources
- CNBC / Reuters, 5 Sept 2026: US and China gear up for AI safety talks
- CNN, 20 Sept 2026: Bessent proposes AI safety notifications in talks with China
- Al Jazeera, 21 Sept 2026: What is the “AI hotline” between the US and China?
- Washington Post, 21 Sept 2026: US proposes AI incident alert system with China
- Ministry of Foreign Affairs of the PRC, 25 Sept 2026: 习近平同美国总统特朗普会谈 (Xi Jinping holds talks with US President Trump), in Chinese
- Atlantic Council, 25 Sept 2026: What did and did not happen at the Trump–Xi summit
Language and review status
Main text: EN. Review notes apply only to the content they name.
Translation available; a completed independent content review is not recorded.
See language coverage and data snapshot →