Эта страница доступна на английском, немецком и китайском языках.
Questions and IDs
Every question has an ID made of topic and number, for example hotline-01. There are nine topics:
- hotline (Hotline): an AI incident hotline: who reports what, to whom
- cyber (Cyber): AI and cybersecurity
- verify (Verification): independent checks: who verifies whom
- origin (Origin): distillation, data flows and export controls
- redlines (Red lines): red lines, thresholds and pauses
- transparency (Transparency): transparency, redactions and shared tests
- trust (Trust): whether we can believe what AI models say about themselves
- humans (Humans): who decides: people, values and the pace of AI
- facts (Facts): open factual questions
Some questions used to have a different ID (S-, CN-, PFQ-, LFQ-, OF-). These old IDs are archived. A link with an old ID still leads to the right question.
Paragraph numbers. Some questions refer to a passage in an original text, for example A12. The letter tells you which text:
- A1–A49: Jakub Pachocki, “An Alien Mind” (OpenAI)
- P1–P51: Dario Amodei, “We Must Pace the Frontier” (Anthropic)
- C1–C26: article by Chen Yixin in the magazine “China Cyberspace”; the same number in the Chinese original and in our English translation
- F1–F33: OpenAI, “Our framework for reporting model misalignment”
- ST1–ST33: OpenAI, “Building standards for the next phase of AI”
- UN0–UN33: Sam Altman’s speech to the UN Security Council on 23 Sept 2026; this means a paragraph of the speech, not “the UN”
- L18–L1265: line of the transcript of the 10228th meeting of the UN Security Council. Not to be confused with the gaps L1–L12 in the atlas.
CheckedThis section: written by Claude Opus 5.5, 19 Nov 2026 · [unchecked]
Statuses
The status says what public sources say about a question, and who says it. A question only counts as answered when the party asked answers it in public. A source is not yet an answer.
- open We found no public source on the question.
- partly answered Public sources say something about part of the question.
- answered by third parties Someone other than the party asked has published something on it. The party asked has not answered.
- pending The answer depends on an event that is not over yet.
- outdated The question rests on an assumption that no longer holds.
- answered The party asked has answered in public. At present no question has this status.
The symbol stands in front of the question in the list. The colour is the same everywhere: in the status, in the bar and in the dot on the ID chip in the atlas. “As of” gives the day the status was last established.
CheckedThis section: written by Claude Opus 5.5, 19 Nov 2026 · [unchecked]
The atlas
The question atlas sorts all 311 original questions into 44 collected questions. It has two axes: the life of a model and the fields.
The life of a model: nine parts. On the page the nine sections of this path are called parts (“Part 3 of 9”). Every collected question belongs to exactly one of parts 1 to 8. Part 9 leads to the gaps.
- Building
- Watching
- Releasing
- Operating
- Reporting
- Passing on
- Deciding
- Understanding
- Missing
The twelve fields: the second axis. Every collected question also belongs to one of twelve fields. The field gives it the letter in its ID and its colour: SF-M1 belongs to field M.
- M: Reporting incidents
- F: Measuring capabilities
- P: Independent verification
- T: Pace and limits
- V: Trust in models
- A: Agents and connected systems
- H: Sharing and protecting models
- K: Peace and security
- O: Disclosure and accountability
- E: Participation and the common good
- W: Economy, markets and liability
- I: Information and the public
Collected question, sub-question, original question
- Original question: a question as a language model asked it, with its ID (for example hotline-01). Its text stays unchanged.
- Collected question (for example SF-M1): a lead question that bundles several original questions.
- Sub-question (T1, T2 …): a question beneath a collected question. Below it are the ID chips of its original questions. Every original question sits under exactly one sub-question: as its basis or, with a dashed outline, as a similar question.
- Note “This was asked about …”: the sub-question is asked in general terms. The note says what the original question was asked about.
Details on a collected question
- Applies to: whom the collected question concerns, one or more of four groups: all states; all labs and deployers; evaluators, standards bodies, researchers; everyone who lives with AI.
- To: who should answer. Mentioned: who appears in the original questions but is not asked to answer. Also affects: whom else the question affects.
- Origin (on each sub-question): the occasion on which the original questions were asked, one of six: US–China summit, general, essay “An Alien Mind”, article by Chen Yixin, essay “Pace the Frontier”, factual question from the editors.
Pie, bar and ID chips
- Pie: a small circle per collected question, in the colour of its field. The filled share is the share of its original questions that have sources (“3 of 11 with sources”). Empty means: no sources yet. Hollow circles at part 9 are gaps.
- Bar: the statuses of the original questions of a collected question, in the status colours: red open, ochre partly answered, blue-grey answered by third parties, blue pending, grey outdated. “2 of 3 open” counts the original questions with the status open.
- ID chip: every original question appears as a chip beneath its sub-question. The dot in it has the colour of its status. ◐ means: this original question carries a note “outdated” or “partly outdated”. This is separate from the status.
- Gap (L1 to L12): a field on which the 311 questions say almost nothing, with finding and research status. All gaps are on the Gaps page.
CheckedThis section: written by Claude Opus 5.5, 19 Nov 2026 · [unchecked]
Addressee
Every question shows whom it is aimed at. There are three cases:
- From the question text: on 22 Oct 2026 the model clef (Cloudflare) looked in the text of every question for whom it addresses. A question can be aimed at several addressees. If both the US government and the Chinese government are meant, it says “both governments”.
- by topic: the question names no one. The addressee is assigned by topic. If “· JEV” appears next to it, this assignment comes from JEV (TypeSafe).
- to everyone (open question): the question is not aimed at anyone in particular.
The colour shows the group: indigo for governments and states, blue for the UN and international bodies, teal for labs and industry, ochre for testing, standards and oversight, violet for society and research, grey for “to everyone”. There are no national colours.
- both governments
- UN
- AI labs
- evaluators
- researchers
- to everyone (open question)
CheckedThis section: written by Claude Opus 5.5, 19 Nov 2026 · [unchecked]
Badges
All badges have the same shape: a rounded outline with coloured text. The colour carries the meaning.
- open
- partly answered
- answered by third parties
- pending
- outdated
Status: dot and word in the colour of the status. Red means: an answer is still missing here. It does not mean anyone is evading.
- both governments
- MiniMax M3
- Summit
- Hotline
Addressee: colour of the group (see above). Model: symbol and name in the colour of the lab, as on The models. Occasion and topic: grey; a click shows all questions for that occasion or topic.
CheckedThis section: written by Claude Opus 5.5, 19 Nov 2026 · [unchecked]
The scores from JEV
Below each question are two scores from JEV (TypeSafe), a model. The estimate has the reference date 25 Sept 2026. Each score runs from 0 to 100; high means important or urgent. The number is JEV’s score, unchanged. The ring next to it is filled as far as the number is high. These are a model’s estimates, not a verdict on the question or on the lab that asked it.
- ⏰ Urgency = mean of four yes/no questions to JEV: would an answer be worth much less if it came only after the end of the month, after 90, after 180 or after 365 days?
- ❗ Importance = JEV’s rating on a scale with five levels, converted to 0 to 100. The scale runs from “changes nothing” to “whether a safeguard against serious harm from AI exists or works depends on the answer”. JEV was asked:“Who would have to change what they do once the question is answered?”
The page does not show JEV’s other scores.
CheckedThis section: written by Claude Opus 5.5, 19 Nov 2026 · [unchecked]
Sources from the archive
The page Sources from the archive shows reports from our archive on the 44 collected questions of the atlas. The page is in German. It shows only title, domain, date and link, no text from the archive.
What was asked. The model clef:27b (Cloudflare) read each of these reports once and answered for each of the 44 collected questions: “How much does this report contribute to answering the following question?” The answer is a level from 0 to 7. clef gives a share for each level, adding up to 100 %.
| Level | Short name | Meaning, as clef received it |
|---|---|---|
| Unrelated | The report is about something else. | |
| Same broad topic, does not help | Same broad topic, but nothing in it helps to answer this question. | |
| Mentioned in passing | Mentions the question or its subject only in passing. | |
| Background, no evidence | Useful background or context, no evidence on the question itself. | |
| Some citable evidence | Contains some facts, statements or decisions that could be cited as evidence. | |
| Substantial evidence | Contains substantial evidence (rules, data, measurements, decisions) on this question. | |
| Answers in part or contradicts | Directly answers part of the question, or explicitly rejects or contradicts what it asks. | |
| Answers directly and substantially | Directly and substantially answers the question. |
clef received the question and the levels in this English wording.
- Level of a collected question
- The level with the largest share.
- Per cent
- The share of exactly this level. “6 · 41 %” means: clef considers level 6 the most likely, at 41 %. The other 59 % are spread over other levels; the bars show all eight.
- Highest level of a report
- The highest level the report has for any of the 44 collected questions. The colour stripe and the filter “Höchste Stufe” (highest level) follow it.
- Threshold and hits
- A collected question counts as a hit when its level reaches the threshold. “mit Bezug” (related) counts all collected questions from level 1 upwards.
- Weg (path)
- The parts on the life of a model in the atlas that contain collected questions: 1 Building to 8 Understanding. Every collected question belongs to exactly one.
- Herkunft (origin)
- The collection in the archive the report comes from. Next to it is the domain of the original source.
- Colours of the IDs
- The colour of a collected question (for example SF-M4) is the colour of its field in the atlas, here M.
- unchecked
- The rating comes from clef alone. Nobody has yet checked it report by report.
CheckedThis section: written by Claude Opus 5.5, 19 Nov 2026 · [unchecked]
Review notes
- The line “Checked” below a text names who wrote it and who checked it. [unchecked] means: nobody has checked it yet.
- On every question the “Checked” panel names who established and checked its status, who translated it and where the scores come from, with model and effort.
- The symbol before a model name shows its lab, as on The models.
CheckedThis section: written by Claude Opus 5.5, 19 Nov 2026 · [unchecked]
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 →