{ "version": "0.5.0-clarity", "reviewedAt": "2026-09-15", "status": "Editorial guide", "camps": [ { "id": "advance", "title": "Keep advancing", "summary": "Build useful AI. Target specific harms.", "note": "The case for continuing development and broad access.", "colour": "blue" }, { "id": "safeguards", "title": "Advance with safeguards", "summary": "Pursue the benefits. Make safety a condition.", "note": "The case for development with stronger checks and governance.", "colour": "yellow" }, { "id": "pause", "title": "Slow or pause", "summary": "Set limits until control can be demonstrated.", "note": "The case for restricting the most powerful systems until conditions are met.", "colour": "pink" } ], "sources": [ { "id": "lecun-interview", "title": "Yann LeCun in conversation with Lex Fridman", "publisher": "Lex Fridman Podcast", "url": "https://lexfridman.com/yann-lecun-3-transcript/", "date": "2024-03-07", "kind": "Primary interview", "locator": "2:08:48, AI doomers; 1:43:48, Open source" }, { "id": "ng-pause", "title": "Why the 6-month AI pause is a bad idea", "publisher": "DeepLearning.AI", "url": "https://www.youtube.com/watch?v=BY9KV8uCtj4", "date": "2023-04-07", "kind": "Primary discussion", "locator": "Andrew Ng and Yann LeCun discuss the proposed pause" }, { "id": "gates-2026", "title": "The turbulent AI era is here. The choices we make now are critical.", "publisher": "Bill Gates, Gates Notes", "url": "https://www.gatesnotes.com/a-turbulent-ai-era-and-critical-choices-to-make", "date": "2026-08-26", "kind": "Primary essay", "locator": "This time really is different; The transition to AI comes with three big risks" }, { "id": "amodei-2026", "title": "The Adolescence of Technology", "publisher": "Dario Amodei", "url": "https://darioamodei.com/essay/the-adolescence-of-technology", "date": "2026-01", "kind": "Primary essay", "locator": "Autonomy risks; misuse; concentration of power" }, { "id": "hinton-nobel", "title": "Geoffrey Hinton: official Nobel interview", "publisher": "Nobel Prize", "url": "https://www.youtube.com/watch?v=66WiF8fXL0k", "date": "2025-03-14", "kind": "Primary interview", "locator": "Recorded 6 December 2024; risks at 3:13, timelines at 9:04" }, { "id": "superintelligence", "title": "Statement on Superintelligence", "publisher": "Future of Life Institute", "url": "https://superintelligence-statement.org/", "date": "2025", "kind": "Joint statement", "locator": "Statement and signatory list; conditional prohibition on superintelligence" }, { "id": "fli-signatories", "title": "AI: statement and initial signatories", "publisher": "Future of Life Institute", "url": "https://futureoflife.org/focus-area/artificial-intelligence/", "date": null, "kind": "Organiser's record", "locator": "Initial signatories include Geoffrey Hinton and Yoshua Bengio; checked 15 September 2026" }, { "id": "tegmark-fli", "title": "Our Position on AI", "publisher": "Future of Life Institute", "url": "https://futureoflife.org/our-position-on-ai/", "date": null, "kind": "Institutional position", "locator": "Pause on frontier experiments and moratorium on artificial superintelligence; checked 15 September 2026" }, { "id": "altman-2025", "title": "The Gentle Singularity", "publisher": "Sam Altman", "url": "https://blog.samaltman.com/the-gentle-singularity", "date": "2025-06-10", "kind": "Primary essay", "locator": "Iterative deployment, alignment and broad access" }, { "id": "hassabis-2025", "title": "Responsible AI: Our 2024 report and ongoing work", "publisher": "Demis Hassabis and James Manyika, Google", "url": "https://blog.google/innovation-and-ai/products/responsible-ai-2024-report-ongoing-work/", "date": "2025-02-04", "kind": "Co-authored company statement", "locator": "Frontier Safety Framework; responsible development and deployment" }, { "id": "pause-2023", "title": "Pause Giant AI Experiments: An Open Letter", "publisher": "Future of Life Institute", "url": "https://futureoflife.org/open-letter/pause-giant-ai-experiments/", "date": "2023-03-22", "kind": "Joint statement", "locator": "Six-month pause proposal and signatories, including Musk and Russell" }, { "id": "grok-4", "title": "Grok 4", "publisher": "xAI", "url": "https://x.ai/news/grok-4", "date": "2025-07-09", "kind": "Company announcement", "locator": "Evidence of model development, not a statement of Musk's personal risk estimate" }, { "id": "mitchell-paper", "title": "Why AI is Harder Than We Think", "publisher": "Melanie Mitchell", "url": "https://arxiv.org/abs/2104.12871", "date": "2021-04-26", "kind": "Research paper", "locator": "Four recurring fallacies about AI; historical context, not a current policy pledge" }, { "id": "gebru-response", "title": "Statement from the listed authors of Stochastic Parrots on the AI pause letter", "publisher": "Timnit Gebru and co-authors, DAIR", "url": "https://www.dair-institute.org/blog/letter-statement-March2023", "date": "2023-03-31", "kind": "Co-authored statement", "locator": "Transparency, accountability and existing harms" }, { "id": "normal-technology", "title": "AI as Normal Technology", "publisher": "Arvind Narayanan and Sayash Kapoor", "url": "https://www.normaltech.ai/p/ai-as-normal-technology", "date": "2025-04-15", "kind": "Research essay", "locator": "Human institutions, gradual adoption and the limits of superintelligence narratives" }, { "id": "safety-report", "title": "International AI Safety Report 2026", "publisher": "International AI Safety Report", "url": "https://internationalaisafetyreport.org/publication/international-ai-safety-report-2026", "date": "2026-02-03", "kind": "Research synthesis", "locator": "Malicious use, loss of control, systemic risks and risk management" }, { "id": "pichai-lex-2025", "title": "Sundar Pichai in conversation with Lex Fridman", "publisher": "Lex Fridman Podcast", "url": "https://lexfridman.com/sundar-pichai-transcript/", "date": "2025-06-05", "kind": "Primary interview", "locator": "44:34–47:32, AI safety. No numerical estimate is attributed to Pichai.", "checkedAt": "2026-09-17" }, { "id": "pichai-summit-2026", "title": "Remarks at the AI Impact Summit", "publisher": "Google", "url": "https://blog.google/company-news/inside-google/message-ceo/sundar-pichai-ai-impact-summit-2026/", "date": "2026-02-19", "kind": "Primary speech", "locator": "Speaker identification: CEO, Google and Alphabet.", "checkedAt": "2026-09-17" }, { "id": "nadella-copilot-2026", "title": "Announcing Copilot leadership update", "publisher": "Microsoft", "url": "https://blogs.microsoft.com/blog/2026/03/17/announcing-copilot-leadership-update/", "date": "2026-03-17", "kind": "Primary signed leadership messages", "locator": "Nadella: human control, agency and governance; signatures identify both leaders.", "checkedAt": "2026-09-17" }, { "id": "huang-g20-2026", "title": "Connect the World: G20 AI remarks", "publisher": "CNN", "url": "https://transcripts.cnn.com/show/ctw/date/2026-09-03/segment/02", "date": "2026-09-03", "kind": "Primary speech excerpt in broadcast transcript", "locator": "Jensen Huang video clip: precise regulation of actual harms and continued innovation. Date is broadcast date.", "checkedAt": "2026-09-17" }, { "id": "zuckerberg-future-2026", "title": "The Future is for Everyone", "publisher": "Meta", "url": "https://about.fb.com/news/2026/08/the-future-is-for-everyone/", "date": "2026-08-10", "kind": "Primary signed essay", "locator": "Mark Zuckerberg: distributed personal superintelligence, openness and the balance of power.", "checkedAt": "2026-09-17" }, { "id": "suleyman-humanist-2025", "title": "Towards Humanist Superintelligence", "publisher": "Microsoft AI", "url": "https://microsoft.ai/news/towards-humanist-superintelligence/", "date": "2025-11-06", "kind": "Primary authored essay", "locator": "By Mustafa Suleyman; A humanist future section.", "checkedAt": "2026-09-17" }, { "id": "su-senate-2025", "title": "Winning the AI Race: Strengthening U.S. Capabilities in Computing and Innovation", "publisher": "US Senate / Government Publishing Office", "url": "https://www.congress.gov/119/chrg/CHRG-119shrg61426/CHRG-119shrg61426.pdf", "date": "2025-05-08", "kind": "Primary testimony and written answers", "locator": "Su: PDF page 84 (printed page 80), regulatory approach; PDF pages 100–101 (printed pages 96–97), open ecosystems. Written answers were submitted after the hearing.", "checkedAt": "2026-09-17" }, { "id": "jassy-letter-2025", "title": "CEO Andy Jassy’s 2025 Letter to Shareholders", "publisher": "Amazon", "url": "https://www.aboutamazon.com/news/company-news/amazon-ceo-andy-jassy-2025-letter-to-shareholders/", "date": null, "kind": "Primary authored shareholder letter", "locator": "Published in 2026, discussing Q1 2026; heading refers to financial year 2025. Section beginning “When you identify disproportionate inflections”. Exact publication date not established from the opened page.", "checkedAt": "2026-09-17" }, { "id": "thinking-machines-human-2026", "title": "The future worth building is human", "publisher": "Thinking Machines Lab", "url": "https://thinkingmachines.ai/blog/the-future-worth-building-is-human/", "date": "2026-07-10", "kind": "Institutional statement", "locator": "Authored by the lab, not a personally signed Murati statement. Use only as institutional context.", "checkedAt": "2026-09-17" }, { "id": "thinking-machines-nvidia-2026", "title": "Thinking Machines Lab and NVIDIA Announce Long-Term Gigawatt-Scale Strategic Partnership", "publisher": "Thinking Machines Lab", "url": "https://thinkingmachines.ai/news/nvidia-partnership/", "date": "2026-03-10", "kind": "Primary company announcement", "locator": "Identifies Murati as co-founder and CEO and Huang as founder and CEO; documents commercial relationship.", "checkedAt": "2026-09-17" }, { "id": "mensch-europe-playbook", "title": "European AI: a playbook to own it", "publisher": "Mistral AI", "url": "https://europe.mistral.ai/", "date": null, "kind": "Primary CEO foreword and company policy paper", "locator": "Words from the CEO, Arthur Mensch. Page does not supply a precise publication date.", "checkedAt": "2026-09-17" }, { "id": "gomez-rules-2026", "title": "Who Gets to Define the Rules for AI?", "publisher": "Cohere", "url": "https://cohere.com/blog/who-gets-to-define-the-rules-for-ai", "date": null, "kind": "Primary authored essay", "locator": "By Aidan Gomez, Co-Founder & CEO. Sections on evidence, assurance and who writes the rules. No publication date visible in the opened text.", "checkedAt": "2026-09-17" }, { "id": "ssi-mission", "title": "Safe Superintelligence mission", "publisher": "Safe Superintelligence Inc.", "url": "https://ssi.inc/", "date": null, "kind": "Institutional mission statement", "locator": "Capabilities and safety in tandem; no date or individual byline on the current page.", "checkedAt": "2026-09-17" }, { "id": "sutskever-ceo-2025", "title": "Message to the SSI team and investors", "publisher": "Safe Superintelligence Inc.", "url": "https://ssi.inc/updates", "date": "2025-07-03", "kind": "Primary signed message", "locator": "Signed Ilya: assumes CEO role and affirms safe-superintelligence objective.", "checkedAt": "2026-09-17" }, { "id": "li-world-models-2026", "title": "Why governing world models is AI’s next big policy challenge", "publisher": "Stanford HAI", "url": "https://hai.stanford.edu/news/why-governing-world-models-is-ais-next-big-policy-challenge", "date": "2026-08-04", "kind": "Primary interview", "locator": "Fei-Fei Li’s answers on concentration/access, benchmarks and rigorous field testing. Do not attribute the other interviewees’ answers to her.", "checkedAt": "2026-09-17" }, { "id": "song-agents-2026", "title": "UC Berkeley computer scientist on the promise and perils of agentic AI", "publisher": "UC Berkeley", "url": "https://news.berkeley.edu/2026/08/19/uc-berkeley-computer-scientist-on-the-promise-and-perils-of-agentic-ai/", "date": "2026-08-19", "kind": "Primary interview", "locator": "Dawn Song’s answers on autonomy, security, evaluation and oversight.", "checkedAt": "2026-09-17" }, { "id": "dragan-framework-2026", "title": "Strengthening our Frontier Safety Framework", "publisher": "Google DeepMind", "url": "https://deepmind.google/blog/strengthening-our-frontier-safety-framework/", "date": "2026-04-17", "kind": "Primary co-authored policy statement", "locator": "Originally 22 September 2025; updated 17 April 2026. By Four Flynn, Helen King and Anca Dragan. Misalignment risks and safety case reviews.", "checkedAt": "2026-09-17" }, { "id": "toner-control-2026", "title": "The A.I.s Are Already Out of Control: interview excerpt", "publisher": "CSET, Georgetown University", "url": "https://cset.georgetown.edu/?p=25269", "date": "2026-08-18", "kind": "Institution-hosted excerpt of primary interview", "locator": "Toner’s quoted comparison of capability and control progress. This is an excerpt from a longer New York Times interview; it does not set out a detailed pause proposal.", "checkedAt": "2026-09-17" }, { "id": "whittaker-ai-now-2023", "title": "What is AI? Part 1, with Meredith Whittaker", "publisher": "AI Now Institute", "url": "https://ainowinstitute.org/publications/collection/what-is-ai-part-1-with-meredith-whittaker-ai-now-salons", "date": "2023-07-19", "kind": "Primary conversation excerpts", "locator": "Democratic control, surveillance and concentration of infrastructure; position is dated, not newly verified advocacy.", "checkedAt": "2026-09-17" }, { "id": "choi-open-minds-2026", "title": "David, Goliath, and the future of AI", "publisher": "Stanford Open Minds", "url": "https://openminds.stanford.edu/support/stories/david-goliath-and-the-future-of-ai", "date": "2026-05-19", "kind": "Institutional research profile with direct interview quotations", "locator": "Common sense, smaller models, access and pluralistic alignment.", "checkedAt": "2026-09-17" }, { "id": "choi-stanford-2025", "title": "NVIDIA’s Yejin Choi joins Stanford HAI", "publisher": "Stanford HAI", "url": "https://hai.stanford.edu/news/nvidias-yejin-choi-joins-stanford-hai", "date": "2025-01-16", "kind": "Institutional appointment announcement", "locator": "Academic appointment and NVIDIA affiliation; affiliation disclosure should retain its source date.", "checkedAt": "2026-09-17" }, { "id": "rossi-accountability-2026", "title": "The ethicist’s rule: Never let the machine take the blame", "publisher": "IBM Think", "url": "https://www.ibm.com/think/news/ethicists-rule-never-let-machine-take-blame-francesca-rossi", "date": "2026-06-22", "kind": "Primary interview", "locator": "Human accountability, limits of pre-release review, misuse and human agency.", "checkedAt": "2026-09-17" }, { "id": "raji-auditing-2026", "title": "Closing the AI Accountability Gap: Audits and Evaluation in the Age of Artificial Intelligence", "publisher": "UC Berkeley EECS", "url": "https://www2.eecs.berkeley.edu/Pubs/TechRpts/2026/EECS-2026-261.html", "date": "2026-08-14", "kind": "Primary doctoral research abstract", "locator": "Inioluwa Raji, report UCB/EECS-2026-261; evaluation methods and institutional accountability.", "checkedAt": "2026-09-17" }, { "id": "grace-pause-2026", "title": "AI pause: the case for ASAP", "publisher": "Katja Grace / LessWrong", "url": "https://www.lesswrong.com/posts/mEhS4wYTy9JXEpe9p/ai-pause-the-case-for-asap", "date": "2026-06-24", "kind": "Primary authored argument", "locator": "Main post by KatjaGrace; exclude comments and their claims.", "checkedAt": "2026-09-17" }, { "id": "crawford-atlas", "title": "Atlas of AI", "publisher": "Kate Crawford", "url": "https://katecrawford.net/atlas", "date": "2021", "kind": "Author’s account of her book", "locator": "Book synopsis: material inputs, inequality and political power. Date identifies the book, not an undated web-page update.", "checkedAt": "2026-09-17" }, { "id": "crawford-bio", "title": "Kate Crawford: biography", "publisher": "Kate Crawford", "url": "https://katecrawford.net/", "date": null, "kind": "Primary biography", "locator": "USC research professor and Microsoft Research senior principal researcher, checked 17 September 2026.", "checkedAt": "2026-09-17" }, { "id": "amodei-pace-2026", "title": "We Must Pace the Frontier", "publisher": "Dario Amodei", "url": "https://darioamodei.com/post/we-must-pace-the-frontier", "date": "2026-09", "kind": "Primary authored essay", "locator": "Coordinated pacing, independent evaluation and explicit qualification that pacing does not mean halting model training.", "checkedAt": "2026-09-17" }, { "id": "ceo-safety-2026", "title": "Advancing AI safety through state and federal action", "publisher": "OpenAI", "url": "https://openai.com/index/advancing-ai-safety-through-state-and-federal-action/", "date": "2026-07-15", "kind": "Institutional statement", "locator": "Identifies Sam Altman as OpenAI CEO and Demis Hassabis as Google DeepMind CEO. Used here only for role identification.", "checkedAt": "2026-09-17" }, { "id": "amodei-bio", "title": "Dario Amodei: biography", "publisher": "Dario Amodei", "url": "https://darioamodei.com/", "date": null, "kind": "Primary biography", "locator": "Identifies Amodei as Anthropic co-founder and CEO.", "checkedAt": "2026-09-17" }, { "id": "li-world-labs", "title": "About World Labs", "publisher": "World Labs", "url": "https://www.worldlabs.ai/about", "date": null, "kind": "Primary company biography", "locator": "Identifies Fei-Fei Li as a founder and describes the company’s commercial world-model research.", "checkedAt": "2026-09-17" }, { "id": "su-amd-bio", "title": "Dr. Lisa Su: Chair and Chief Executive Officer", "publisher": "AMD", "url": "https://www.amd.com/en/corporate/leadership/lisa-su.html", "date": null, "kind": "Primary company biography", "locator": "Current title, checked 17 September 2026.", "checkedAt": "2026-09-17" }, { "id": "grace-ai-impacts", "title": "About AI Impacts", "publisher": "AI Impacts", "url": "https://aiimpacts.org/about/", "date": null, "kind": "Primary organisation biography", "locator": "Katja Grace: co-founder and lead researcher; organisation based at MIRI.", "checkedAt": "2026-09-17" }, { "id": "whittaker-open-source-2026", "title": "Reframing Impact: AI Summit 2026 — Open Source", "publisher": "AI Now Institute, Aapti Institute and The Maybe", "url": "https://ainowinstitute.org/wp-content/uploads/2026/02/Reframing-Impact_Open-Source_Meredith-Whittaker-1.pdf", "date": "2026-02", "kind": "Primary interview excerpts", "locator": "PDF pages 2–6: Signal presidency; openness, infrastructure concentration and privacy-preserving applications.", "checkedAt": "2026-09-17" }, { "id": "military-wargames-2024", "title": "Escalation Risks from Language Models in Military and Diplomatic Decision-Making", "publisher": "Rivera et al. · arXiv", "url": "https://arxiv.org/abs/2401.03408", "date": "2024-01-07", "kind": "Primary research · simulated wargames", "locator": "Abstract and experimental results. Five models in simulations; the setup does not establish a real-world probability of escalation.", "checkedAt": "2026-09-17" }, { "id": "sacks-action-plan-2025", "title": "White House Unveils America’s AI Action Plan", "publisher": "The White House", "url": "https://www.whitehouse.gov/releases/2025/07/white-house-unveils-americas-ai-action-plan/", "date": "2025-07-23", "kind": "Primary attributed statement", "locator": "David Sacks’s named quotation on innovation, infrastructure, partnerships, workers and avoiding authoritarian uses. Distinguish his words from the surrounding administration announcement.", "checkedAt": "2026-10-01" }, { "id": "andreessen-save-world-2023", "title": "Why AI Will Save the World", "publisher": "Marc Andreessen / Andreessen Horowitz", "url": "https://a16z.com/ai-will-save-the-world/", "date": "2023-06-06", "kind": "Primary authored essay", "locator": "AI Risk #1: Will AI Kill Us All?; AI Risk #4: Will AI Lead To Bad People Doing Bad Things?; What Is To Be Done? A dated argument, not a newly issued position.", "checkedAt": "2026-10-01" }, { "id": "harari-interview-2026", "title": "Yuval Noah Harari: AI could be a great manipulative psychopath", "publisher": "EL PAÍS English", "url": "https://english.elpais.com/culture/2026-07-25/yuval-noah-harari-ai-could-be-a-great-manipulative-psychopath.html", "date": "2026-07-25", "kind": "Primary interview", "locator": "Harari’s answers on agents, financial control and the final AI-policy answer: disclosure, children’s AI relationships and legal personhood. Interview conducted 8 July; English publication dated 25 July, not the 26 July date on his media index.", "checkedAt": "2026-10-01" }, { "id": "harari-faqs", "title": "Frequently Asked Questions, answered by Yuval Noah Harari", "publisher": "Yuval Noah Harari", "url": "https://www.ynharari.com/faqs/", "date": null, "kind": "Primary authored answers", "locator": "Do you believe in technological determinism? The worst-case scenario is described as a possibility, not a prophecy. The page supplies no publication date.", "checkedAt": "2026-10-01" }, { "id": "harris-roadmap-2026", "title": "Here’s Our Roadmap to a Better AI Future", "publisher": "Center for Humane Technology / Your Undivided Attention", "url": "https://centerforhumanetechnology.substack.com/p/heres-our-roadmap-to-a-better-ai", "date": "2026-04-02", "kind": "Primary podcast transcript", "locator": "Tristan Harris’s own turns beginning “what we’re going to need ultimately are laws” and “AI agents should have driver’s licenses”. Do not attribute Aza Raskin’s or the policy team’s answers to Harris.", "checkedAt": "2026-10-01" }, { "id": "harris-slowdown-2026", "title": "The Most Hopeful (And Concerning) Moment Yet in AI", "publisher": "Center for Humane Technology / Your Undivided Attention", "url": "https://centerforhumanetechnology.substack.com/p/the-most-hopeful-and-concerning-moment", "date": "2026-09-17", "kind": "Primary podcast transcript", "locator": "Harris’s own proposals: shared safety research, careful testing, pre-approval for recursive self-improvement, insurance, liability and a right to warn. Read the end-of-transcript correction about Ajeya Cotra. Used for Harris’s advocacy, not independent verification of the incidents or percentages he describes.", "checkedAt": "2026-10-01" } ], "people": [ { "id": "yann-lecun", "name": "Yann LeCun", "role": "AI researcher · Turing Award recipient", "camp": "advance", "featured": true, "position": "Intelligence need not mean a desire to dominate.", "summary": "LeCun challenges the idea that greater intelligence inevitably creates a drive to seize power. He argues for open development and architectures that can be designed to remain controllable.", "action": "Continue research and broad access to AI.", "nuance": "His scepticism about takeover scenarios does not mean he dismisses misuse or concentrated corporate power.", "placement": "His opposition to a broad pause and support for open development place him here. This is an editorial grouping, not a self-applied label.", "counterpoint": "How would we establish that future, more capable systems will reliably stay within their intended limits?", "interests": "Helped develop deep learning and has championed open AI systems. The cited interview was given while he was Meta's chief AI scientist.", "sourceIds": [ "lecun-interview", "ng-pause" ], "connections": [ { "pathway": "control", "relation": "questions", "sourceIds": [ "lecun-interview" ] }, { "pathway": "power", "relation": "discusses", "sourceIds": [ "lecun-interview" ] } ], "history": [ { "date": "2024-03-07", "text": "Defends open development and challenges takeover arguments in a long-form interview.", "sourceIds": [ "lecun-interview" ] } ], "group": "research", "cardRole": "AI researcher", "aiScope": "Open AI research and future systems.", "actionKind": "Open development", "positionSourceIds": [ "lecun-interview", "ng-pause" ] }, { "id": "andrew-ng", "name": "Andrew Ng", "role": "AI researcher & educator", "camp": "advance", "featured": false, "position": "Stopping progress carries risks of its own.", "summary": "Ng argues that a broad pause would delay useful applications and that safety work should address specific harms. His discussion with LeCun makes the case for continuing development.", "action": "Keep building; tackle concrete risks.", "nuance": "Opposition to a general pause is not opposition to all safety measures or regulation.", "placement": "Placed here for his stated opposition to the six-month pause. The linked discussion is historical and dated.", "counterpoint": "How should we handle risks where waiting for clear evidence could leave too little time to prevent irreversible harm?", "interests": "Founded DeepLearning.AI and invests in AI applications through AI Fund.", "sourceIds": [ "ng-pause" ], "connections": [ { "pathway": "control", "relation": "questions", "sourceIds": [ "ng-pause" ] } ], "history": [ { "date": "2023-04-07", "text": "Publicly argues against a six-month pause in frontier AI development.", "sourceIds": [ "ng-pause" ] } ], "group": "research", "cardRole": "AI researcher & educator", "aiScope": "AI development broadly, including advanced models.", "actionKind": "Continue development", "positionSourceIds": [ "ng-pause" ] }, { "id": "bill-gates", "name": "Bill Gates", "role": "Microsoft co-founder & philanthropist", "camp": "safeguards", "featured": false, "qualifier": "Conditional position", "position": "The benefits are enormous. We are not ready for the risks.", "summary": "Gates calls for urgent preparation for AI-driven disruption, malicious use and possible loss of control. He sees major benefits, but says institutions are not preparing adequately.", "action": "Build national and international capacity to manage the transition.", "nuance": "In August 2026 he said he would likely support a credible global slowdown, while doubting one could be achieved. This makes his placement conditional.", "placement": "His practical emphasis is on safeguards and preparation during continuing development. His qualified support for slowing down is retained here.", "counterpoint": "Can institutions prepare quickly enough to keep pace with capabilities and commercial pressure?", "interests": "Gates discloses financial ties to technology and work with Microsoft and other AI companies through his foundation.", "sourceIds": [ "gates-2026" ], "connections": [ { "pathway": "control", "relation": "discusses", "sourceIds": [ "gates-2026" ] }, { "pathway": "biology", "relation": "discusses", "sourceIds": [ "gates-2026" ] }, { "pathway": "infrastructure", "relation": "discusses", "sourceIds": [ "gates-2026" ] } ], "history": [ { "date": "2026-08-26", "text": "Sets out a more urgent call for preparation and expresses conditional support for a credible slowdown.", "sourceIds": [ "gates-2026" ] } ], "group": "research", "cardRole": "Technology founder & philanthropist", "aiScope": "AI adoption and its wider economic and security effects.", "actionKind": "Prepare for the transition", "positionSourceIds": [ "gates-2026" ] }, { "id": "dario-amodei", "name": "Dario Amodei", "role": "Anthropic co-founder & CEO", "camp": "pause", "featured": true, "position": "Slow capability gains so safety can catch up.", "summary": "Amodei now argues for coordinated pacing of frontier capability progress, alongside independent evaluation and stronger safeguards.", "action": "Use independent evaluation and coordinated pacing across labs and governments.", "nuance": "He explicitly distinguishes pacing from halting model training or technical progress.", "placement": "The new essay advocates slowing capability advancement. This is an editorial placement in “Slow or pause”, qualified as pacing rather than a halt.", "counterpoint": "Can coordinated pacing be verified without giving incumbent labs excessive control?", "interests": "Leads Anthropic, a frontier AI developer. The proposal would affect its competitors as well as its own development.", "sourceIds": [ "amodei-2026", "amodei-pace-2026", "amodei-bio" ], "connections": [ { "pathway": "control", "relation": "discusses", "sourceIds": [ "amodei-2026" ] }, { "pathway": "biology", "relation": "discusses", "sourceIds": [ "amodei-2026" ] }, { "pathway": "power", "relation": "discusses", "sourceIds": [ "amodei-2026" ] } ], "history": [ { "date": "2026-01", "text": "Publishes an extended argument about managing the risks of powerful AI.", "sourceIds": [ "amodei-2026" ] }, { "date": "2026-09", "text": "Argues for coordinated pacing; explicitly distinguishes this from halting training.", "sourceIds": [ "amodei-pace-2026" ] } ], "qualifier": "Coordinated slowdown", "group": "leaders", "cardRole": "AI company leader", "aiScope": "Capability gains in the most powerful AI models.", "actionKind": "Coordinated slowdown", "positionSourceIds": [ "amodei-2026", "amodei-pace-2026" ] }, { "id": "geoffrey-hinton", "name": "Geoffrey Hinton", "role": "AI researcher · Nobel laureate", "camp": "pause", "featured": true, "position": "We may build something we cannot keep under control.", "summary": "Hinton warns that systems more capable than humans could become difficult to control. He has joined a call to prohibit superintelligence until there is broad scientific confidence in its safety and public support.", "action": "Put conditions on superintelligence and expand control research.", "nuance": "The statement concerns superintelligence: systems that substantially outperform humans across cognitive tasks. It is not a call to switch off every current AI application.", "placement": "His participation in the conditional prohibition statement grounds this placement; the organiser names him among the initial signatories.", "counterpoint": "What evidence would show that greater capability leads to uncontrollable power, and what might prevent that chain?", "interests": "A deep-learning pioneer and former Google researcher. His research background is relevant expertise, not proof that his predictions will occur.", "sourceIds": [ "hinton-nobel", "superintelligence", "fli-signatories" ], "connections": [ { "pathway": "control", "relation": "supports", "sourceIds": [ "hinton-nobel" ] } ], "history": [ { "date": "2024-12-06", "text": "Explains concerns about systems becoming smarter than humans in his Nobel interview.", "sourceIds": [ "hinton-nobel" ] }, { "date": "2025", "text": "Joins the conditional prohibition statement on superintelligence.", "sourceIds": [ "superintelligence", "fli-signatories" ] } ], "group": "research", "cardRole": "AI researcher", "aiScope": "Future AI more capable than humans across many tasks.", "actionKind": "Conditional restrictions", "positionSourceIds": [ "hinton-nobel", "superintelligence" ], "qualifier": "Conditional restrictions" }, { "id": "max-tegmark", "name": "Max Tegmark", "role": "Physicist · Future of Life Institute co-founder", "camp": "pause", "featured": false, "position": "Keep humanity in control. Set the boundaries first.", "summary": "Tegmark and the Future of Life Institute advocate limits on the most advanced AI systems, including a halt to superintelligence: systems that substantially outperform humans across cognitive tasks.", "action": "Restrict superintelligence until safety and public-consent conditions are met.", "nuance": "This position distinguishes beneficial AI tools from a race to systems that outperform humanity across cognitive tasks.", "placement": "Grounded in his involvement in the superintelligence statement and FLI's explicit institutional position.", "counterpoint": "How could a global restriction be agreed, verified and enforced?", "interests": "Co-founded an organisation that funds research and campaigns for restrictions on advanced AI.", "sourceIds": [ "superintelligence", "tegmark-fli" ], "connections": [ { "pathway": "control", "relation": "supports", "sourceIds": [ "tegmark-fli" ] }, { "pathway": "power", "relation": "discusses", "sourceIds": [ "tegmark-fli" ] } ], "history": [ { "date": "2025", "text": "Supports the call for a conditional prohibition on superintelligence.", "sourceIds": [ "superintelligence" ] } ], "group": "research", "cardRole": "AI researcher & campaigner", "aiScope": "Future AI more capable than humans across many tasks.", "actionKind": "Restrict superintelligence", "positionSourceIds": [ "superintelligence", "tegmark-fli" ], "qualifier": "Restrict superintelligence" }, { "id": "sam-altman", "name": "Sam Altman", "role": "OpenAI co-founder & CEO", "camp": "safeguards", "featured": false, "position": "Deploy, learn and widen access as AI gets more capable.", "summary": "Altman argues for releasing systems in stages, learning from their use and keeping them aligned with human intentions. He advocates broad access to the benefits of powerful AI.", "action": "Continue deployment while improving control and access.", "nuance": "This is a developer's proposed approach, not independent evidence that deployment is safe.", "placement": "The cited essay combines continued progress with work to keep systems aligned with human intentions.", "counterpoint": "Which failures can we safely learn from through deployment, and which would be irreversible?", "interests": "Co-founded OpenAI, which develops and deploys frontier AI.", "sourceIds": [ "altman-2025", "ceo-safety-2026" ], "connections": [ { "pathway": "control", "relation": "discusses", "sourceIds": [ "altman-2025" ] }, { "pathway": "power", "relation": "discusses", "sourceIds": [ "altman-2025" ] } ], "history": [], "group": "leaders", "cardRole": "AI company leader", "aiScope": "Increasingly capable AI systems and their deployment.", "actionKind": "Deploy with safeguards", "positionSourceIds": [ "altman-2025" ] }, { "id": "demis-hassabis", "name": "Demis Hassabis", "role": "Google DeepMind co-founder & CEO", "camp": "safeguards", "featured": false, "position": "Advance the science and build safety into the process.", "summary": "In a co-authored statement, Hassabis supports continued AI development alongside risk evaluations, deployment safeguards and governance.", "action": "Use capability evaluations and safety frameworks during development.", "nuance": "A company's safety framework describes its intended process; it does not establish that every risk is controlled.", "placement": "His co-authored statement explicitly combines bold innovation with responsible development.", "counterpoint": "How effective are evaluations at detecting dangerous behaviour before a system is widely used?", "interests": "Co-founded DeepMind. The source is a Google statement and should be read in that context.", "sourceIds": [ "hassabis-2025", "ceo-safety-2026" ], "connections": [ { "pathway": "control", "relation": "discusses", "sourceIds": [ "hassabis-2025" ] } ], "history": [], "group": "leaders", "cardRole": "AI company leader & researcher", "aiScope": "The most capable AI models under development.", "actionKind": "Safety checks during development", "positionSourceIds": [ "hassabis-2025" ] }, { "id": "yoshua-bengio", "name": "Yoshua Bengio", "role": "AI researcher · Turing Award recipient", "camp": "pause", "featured": false, "position": "Safety must be established before superintelligence.", "summary": "Bengio is an initial signatory of the statement calling for a conditional prohibition on superintelligence.", "action": "Require scientific safety confidence and public support first.", "nuance": "The proposal targets superintelligence, not all AI research or present-day uses.", "placement": "Based on an explicit public statement he signed, as recorded by its organiser.", "counterpoint": "What evidence would be sufficient to establish safety, and who should decide?", "interests": "A leading deep-learning researcher involved in AI safety research and advocacy.", "sourceIds": [ "superintelligence", "fli-signatories" ], "connections": [ { "pathway": "control", "relation": "discusses", "sourceIds": [ "superintelligence" ] } ], "history": [], "group": "research", "cardRole": "AI researcher", "aiScope": "Future AI more capable than humans across many tasks.", "actionKind": "Conditional restrictions", "positionSourceIds": [ "superintelligence" ], "qualifier": "Conditional restrictions" }, { "id": "stuart-russell", "name": "Stuart Russell", "role": "AI researcher · UC Berkeley", "camp": "pause", "featured": false, "qualifier": "Temporary training pause", "position": "Do not race ahead of our ability to control the systems.", "summary": "Russell signed the 2023 call for a pause on training systems more powerful than GPT-4, while safety protocols and governance were developed.", "action": "Pause the largest training runs while establishing safeguards.", "nuance": "This placement records a dated proposal. It does not imply that the original six-month proposal is still his complete position.", "placement": "Based on his signature on the 2023 pause letter. A fuller current profile remains part of the next research pass.", "counterpoint": "Would a pause meaningfully improve safety, and how could it be enforced internationally?", "interests": "Studies human-compatible AI and has advocated changes to how AI objectives are specified.", "sourceIds": [ "pause-2023" ], "connections": [ { "pathway": "control", "relation": "discusses", "sourceIds": [ "pause-2023" ] } ], "history": [], "group": "research", "cardRole": "AI researcher", "aiScope": "Training systems more powerful than GPT-4, in the 2023 letter.", "actionKind": "Temporary training pause", "positionSourceIds": [ "pause-2023" ] }, { "id": "elon-musk", "name": "Elon Musk", "role": "Technology founder", "camp": null, "featured": false, "qualifier": "Mixed signals", "position": "Signed a pause call, then founded a company building advanced AI.", "summary": "Musk signed the 2023 pause letter. The company he founded subsequently developed and released Grok models. Those facts do not resolve his present policy position into a single camp.", "action": "Current policy position left unplaced in this atlas.", "nuance": "Stated policy preferences and commercial activity are separate kinds of evidence.", "placement": "Unplaced: a historic signature should not be treated as a permanent policy commitment, and company activity alone cannot establish a personal view.", "counterpoint": "How do the dates and scope of his policy proposals relate to the systems his company develops?", "interests": "Founded xAI and has commercial interests in AI development.", "sourceIds": [ "pause-2023", "grok-4" ], "connections": [], "history": [ { "date": "2023-03-22", "text": "Signs the six-month pause letter.", "sourceIds": [ "pause-2023" ] }, { "date": "2025-07-09", "text": "xAI announces Grok 4, evidence of continuing frontier development.", "sourceIds": [ "grok-4" ] } ], "group": "leaders", "cardRole": "AI company leader", "aiScope": "The 2023 frontier-training pause call and later model development.", "actionKind": "Mixed position", "positionSourceIds": [ "pause-2023" ], "placementStatus": "mixed" }, { "id": "timnit-gebru", "name": "Timnit Gebru", "role": "AI researcher · DAIR founder", "camp": null, "featured": false, "qualifier": "Questions the framing", "position": "Whose harms get attention, and whose power gets protected?", "summary": "Gebru and co-authors challenge the pause letter's framing, emphasising existing harms, transparency and accountability.", "action": "Address present harms and strengthen accountability.", "nuance": "Criticism of extinction-focused narratives is not an endorsement of unrestricted development.", "placement": "Unplaced: a different framing of the problem does not map reliably onto a pace-of-development camp.", "counterpoint": "How should attention and resources be divided between harms already occurring and possible future catastrophes?", "interests": "Founded the Distributed AI Research Institute, which studies AI's social impacts.", "sourceIds": [ "gebru-response" ], "connections": [ { "pathway": "power", "relation": "discusses", "sourceIds": [ "gebru-response" ] } ], "history": [], "group": "research", "cardRole": "AI researcher", "aiScope": "Existing AI systems, their makers and the people affected.", "actionKind": "Accountability for present harms", "positionSourceIds": [ "gebru-response" ], "placementStatus": "different" }, { "id": "melanie-mitchell", "name": "Melanie Mitchell", "role": "AI & complexity researcher", "camp": null, "featured": false, "qualifier": "Questions the assumptions", "position": "First, be precise about what intelligence means.", "summary": "Mitchell examines recurring mistakes in how people interpret and predict AI capabilities, including assumptions about understanding and general intelligence.", "action": "Test capability claims and the assumptions behind predictions.", "nuance": "The cited paper is a research critique, not a policy pledge about a current pause.", "placement": "Unplaced: a technical critique alone is insufficient evidence of a preferred development pace.", "counterpoint": "How should we prepare for a severe outcome when its likelihood and the capabilities it requires are uncertain?", "interests": "Researches intelligence and complex systems. The cited 2021 paper provides historical conceptual context.", "sourceIds": [ "mitchell-paper" ], "connections": [], "history": [], "group": "research", "cardRole": "AI researcher", "aiScope": "Claims about general intelligence and future AI capability.", "actionKind": "Test the assumptions", "positionSourceIds": [ "mitchell-paper" ], "placementStatus": "different" }, { "id": "arvind-narayanan", "name": "Arvind Narayanan", "role": "Computer scientist · Princeton", "camp": null, "featured": false, "qualifier": "Questions the framing", "position": "AI's effects depend on institutions, not just intelligence.", "summary": "Narayanan and Kapoor argue for understanding AI through its adoption, human institutions and practical limitations, rather than assuming an autonomous superintelligence transformation.", "action": "Build institutional capacity and evaluate systems in their real settings.", "nuance": "Their argument challenges a model of technological change; it is not a claim that AI cannot cause harm.", "placement": "Unplaced: this framework does not establish a simple preference for acceleration or a pause.", "counterpoint": "Could capabilities change faster than the institutional patterns this analysis emphasises?", "interests": "A computer scientist and co-author of AI Snake Oil and AI as Normal Technology.", "sourceIds": [ "normal-technology" ], "connections": [ { "pathway": "control", "relation": "questions", "sourceIds": [ "normal-technology" ] }, { "pathway": "power", "relation": "discusses", "sourceIds": [ "normal-technology" ] } ], "history": [], "group": "research", "cardRole": "AI researcher", "aiScope": "AI deployed through human institutions.", "actionKind": "Evaluate real-world effects", "positionSourceIds": [ "normal-technology" ], "placementStatus": "different" }, { "id": "sundar-pichai", "name": "Sundar Pichai", "role": "Google & Alphabet CEO", "camp": "safeguards", "featured": false, "position": "Take serious risks seriously, while pursuing AI’s benefits.", "summary": "Pichai combines optimism about AI with a call for collective work to prevent serious harm.", "action": "Develop AI while investing in safety and collective coordination.", "nuance": "The interview’s numerical risk example comes from the interviewer; it is not Pichai’s estimate.", "placement": "Editorial reading of his call to manage risks alongside development.", "counterpoint": "What evidence would justify slowing deployment?", "interests": "Leads Google and Alphabet, which develop and commercialise frontier AI.", "sourceIds": [ "pichai-lex-2025", "pichai-summit-2026" ], "connections": [ { "pathway": "control", "relation": "discusses", "sourceIds": [ "pichai-lex-2025" ] } ], "history": [], "group": "leaders", "cardRole": "Technology company leader", "aiScope": "AI development and adoption, including powerful models.", "actionKind": "Develop with safeguards", "positionSourceIds": [ "pichai-lex-2025" ] }, { "id": "satya-nadella", "name": "Satya Nadella", "role": "Microsoft chairman & CEO", "camp": "safeguards", "featured": false, "position": "More capable AI must preserve human control.", "summary": "Nadella couples investment in superintelligence with human agency, security and governance.", "action": "Build control and governance into AI products.", "nuance": "A leadership commitment is not evidence that deployed systems satisfy it.", "placement": "Editorial interpretation of his signed development strategy.", "counterpoint": "Who can verify that people retain meaningful control?", "interests": "Leads Microsoft, a major AI developer and cloud provider.", "sourceIds": [ "nadella-copilot-2026" ], "connections": [ { "pathway": "control", "relation": "discusses", "sourceIds": [ "nadella-copilot-2026" ] } ], "history": [], "group": "leaders", "cardRole": "Technology company leader", "aiScope": "AI products and increasingly autonomous systems.", "actionKind": "Preserve human control", "positionSourceIds": [ "nadella-copilot-2026" ] }, { "id": "jensen-huang", "name": "Jensen Huang", "role": "NVIDIA founder & CEO", "camp": "advance", "featured": true, "position": "Keep advancing; target concrete harms with precise rules.", "summary": "Huang argues for rapid innovation and regulation focused on actual harms rather than hypothetical ones.", "action": "Adapt existing sector rules and make developers responsible for safety.", "nuance": "He supports safety responsibilities; this is not a claim that he opposes every regulation.", "placement": "His speech explicitly prioritises continued progress and narrowly targeted rules.", "counterpoint": "Can rules act early enough when a harm could be irreversible?", "interests": "NVIDIA supplies AI computing hardware and partners with model developers.", "sourceIds": [ "huang-g20-2026", "thinking-machines-nvidia-2026" ], "connections": [], "history": [], "group": "leaders", "cardRole": "Computing company leader", "aiScope": "AI development and specific harms in its applications.", "actionKind": "Target concrete harms", "positionSourceIds": [ "huang-g20-2026" ] }, { "id": "mark-zuckerberg", "name": "Mark Zuckerberg", "role": "Meta founder & CEO", "camp": "advance", "featured": false, "position": "Spread powerful AI widely to prevent concentrated control.", "summary": "Zuckerberg argues that distributed personal superintelligence can support freedom and a safer balance of power.", "action": "Expand access and pursue open AI development.", "nuance": "He also advocates safety boundaries; distribution does not establish safety by itself.", "placement": "Editorial reading of his development and distribution agenda.", "counterpoint": "How can broad access be reconciled with irreversible misuse?", "interests": "Leads Meta, which builds and distributes AI models and products.", "sourceIds": [ "zuckerberg-future-2026" ], "connections": [ { "pathway": "power", "relation": "discusses", "sourceIds": [ "zuckerberg-future-2026" ] } ], "history": [], "group": "leaders", "cardRole": "Technology company leader", "aiScope": "Powerful personal AI and access to its capabilities.", "actionKind": "Broaden access", "positionSourceIds": [ "zuckerberg-future-2026" ] }, { "id": "mustafa-suleyman", "name": "Mustafa Suleyman", "role": "Microsoft AI CEO", "camp": "safeguards", "featured": false, "position": "Build advanced AI within limits people control.", "summary": "Suleyman advocates powerful, purpose-specific systems that serve people and remain controllable.", "action": "Constrain autonomy and orient development around concrete human needs.", "nuance": "This states a design objective, not a demonstrated solution to control.", "placement": "His authored essay combines development with explicit limits.", "counterpoint": "How would those limits hold as capabilities increase?", "interests": "Leads Microsoft AI and its superintelligence team.", "sourceIds": [ "suleyman-humanist-2025", "nadella-copilot-2026" ], "connections": [ { "pathway": "control", "relation": "discusses", "sourceIds": [ "suleyman-humanist-2025" ] } ], "history": [], "group": "leaders", "cardRole": "AI company leader", "aiScope": "Advanced AI designed around human goals and limits.", "actionKind": "Limit autonomy", "positionSourceIds": [ "suleyman-humanist-2025" ] }, { "id": "lisa-su", "name": "Lisa Su", "role": "AMD chair & CEO", "camp": null, "featured": false, "position": "Expand access to computing, with clear security rules.", "summary": "Su supports AI adoption, interoperable technology and a consistent federal regulatory approach.", "action": "Promote open ecosystems and workable national-security restrictions.", "nuance": "Testimony on chips and regulation does not establish an extinction-risk position.", "placement": "Unplaced: the source addresses computing access more directly than frontier model pacing.", "counterpoint": "How should hardware suppliers share responsibility for downstream risks?", "interests": "Leads AMD, which sells processors and AI accelerators.", "sourceIds": [ "su-senate-2025", "su-amd-bio" ], "connections": [ { "pathway": "power", "relation": "discusses", "sourceIds": [ "su-senate-2025" ] } ], "history": [], "qualifier": "Risk position not established", "group": "leaders", "cardRole": "Computing company leader", "aiScope": "AI computing hardware, access and security restrictions.", "actionKind": "Development-speed view unclear", "positionSourceIds": [ "su-senate-2025" ], "placementStatus": "unclear" }, { "id": "andy-jassy", "name": "Andy Jassy", "role": "Amazon CEO", "camp": null, "featured": false, "position": "Invest heavily in AI’s economic potential.", "summary": "Jassy argues that AI is a major technological shift warranting substantial investment.", "action": "Expand AI infrastructure and applications.", "nuance": "Commercial optimism is not evidence of a view on extinction or a frontier pause.", "placement": "Unplaced: this shareholder letter does not resolve the relevant risk-policy question.", "counterpoint": "What safety threshold would override the investment case?", "interests": "Leads Amazon, including its AWS cloud and AI businesses.", "sourceIds": [ "jassy-letter-2025" ], "connections": [], "history": [], "qualifier": "Risk position not established", "group": "leaders", "cardRole": "Technology company leader", "aiScope": "AI infrastructure and commercial applications.", "actionKind": "Risk policy unclear", "positionSourceIds": [ "jassy-letter-2025" ], "placementStatus": "unclear" }, { "id": "mira-murati", "name": "Mira Murati", "role": "Thinking Machines Lab co-founder & CEO", "camp": null, "featured": false, "position": "A human-centred lab; personal risk position still unclear.", "summary": "Her lab’s published mission emphasises human judgment and participation in AI development.", "action": "Institutional approach: build customisable AI around human participation.", "nuance": "The mission is a company statement, not a personally signed Murati position.", "placement": "Unplaced until a sufficiently specific personal policy statement is sourced.", "counterpoint": "What limits would the lab impose on increasingly autonomous systems?", "interests": "Leads Thinking Machines Lab; its NVIDIA partnership expands access to compute.", "sourceIds": [ "thinking-machines-human-2026", "thinking-machines-nvidia-2026" ], "connections": [], "history": [], "qualifier": "Institutional view only", "group": "leaders", "cardRole": "AI company leader", "aiScope": "Customisable AI, as described by her lab.", "actionKind": "Personal position unclear", "positionSourceIds": [ "thinking-machines-human-2026" ], "placementStatus": "unclear" }, { "id": "arthur-mensch", "name": "Arthur Mensch", "role": "Mistral AI co-founder & CEO", "camp": "advance", "featured": false, "position": "Build European AI capacity to preserve autonomy.", "summary": "Mensch argues that dependence on foreign AI providers threatens Europe’s control over its future.", "action": "Accelerate domestic AI capability, infrastructure and adoption.", "nuance": "This is a sovereignty argument, not an extinction forecast.", "placement": "His CEO foreword explicitly calls for faster capacity building.", "counterpoint": "Can a competitive AI race coexist with shared safety limits?", "interests": "Leads Mistral AI, which benefits from demand for European AI suppliers.", "sourceIds": [ "mensch-europe-playbook" ], "connections": [ { "pathway": "power", "relation": "discusses", "sourceIds": [ "mensch-europe-playbook" ] } ], "history": [], "qualifier": "Sovereignty-led argument", "group": "leaders", "cardRole": "AI company leader", "aiScope": "European AI capability, infrastructure and adoption.", "actionKind": "Build European capacity", "positionSourceIds": [ "mensch-europe-playbook" ] }, { "id": "aidan-gomez", "name": "Aidan Gomez", "role": "Cohere co-founder & CEO", "camp": "advance", "featured": false, "position": "Build responsibly; base the rules on evidence.", "summary": "Gomez challenges speculative extinction claims and lab-led rulemaking, while supporting enforceable safeguards.", "action": "Use independent, proportionate testing and incident reporting.", "nuance": "He recognises serious misuse and infrastructure risks; he does not dismiss all AI danger.", "placement": "Editorial reading of his objection to incumbent-led pacing and preference for continued responsible development.", "counterpoint": "How should standards handle uncertain but potentially catastrophic risks?", "interests": "Leads Cohere, a competitor to larger AI labs with an interest in open market access.", "sourceIds": [ "gomez-rules-2026" ], "connections": [ { "pathway": "control", "relation": "questions", "sourceIds": [ "gomez-rules-2026" ] }, { "pathway": "infrastructure", "relation": "discusses", "sourceIds": [ "gomez-rules-2026" ] }, { "pathway": "power", "relation": "discusses", "sourceIds": [ "gomez-rules-2026" ] } ], "history": [], "group": "leaders", "cardRole": "AI company leader", "aiScope": "AI development and evidence-based evaluation.", "actionKind": "Proportionate safeguards", "positionSourceIds": [ "gomez-rules-2026" ] }, { "id": "ilya-sutskever", "name": "Ilya Sutskever", "role": "Safe Superintelligence co-founder & CEO", "camp": "safeguards", "featured": false, "position": "Advance capability only with safety staying ahead.", "summary": "Sutskever affirms building safe superintelligence; SSI’s mission couples capability progress with prior safety.", "action": "Focus a dedicated research lab on safe superintelligence.", "nuance": "The detailed approach is a lab mission, not evidence that safety is solved.", "placement": "Conditional placement based on his signed message and the lab’s stated mission.", "counterpoint": "How will outsiders assess whether safety really stays ahead?", "interests": "Leads a funded company developing superintelligence.", "sourceIds": [ "sutskever-ceo-2025", "ssi-mission" ], "connections": [ { "pathway": "control", "relation": "discusses", "sourceIds": [ "ssi-mission" ] } ], "history": [], "qualifier": "Conditional on safety", "group": "leaders", "cardRole": "AI researcher & company leader", "aiScope": "Future AI more capable than humans across many tasks.", "actionKind": "Safety before capability", "positionSourceIds": [ "sutskever-ceo-2025", "ssi-mission" ] }, { "id": "fei-fei-li", "name": "Fei-Fei Li", "role": "Stanford AI researcher & HAI founding director", "camp": "safeguards", "featured": true, "position": "AI acting in the world needs real-world validation.", "summary": "Li warns that impressive simulations do not establish reliable physical-world behaviour. She also highlights barriers to access.", "action": "Require rigorous field testing and improve evaluation.", "nuance": "These claims concern world models and deployment; they are not an extinction estimate.", "placement": "Editorial reading of her support for development with substantive validation.", "counterpoint": "What evidence is enough for a safety-critical deployment?", "interests": "Stanford researcher and World Labs co-founder. World Labs commercialises the world-model technology discussed in her policy interview.", "sourceIds": [ "li-world-models-2026", "li-world-labs" ], "connections": [ { "pathway": "infrastructure", "relation": "discusses", "sourceIds": [ "li-world-models-2026" ] }, { "pathway": "power", "relation": "discusses", "sourceIds": [ "li-world-models-2026" ] } ], "history": [], "group": "leaders", "cardRole": "AI researcher & company founder", "aiScope": "AI that models and acts in the physical world.", "actionKind": "Real-world validation", "positionSourceIds": [ "li-world-models-2026" ] }, { "id": "dawn-song", "name": "Dawn Song", "role": "UC Berkeley computer scientist", "camp": "safeguards", "featured": true, "position": "Build security and oversight alongside autonomous capability.", "summary": "Song warns that agents can amplify errors and manipulation because they take actions in the world.", "action": "Invest in rigorous evaluation, secure design and human oversight.", "nuance": "Agent-security concerns do not establish a probability of human extinction.", "placement": "Her interview explicitly calls for safeguards alongside development.", "counterpoint": "Can oversight remain effective over long autonomous tasks?", "interests": "University researcher working on AI security; co-director of Berkeley RDI.", "sourceIds": [ "song-agents-2026" ], "connections": [ { "pathway": "control", "relation": "discusses", "sourceIds": [ "song-agents-2026" ] }, { "pathway": "infrastructure", "relation": "discusses", "sourceIds": [ "song-agents-2026" ] } ], "history": [], "group": "research", "cardRole": "AI security researcher", "aiScope": "AI agents that carry out tasks with some independence.", "actionKind": "Secure autonomous systems", "positionSourceIds": [ "song-agents-2026" ] }, { "id": "anca-dragan", "name": "Anca Dragan", "role": "AI safety researcher, Google DeepMind", "camp": "safeguards", "featured": false, "position": "Test whether powerful systems remain under human direction.", "summary": "A framework Dragan co-authored addresses misalignment and models interfering with correction or shutdown.", "action": "Use capability thresholds, risk assessments and safety case reviews.", "nuance": "This is a co-authored institutional framework, not proof of effective control.", "placement": "Conditional development under a published safety framework.", "counterpoint": "Who independently judges the adequacy of a safety case?", "interests": "Works at Google DeepMind, a frontier AI developer.", "sourceIds": [ "dragan-framework-2026" ], "connections": [ { "pathway": "control", "relation": "discusses", "sourceIds": [ "dragan-framework-2026" ] }, { "pathway": "power", "relation": "discusses", "sourceIds": [ "dragan-framework-2026" ] } ], "history": [], "group": "research", "cardRole": "AI safety researcher", "aiScope": "The most capable models, especially risks of losing control.", "actionKind": "Safety reviews at capability thresholds", "positionSourceIds": [ "dragan-framework-2026" ] }, { "id": "helen-toner", "name": "Helen Toner", "role": "CSET executive director, Georgetown", "camp": null, "featured": false, "position": "Capability is improving faster than reliable control.", "summary": "Toner warns that techniques for making AI more capable are outpacing techniques for making it reliably follow constraints.", "action": "Prioritise closing the gap between capability and control.", "nuance": "The published excerpt does not say what kind of slowdown she would favour.", "placement": "Not placed yet. The sources here show concern, but not a full policy argument, and concern alone does not decide a camp.", "counterpoint": "What measurable threshold would demonstrate adequate control?", "interests": "Leads an AI policy research centre; institutional affiliation is disclosed in the cited interview excerpt.", "sourceIds": [ "toner-control-2026" ], "connections": [ { "pathway": "control", "relation": "discusses", "sourceIds": [ "toner-control-2026" ] } ], "history": [], "qualifier": "Policy detail unclear", "group": "research", "cardRole": "AI policy researcher", "aiScope": "Increasingly capable AI and the ability to control it.", "actionKind": "Policy detail unclear", "positionSourceIds": [ "toner-control-2026" ], "placementStatus": "unclear" }, { "id": "meredith-whittaker", "name": "Meredith Whittaker", "role": "Signal president & AI Now co-founder", "camp": null, "featured": false, "position": "Ask who controls AI—and whom it gives power over.", "summary": "Whittaker argues that releasing open models does not by itself disperse the power held by companies controlling data, compute and distribution.", "action": "Assess openness concretely and build privacy-preserving, mission-specific technology.", "nuance": "She accepts useful, focused AI applications; her critique is not a blanket rejection of AI.", "placement": "Unplaced: this critique cuts across the development-speed camps.", "counterpoint": "Which institutional reforms would redistribute power effectively?", "interests": "Leads Signal and co-founded AI Now; has a stated privacy and accountability advocacy perspective.", "sourceIds": [ "whittaker-open-source-2026", "whittaker-ai-now-2023" ], "connections": [ { "pathway": "power", "relation": "discusses", "sourceIds": [ "whittaker-open-source-2026" ] } ], "history": [ { "date": "2023-07-19", "text": "Calls for democratic control over AI infrastructure.", "sourceIds": [ "whittaker-ai-now-2023" ] }, { "date": "2026-02", "text": "Challenges claims that open models alone decentralise power.", "sourceIds": [ "whittaker-open-source-2026" ] } ], "qualifier": "Questions the framing", "group": "research", "cardRole": "Technology & privacy advocate", "aiScope": "AI infrastructure, access, surveillance and privacy.", "actionKind": "Challenge concentrated power", "positionSourceIds": [ "whittaker-open-source-2026", "whittaker-ai-now-2023" ], "placementStatus": "different" }, { "id": "yejin-choi", "name": "Yejin Choi", "role": "Stanford AI researcher", "camp": null, "featured": false, "position": "Better AI needs more than bigger models.", "summary": "Choi studies common sense and argues for a more varied, accessible AI ecosystem with room for different human values.", "action": "Develop smaller models, open research resources and pluralistic alignment.", "nuance": "Criticism of scaling assumptions is not a claim that catastrophic risk is impossible.", "placement": "Unplaced: her research agenda does not specify a general frontier pacing policy.", "counterpoint": "How can broader participation coexist with safeguards against misuse?", "interests": "Stanford professor; Stanford’s 2025 appointment announcement also disclosed an NVIDIA affiliation.", "sourceIds": [ "choi-open-minds-2026", "choi-stanford-2025" ], "connections": [ { "pathway": "power", "relation": "discusses", "sourceIds": [ "choi-open-minds-2026" ] } ], "history": [], "qualifier": "Alternative research direction", "group": "research", "cardRole": "AI researcher", "aiScope": "Language models, common sense and access to research.", "actionKind": "Alternative research direction", "positionSourceIds": [ "choi-open-minds-2026" ], "placementStatus": "different" }, { "id": "francesca-rossi", "name": "Francesca Rossi", "role": "IBM responsible AI & governance leader", "camp": null, "featured": false, "position": "People must remain accountable for what AI does.", "summary": "Rossi argues that human responsibility cannot be delegated to software and that pre-release review cannot guarantee zero misuse.", "action": "Retain human judgment, verify outputs and govern deployment.", "nuance": "This practical governance view does not settle her preferred pace of frontier research.", "placement": "Unplaced to avoid treating responsible-use advice as a full existential-risk policy.", "counterpoint": "How can accountability work when no person can meaningfully supervise the system?", "interests": "IBM Fellow and its global responsible AI and governance leader.", "sourceIds": [ "rossi-accountability-2026" ], "connections": [ { "pathway": "control", "relation": "discusses", "sourceIds": [ "rossi-accountability-2026" ] }, { "pathway": "power", "relation": "discusses", "sourceIds": [ "rossi-accountability-2026" ] } ], "history": [], "qualifier": "Human accountability", "group": "research", "cardRole": "AI ethics researcher", "aiScope": "AI use in decisions for which people remain responsible.", "actionKind": "Human accountability", "positionSourceIds": [ "rossi-accountability-2026" ], "placementStatus": "different" }, { "id": "inioluwa-deborah-raji", "name": "Inioluwa Deborah Raji", "role": "AI auditing researcher", "camp": null, "featured": false, "position": "Audit what systems actually do, with consequences.", "summary": "Raji’s research connects evaluations of deployed systems to institutional and legal accountability.", "action": "Combine audits with incident reporting, workflow simulations and pilot studies.", "nuance": "This is evaluation research, not an extinction forecast or a pause endorsement.", "placement": "Unplaced: provides evidence and safeguards that can inform multiple policy positions.", "counterpoint": "What gives auditors enough access and independence to find serious failures?", "interests": "The cited work is her UC Berkeley doctoral research.", "sourceIds": [ "raji-auditing-2026" ], "connections": [], "history": [], "qualifier": "Evidence and accountability", "group": "research", "cardRole": "AI auditing researcher", "aiScope": "Systems evaluated in the settings where they are used.", "actionKind": "Audits and accountability", "positionSourceIds": [ "raji-auditing-2026" ], "placementStatus": "different" }, { "id": "katja-grace", "name": "Katja Grace", "role": "AI Impacts co-founder & lead researcher", "camp": "pause", "featured": false, "position": "Build the ability to pause before danger becomes obvious.", "summary": "Grace argues that waiting until the last safe moment is unreliable because collective action takes time.", "action": "Pursue an AI pause as soon as possible.", "nuance": "Her argument is precautionary; its assumptions about danger and coordination remain contestable.", "placement": "Her June 2026 essay explicitly argues for pausing promptly.", "counterpoint": "How would a pause be enforced, and what would justify restarting?", "interests": "Leads research at AI Impacts, based at MIRI. Her pause essay is personal advocacy, not a consensus finding.", "sourceIds": [ "grace-pause-2026", "grace-ai-impacts" ], "connections": [ { "pathway": "control", "relation": "supports", "sourceIds": [ "grace-pause-2026" ] } ], "history": [], "group": "research", "cardRole": "AI risk researcher", "aiScope": "Advanced AI development addressed by her pause proposal.", "actionKind": "Precautionary pause", "positionSourceIds": [ "grace-pause-2026" ], "qualifier": "Precautionary pause" }, { "id": "kate-crawford", "name": "Kate Crawford", "role": "AI researcher & author of Atlas of AI", "camp": null, "featured": false, "position": "Examine the material and political costs of AI.", "summary": "Crawford traces how AI’s resources, labour and institutions can entrench inequality and concentrated power.", "action": "Scrutinise the institutions and resource choices behind AI systems.", "nuance": "Her 2021 book’s argument is not a current extinction forecast.", "placement": "Unplaced: political and material harms do not map neatly onto a single development-speed policy.", "counterpoint": "Which interventions address these harms without concentrating power further?", "interests": "USC research professor and Microsoft Research senior principal researcher, according to her biography.", "sourceIds": [ "crawford-atlas", "crawford-bio" ], "connections": [ { "pathway": "power", "relation": "discusses", "sourceIds": [ "crawford-atlas" ] } ], "history": [], "qualifier": "Power and present harms", "group": "research", "cardRole": "AI researcher & author", "aiScope": "Existing AI, its material resources and political power.", "actionKind": "Power and present harms", "positionSourceIds": [ "crawford-atlas" ], "placementStatus": "different" }, { "id": "david-sacks", "name": "David Sacks", "role": "Investor & policy adviser · White House AI adviser in the cited 2025 statement", "cardRole": "Investor & policy adviser", "group": "policy", "camp": "advance", "featured": false, "position": "Build AI capacity to secure national competitiveness.", "summary": "Sacks presents AI leadership as an economic and security priority. His July 2025 statement advocates investment in innovation, infrastructure and international partnerships.", "action": "Expand AI innovation, infrastructure and international partnerships.", "aiScope": "National AI capacity, infrastructure and adoption.", "actionKind": "National competitiveness", "nuance": "He also calls for protecting workers and avoiding authoritarian uses. This statement does not establish a personal extinction-risk estimate.", "placement": "His emphasis on building capacity places this dated statement in Keep advancing. The label is editorial.", "qualifier": "National competitiveness", "counterpoint": "How should national competition be balanced against shared safety requirements?", "interests": "An investor speaking in a government advisory role when this source was published.", "sourceIds": [ "sacks-action-plan-2025" ], "positionSourceIds": [ "sacks-action-plan-2025" ], "connections": [], "history": [ { "date": "2025-07-23", "text": "Advocates national AI leadership, with worker and freedom protections.", "sourceIds": [ "sacks-action-plan-2025" ] } ] }, { "id": "marc-andreessen", "name": "Marc Andreessen", "role": "Technology investor · Andreessen Horowitz co-founder", "cardRole": "Technology investor", "group": "leaders", "camp": "advance", "featured": false, "position": "AI’s benefits justify rapid development and broad access.", "summary": "Andreessen rejects the argument that AI will choose to destroy humanity. His 2023 essay favours rapid development, competition and open access while recognising misuse.", "action": "Develop AI rapidly, keep competition open and tackle harmful uses.", "aiScope": "AI development broadly, including powerful and open models.", "actionKind": "Rapid development", "nuance": "He favours action against concrete misuse. This is a historical essay, not a newly verified statement of his latest position.", "placement": "His explicit case for rapid development places the cited essay in Keep advancing.", "qualifier": "Historical statement", "counterpoint": "Does rejecting human-like motives settle the risks of autonomous systems?", "interests": "Co-founder of an investment firm with a commercial interest in technology companies.", "sourceIds": [ "andreessen-save-world-2023" ], "positionSourceIds": [ "andreessen-save-world-2023" ], "connections": [ { "pathway": "control", "relation": "questions", "sourceIds": [ "andreessen-save-world-2023" ] }, { "pathway": "power", "relation": "discusses", "sourceIds": [ "andreessen-save-world-2023" ] } ], "history": [ { "date": "2023-06-06", "text": "Argues for rapid development, competition and action against misuse.", "sourceIds": [ "andreessen-save-world-2023" ] } ] }, { "id": "yuval-noah-harari", "name": "Yuval Noah Harari", "role": "Historian & author of Sapiens and Nexus", "cardRole": "Historian & author", "group": "commentary", "camp": null, "placementStatus": "different", "featured": false, "position": "Protect human agency as AI gains the ability to act.", "summary": "Harari warns that AI agents could reshape financial and social power. His July 2026 interview calls for identifying AI interactions, protecting children and withholding legal personhood from AI.", "action": "Identify AI interactions, protect children and deny AI legal personhood for now.", "aiScope": "AI agents, financial systems and relationships with people.", "actionKind": "Limits on specific uses", "nuance": "These are restrictions on uses and legal status, not a blanket training pause. His authored answers describe dystopian outcomes as possibilities, not prophecies.", "placement": "Different framing: the cited proposals concern human agency and legal status, rather than one development-speed policy.", "qualifier": "Human agency and legal status", "counterpoint": "Which limits would protect people without blocking beneficial applications?", "interests": "A historian and public author offering social and political analysis.", "sourceIds": [ "harari-interview-2026", "harari-faqs" ], "positionSourceIds": [ "harari-interview-2026", "harari-faqs" ], "connections": [ { "pathway": "power", "relation": "discusses", "sourceIds": [ "harari-interview-2026", "harari-faqs" ] }, { "pathway": "control", "relation": "discusses", "sourceIds": [ "harari-interview-2026" ] } ], "history": [ { "date": "2026-07-25", "text": "Proposes limits on AI relationships and legal personhood in a published interview.", "sourceIds": [ "harari-interview-2026" ] } ] }, { "id": "tristan-harris", "name": "Tristan Harris", "role": "Center for Humane Technology co-founder & podcast host", "cardRole": "Technology advocate & podcast host", "group": "commentary", "camp": "pause", "featured": false, "position": "Set enforceable limits before AI can improve itself without oversight.", "summary": "Harris advocates shared safety research and independent scrutiny. In September 2026 he calls for pre-approval before labs attempt AI development loops without human oversight.", "action": "Require approval before unchecked AI self-improvement, alongside safety review and liability.", "aiScope": "Powerful AI agents and systems that could build their own successors.", "actionKind": "Conditional restrictions", "nuance": "This is a restriction on specific frontier work. His accounts of incidents are commentary, not independent findings or a measured probability of catastrophe.", "placement": "His call to withhold permission for unchecked self-improvement until it is safe places him in Slow or pause.", "qualifier": "Conditional restrictions", "counterpoint": "Who could assess safety independently and enforce those boundaries?", "interests": "A technology-reform advocate and podcast host. His April discussion also supports changing laws and incentives.", "sourceIds": [ "harris-slowdown-2026", "harris-roadmap-2026" ], "positionSourceIds": [ "harris-slowdown-2026", "harris-roadmap-2026" ], "connections": [ { "pathway": "control", "relation": "supports", "sourceIds": [ "harris-slowdown-2026" ] }, { "pathway": "power", "relation": "discusses", "sourceIds": [ "harris-roadmap-2026" ] } ], "history": [ { "date": "2026-04-02", "text": "Advocates laws and incentives that protect people.", "sourceIds": [ "harris-roadmap-2026" ] }, { "date": "2026-09-17", "text": "Calls for approval before unchecked self-improvement and stronger safety scrutiny.", "sourceIds": [ "harris-slowdown-2026" ] } ] } ], "pathways": [ { "id": "control", "title": "Loss of control", "short": "A powerful system could pursue goals people can no longer change.", "outcome": "Potential existential risk", "steps": [ "Capable systems act independently", "Oversight or control fails", "Human control could be lost" ], "evidence": "The February 2026 safety report describes early warning behaviours in laboratory tests. It does not establish an irreversible real-world takeover. Researchers disagree about how far those findings generalise.", "uncertainty": "What capabilities, access and incentives would be sufficient? Could reliable safeguards prevent the chain?", "safeguards": "Limit autonomy and access, evaluate dangerous capabilities, and research reliable oversight.", "sourceIds": [ "safety-report", "hinton-nobel", "lecun-interview" ] }, { "id": "biology", "title": "Biological misuse", "short": "AI could make it easier for people to cause biological harm.", "outcome": "Potential global catastrophe", "steps": [ "AI assists a malicious actor", "Barriers to biological misuse fall", "A severe outbreak could spread" ], "evidence": "Capability tests have raised concern that AI could assist biological misuse. The February 2026 report stresses uncertainty about real-world effects, with physical access and practical expertise still important barriers.", "uncertainty": "Real-world harm depends on physical access, expertise and many steps beyond receiving information.", "safeguards": "Evaluate misuse capabilities, apply access safeguards and strengthen public-health resilience.", "sourceIds": [ "safety-report", "gates-2026", "amodei-2026" ] }, { "id": "conflict", "title": "Military escalation", "short": "Faster decisions and misplaced trust could amplify conflict.", "outcome": "Potential global catastrophe", "steps": [ "AI informs military decisions", "Error or escalation compounds", "Conflict could become catastrophic" ], "evidence": "An exploratory 2024 study found escalatory behaviour in simulated wargames with five language models. This is evidence about those models in that setting, not a forecast of real military decisions.", "uncertainty": "The link to nuclear catastrophe depends on deployment choices and human decision-making; it is not an inevitable consequence of military AI.", "safeguards": "Preserve accountable human control and clear limits on consequential decisions.", "sourceIds": [ "tegmark-fli", "military-wargames-2024" ] }, { "id": "infrastructure", "title": "Infrastructure failures", "short": "Failures could spread through the services societies depend on.", "outcome": "Severe societal harm", "steps": [ "Essential services rely on AI", "Attacks or failures spread", "Vital services could be disrupted" ], "evidence": "Cyber misuse and dependence on shared systems create routes to wider disruption.", "uncertainty": "Infrastructure disruption is not itself proof of an extinction pathway. Scale and resilience matter.", "safeguards": "Reduce shared points of failure, preserve recovery options and strengthen defensive security.", "sourceIds": [ "safety-report", "gates-2026" ] }, { "id": "power", "title": "Human disempowerment", "short": "Power could become concentrated and difficult to challenge.", "outcome": "Potential permanent disempowerment", "steps": [ "AI concentrates capability", "Accountability weakens", "People could lose lasting control" ], "evidence": "Researchers and commentators discuss surveillance, concentrated influence and the distribution of AI's benefits.", "uncertainty": "Persistent or irreversible disempowerment is a stronger claim than present inequality, and needs its own supporting argument.", "safeguards": "Build accountability, public oversight, meaningful participation and broad access to benefits.", "sourceIds": [ "amodei-2026", "normal-technology", "lecun-interview", "whittaker-open-source-2026" ] } ], "revisions": [ { "date": "2026-10-01", "summary": "Will approved David Sacks, Marc Andreessen, Yuval Noah Harari and Tristan Harris. Added six checked primary sources, explicit AI scope and roles, distinct reasons for being outside the groups, and clearer action and source-date labels. Existing historical placements retain their evidence and qualifications." }, { "date": "2026-09-17", "summary": "Renamed AI Risk Atlas; expanded editorial comparisons from three to nine. Preserved sourced positions and qualifications." }, { "date": "2026-09-17", "summary": "Will approved 21 additional voices and Dario Amodei’s qualified move to coordinated slowing. Expanded interactive infographic preview. Original source-review date retained." }, { "date": "2026-09-15", "version": "0.1.0-prototype", "summary": "First HTML prototype. Dated source-based positions; not a comprehensive review of every thinker's latest statements. Military escalation is a conceptual preview awaiting fuller research." }, { "date": "2026-09-16", "version": "0.2.0-design-preview", "summary": "Public design-review preview authorised by Will. Clarified policy-based camps, dated paraphrases, editorial questions and conditional pathway language. No claim of a new comprehensive source review." } ], "rosterUpdatedAt": "2026-10-01", "name": "AI Risk Atlas", "comparisons": [ { "id": "open-ai-power", "a": "mark-zuckerberg", "b": "meredith-whittaker", "question": "Does open AI distribute power?", "note": "Compare the case for wider access with the argument that infrastructure can still concentrate control.", "sourceIds": [ "zuckerberg-future-2026", "whittaker-open-source-2026", "whittaker-ai-now-2023" ] }, { "id": "intelligence-control", "a": "geoffrey-hinton", "b": "yann-lecun", "question": "Can we keep control of smarter AI?", "note": "Compare a warning about losing control with a different account of how intelligence and goals relate.", "sourceIds": [ "hinton-nobel", "superintelligence", "fli-signatories", "lecun-interview", "ng-pause" ] }, { "id": "pace-and-rules", "a": "dario-amodei", "b": "aidan-gomez", "question": "Who should set the pace?", "note": "Compare coordinated slowing with objections to letting the largest labs define the rules.", "sourceIds": [ "amodei-2026", "amodei-pace-2026", "amodei-bio", "gomez-rules-2026" ] }, { "id": "cost-of-pausing", "a": "andrew-ng", "b": "katja-grace", "question": "What is the cost of waiting?", "note": "Weigh the risks of delaying useful AI against the case for pausing before dangerous capabilities become obvious.", "sourceIds": [ "ng-pause", "grace-pause-2026", "grace-ai-impacts" ] }, { "id": "deploy-or-demonstrate", "a": "sam-altman", "b": "yoshua-bengio", "question": "Learn by deploying, or establish safety first?", "note": "Compare learning through real-world deployment with requiring scientific safety confidence before superintelligence.", "sourceIds": [ "altman-2025", "ceo-safety-2026", "superintelligence", "fli-signatories" ] }, { "id": "who-tests-safety", "a": "demis-hassabis", "b": "inioluwa-deborah-raji", "question": "Who gets to say an AI is safe?", "note": "Compare capability evaluations during development with audits, incident reporting and testing in real workflows.", "sourceIds": [ "hassabis-2025", "ceo-safety-2026", "raji-auditing-2026" ] }, { "id": "evidence-and-forecasts", "a": "geoffrey-hinton", "b": "melanie-mitchell", "question": "What would count as evidence of danger?", "note": "Compare a warning about future control with scrutiny of what intelligence claims and capability tests establish.", "sourceIds": [ "hinton-nobel", "superintelligence", "fli-signatories", "mitchell-paper" ] }, { "id": "risk-priorities", "a": "sundar-pichai", "b": "timnit-gebru", "question": "Which harms should shape the agenda?", "note": "Compare pursuing AI’s benefits while coordinating on serious risks with prioritising present harms and accountability.", "sourceIds": [ "pichai-lex-2025", "pichai-summit-2026", "gebru-response" ] }, { "id": "human-control", "a": "satya-nadella", "b": "francesca-rossi", "question": "What does human control really require?", "note": "Explore two approaches to human control: governance built into products, and people remaining responsible for decisions.", "sourceIds": [ "nadella-copilot-2026", "rossi-accountability-2026" ] } ] }