Promises are easy to make. Each entry says what was asked for, who asked, and what has happened since.
United Nations General Assembly · September 2024
The first global agreement, signed by every UN member state, creating a standing international process for AI governance.
White House Office of Science and Technology Policy · October 2022
Americas first national statement of five rights every person should have in the age of AI, never enacted into law.
Beijing Academy of Artificial Intelligence · May 2019
Chinas first statement of AI ethics principles, later followed by some of the worlds first binding AI regulations.
Access Now and Amnesty International · May 2018
One of the first major documents to frame AI harms as human-rights violations rather than technical glitches.
Current and former employees of frontier AI labs · June 2024
The letter that broke the silence inside the labs: employees demanding the right to warn the public without losing their livelihoods.
Pontifical Academy for Life, with Microsoft, IBM and others · February 2020
The Vatican-brokered pledge putting AI ethics in explicitly moral and spiritual terms.
Anthropic · 26 February 2026
Anthropic kept its rule against using its AI for autonomous weapons or mass surveillance even when the Pentagon demanded otherwise and blacklisted it. A federal judge later ruled the blacklisting was illegal retaliation.
Government of India (summit host) · 21 February 2026
The 2026 summit in India produced a declaration on AI for development signed by 88 countries and bodies, plus large investment pledges. It continues the shift of the summit series from safety toward growth and access.
New York State Legislature; signed by Governor Kathy Hochul · 2025-12-19 (signed); 2026-03-27 (chapter amendment); 2027-01-01 (effective)
New York's frontier-AI law requires big developers to publish safety frameworks and report critical incidents within 72 hours, starting in 2027. Whether it survives federal preemption efforts is an open question.
Landgericht München I · 11 November 2025
A German court ruled that when ChatGPT memorises and reproduces song lyrics, OpenAI infringes copyright, and ordered it to stop and pay. It is the first major European win for rights holders against an AI developer, and is under appeal.
Future of Life Institute · 22 October 2025
A 2025 statement, signed by AI pioneers and public figures across the political spectrum, calling for a ban on building superintelligence until it is proven safe and the public agrees. No country has acted on it, and labs continue the race.
California Legislature; signed by Governor Gavin Newsom · 2025-09-29 (signed); 2026-01-01 (effective)
California now requires the largest AI developers to publish their safety plans and report serious incidents, and protects employees who raise concerns. The federal government is trying to override it.
CeSIA (French Center for AI Safety), The Future Society, and UC Berkeley CHAI · 22 September 2025
Nobel laureates and ex-leaders asked governments to agree by the end of 2026 on things AI must never do, such as launching nuclear weapons or mass surveillance. With four months left, no negotiation toward such a treaty has begun.
UN General Assembly · 26 August 2025
The UN created its first scientific panel on AI and an annual forum for governments to discuss AI rules. Both are now up and running, but they advise rather than regulate.
European Commission AI Office (drafted by independent experts); voluntary signatures by model providers · 2025-07-10 (final); obligations applied 2025-08-02
A voluntary rulebook that shows how big AI model makers can meet the EU AI Act's requirements on safety, copyright and transparency. Most major labs signed; Meta refused and xAI signed only part of it.
France and India (summit co-chairs) · 11 February 2025
The 2025 Paris summit statement emphasised inclusive, sustainable and open AI rather than safety. The US and UK refused to sign, and the summit process has since focused more on investment than on risk.
Independent expert panel chaired by Yoshua Bengio, mandated by the Bletchley summit; secretariat hosted by the UK AISI · 2025-01-29 (first full report); 2026-02-03 (second edition)
An IPCC-style scientific assessment of AI capabilities and risks, produced annually by an international expert panel. It has become the reference document for summits, though the US stopped endorsing it in 2026.
President Donald J. Trump · 2025-01-23 (EO 14179); 2025-07-23 (AI Action Plan); 2025-12-11 (preemption EO)
The 2025 US policy reset: cancel the Biden safety order, publish a growth-first action plan, and try to override state AI laws. It replaced mandatory reporting with voluntary testing agreements while attacking laws like California's and New York's.
Ed Newton-Rex / Fairly Trained (individual creators and creative organisations) · 22 October 2024
A one-sentence statement signed by tens of thousands of artists, authors and actors saying that training AI on their work without a licence is unjust. It helped block a UK copyright carve-out, but licensing remains voluntary and courts are split.
US District Court, N.D. California (Judge William Alsup; settlement approved by Judge Araceli Martínez-Olguín) · 2024-08-19 (filed); 2025-06-23 (fair-use ruling); 2026-07-22 (final approval)
Authors sued Anthropic over books used to train Claude. A judge said training on bought books is legal but piracy is not, and Anthropic agreed to the largest copyright payout in history, about $3,000 per book.
European Parliament and Council · 2024-08-01 (in force); amended 2026-07-27
The world's first comprehensive AI law, banning some uses and requiring safety and transparency for risky ones. Bans and rules for large models are live, but in 2026 the EU delayed its high-risk rules by 16-24 months under industry pressure.
UK and Republic of Korea governments; 16 companies signed the Frontier AI Safety Commitments · 21 May 2024
At the 2024 Seoul summit sixteen AI companies promised to publish safety plans with clear red lines within nine months. Independent tracking found only about a third fully delivered on time.
Google DeepMind · 2024-05-17 (v1); 2025-09-22 (v3); 2026-04-17 (v3.1)
Google DeepMind's protocol for spotting when a model crosses dangerous capability thresholds. It has been updated repeatedly and model reports are published, but it sits alongside Google's 2025 decision to drop its no-weapons pledge.
Council of Europe · 2024-05-17 (adopted); 2024-09-05 (opened for signature)
The first legally binding international treaty on AI, focused on human rights and democracy. The EU ratified it in 2026; the US signed but has not ratified, and the treaty mostly lets each country decide how to apply it to private companies.
US District Court, S.D. New York (Judge Sidney Stein), consolidated MDL with Daily News, CIR, and authors' cases · 2023-12-27 (filed)
The most-watched news-publisher lawsuit against OpenAI and Microsoft, filed in 2023, is still grinding through discovery. Meanwhile OpenAI has paid other publishers for licences rather than settle with the Times.
OpenAI · 2023-12-18 (v1 beta); 2025-04-15 (v2)
OpenAI's internal system for measuring dangerous capabilities before release. Its 2025 update gave the company explicit room to relax safeguards if competitors do, which critics read as a loophole.
UK AI Safety Summit (28 countries + EU) · 1 November 2023
The 2023 UK summit statement in which 28 governments, including the US and China, agreed that frontier AI poses serious risks worth cooperating on. It produced a standing scientific report and national safety institutes, but the follow-up summits moved away from safety toward 'action' and 'impact'.
President Joseph R. Biden · 30 October 2023
The Biden administration's sweeping 2023 order on AI safety, civil rights, and reporting by frontier developers. It was cancelled on the first day of the Trump administration, though some institutions it created live on in altered form.
G7 (Japan presidency) · 30 October 2023
The G7's code of conduct for companies building advanced AI, backed by a public reporting system run by the OECD. Around twenty firms now file reports, but reporting is voluntary and self-graded.
Anthropic · 2023-09-19 (v1.0); 2024-10-15 (v2.0); 2026-02-24 (v3.0)
Anthropic's rulebook for when it will add safeguards as its models get more capable. The 2026 rewrite replaced the unconditional promise to pause if safeguards fall behind with a conditional one, and added more public reporting.
Anthropic, Google, Microsoft, OpenAI (founders); Amazon and Meta joined May 2024 · 26 July 2023
An industry club of six large AI companies that shares safety research and funds some external work. It has produced technical guidance but has no power over its members.
Biden-Harris White House with Amazon, Anthropic, Google, Inflection, Meta, Microsoft, OpenAI (7 initial; 8 more Sept 2023) · 21 July 2023
In 2023 the biggest AI companies promised the White House they would test models, secure model weights, and label AI content. An independent 2025 audit found they kept about half of those promises on average, with weight security the weakest.
Center for AI Safety · 30 May 2023
A single-sentence 2023 statement that AI extinction risk deserves the same priority as pandemics and nuclear war, signed by the heads of the top labs. It shifted public debate but committed the signers to nothing specific.
Future of Life Institute · 22 March 2023
A 2023 letter asking labs to stop training systems more powerful than GPT-4 for six months. No lab paused, and several signatories went on to build bigger models, but it made AI risk a mainstream political topic.
UNESCO General Conference · 23 November 2021
The first global standard on AI ethics, agreed by nearly every country. UNESCO helps governments assess their readiness, but the recommendation has no enforcement and progress varies widely.
Google / Alphabet · 14 September 2020
Google pledged to halve its emissions and run entirely on clean power around the clock by 2030. Its emissions are now far above the baseline because of AI, and its clean-energy share has stopped improving.
Microsoft · 16 January 2020
Microsoft promised in 2020 to remove more carbon than it emits by 2030. Its emissions instead rose 25% in its 2025 fiscal year because of AI data centres, and the company now concedes the target is getting harder.
OECD Council · 2019-05-22 (updated 2024-05-03)
The first intergovernmental AI principles, adopted by dozens of countries and updated in 2024 for generative AI. They shape national policies and reporting, but are recommendations rather than binding rules.
Université de Montréal (with Fonds de recherche du Québec) · 4 December 2018
A Quebec-led set of ten principles asking AI developers to respect well-being, autonomy, privacy, and the environment. Widely cited, but it binds no one and Canada's attempt at an AI law lapsed in 2025.
Google / Google DeepMind · 2018-06-07 (original); 2025-02-04 (revised)
In 2018 Google promised not to build AI for weapons or norm-violating surveillance. In 2025 it quietly deleted that promise and has since expanded military AI work, over objections from hundreds of its own staff.
OpenAI · 9 April 2018
OpenAI's founding promise to develop AGI for all of humanity, avoid a dangerous race, and put safety first. The company has since converted to a for-profit structure, and watchdogs have documented a long list of gaps between the charter and its conduct.
Future of Life Institute (Beneficial AI 2017 conference) · 6 January 2017
A 2017 list of 23 principles for beneficial AI that many leading researchers and executives signed. It set the vocabulary for later debates but created no obligations, and nobody tracks whether signatories follow it.