Commitment tracker

Anthropic refused Defense Secretary Hegseth's 2026-02-24 ultimatum to allow 'all legal purposes'; on 2026-02-27 the Pentagon designated it a supply-chain risk and the President ordered agencies to stop using its products (https://www.techpolicy.press/a-timeline-of-the-anthropic-pentagon-dispute/). Anthropic sued; Judge Rita Lin granted a preliminary injunction (Mar 26) and on the evening of 2026-08-27 (reported 2026-08-28) ruled the designation unlawful First Amendment retaliation (https://techcrunch.com/2026/08/28/anthropic-gets-its-first-court-win-over-the-pentagons-supply-chain-risk-label/). A D.C. Circuit case remains pending. This is a rare documented case of a lab bearing a large commercial cost to keep a stated limit. Additional coverage: https://www.npr.org/2026/08/28/nx-s1-5947951/judge-says-the-pentagon-cant-designate-ai-company-anthropic-a-supply-chain-risk

The 40-member panel co-chaired by Yoshua Bengio and Maria Ressa held its first meeting on 2026-03-03 (https://en.wikipedia.org/wiki/Independent_International_Scientific_Panel_on_AI). The first Global Dialogue on AI Governance took place in Geneva on 6-7 July 2026 with the panel presenting preliminary findings (https://www.unesco.org/en/articles/global-dialogue-ai-governance-geneva-6-7-july). The bodies are advisory; the US has said it opposes 'centralized control and global governance of AI'. Looked for: any binding outcome from Geneva - none reported.

Signing gives providers a presumption of compliance with AI Act Articles 53/55. Meta's Joel Kaplan announced Meta would not sign, calling the code overreach (https://www.theregister.com/2025/07/18/meta_declines_eu_ai_guidelines/); xAI signed the safety chapter only (https://techinformed.com/meta-rejects-eu-ai-code-of-practice/). FLI's Winter 2025 AI Safety Index found even signatories' practices fall short of the code's depth (https://futureoflife.org/ai-safety-index-winter-2025/). AI Office enforcement powers (fines) begin August 2026.

v3 added a 'harmful manipulation' CCL and shutdown-resistance protocols; v3.1 added Tracked Capability Levels (https://deepmind.google/blog/strengthening-our-frontier-safety-framework/). DeepMind published FSF reports per model, e.g. Gemini 3 Pro (Nov 2025, https://deepmind.google/models/fsf-reports/gemini-3-pro/). Context: Google deleted its 2018 pledge not to build weapons or surveillance AI in Feb 2025 (https://www.vice.com/en/article/google-abandoned-core-parts-ai-pledge-weapons/) and 580+ employees protested a classified Pentagon Gemini deal in April 2026 (https://thenextweb.com/news/google-employees-classified-military-ai-pentagon).

Direct outputs: the International AI Safety Report chaired by Yoshua Bengio (interim May 2024, full Jan 2025, second edition Feb 2026, https://internationalaisafetyreport.org/) and the UK and US AI Safety Institutes. The summit series continued (Seoul May 2024, Paris Feb 2025, New Delhi Feb 2026) but the 'safety' framing was diluted: the US and UK refused the Paris statement (https://techcrunch.com/2025/02/11/as-us-and-uk-refuse-to-sign-ai-action-summit-statement-countries-fail-to-agree-on-the-basics) and the US withheld support from the 2026 safety report. Looked for: any binding follow-on agreement - none.

The OECD launched a reporting framework on 2025-02-07 and published the first 19 company reports on 2025-04-24 (Anthropic, Google, Microsoft, OpenAI among them; Meta and xAI absent) (https://oecd.ai/en/wonk/initial-submissions-g7-hiroshima-ai-process-reporting-framework). A streamlined SME framework followed in May 2026 (https://www.oecd.org/en/about/news/press-releases/2026/05/oecd-launches-streamlined-hiroshima-ai-process-reporting-framework-to-help-small-and-medium-sized-enterprises-participate.html). Reports are self-attested and unaudited. Looked for: consequences for non-reporters - none.

The FMF launched a $10M+ AI Safety Fund (Oct 2023) and publishes issue briefs on evaluations and bio/cyber thresholds (https://www.frontiermodelforum.org/). It has not admitted xAI, Mistral, or any Chinese lab, and its members were also the only companies scoring >=60% on the White House commitments audit (https://arxiv.org/abs/2508.08345). Looked for: any binding member standard or public compliance review - none; the FMF describes its outputs as best-practice sharing.

UNESCO runs the Readiness Assessment Methodology (RAM) with member states and held its 3rd Global Forum on the Ethics of AI in Bangkok, 24-27 June 2025, launching a Global Network of Supervisory Authorities on AI (https://www.unesco.org/en/articles/advancing-ethical-ai-unescos-3rd-global-forum-bangkok-marks-new-chapter-global-cooperation). UNESCO's Global AI Ethics and Governance Observatory tracks country RAM reports (https://www.unesco.org/ethics-ai/en). Implementation is voluntary and uneven; UNESCO does not publish a count of countries that have completed a RAM in the forum summary. Looked for: a public RAM completion tally - not found in the 2025 forum materials.

The 2024 update added general-purpose/generative AI language, misinformation, and environmental sustainability (https://www.oecd.org/en/about/news/press-releases/2024/05/oecd-updates-ai-principles-to-stay-abreast-of-rapid-technological-developments.html). The OECD operates the OECD.AI policy observatory and now hosts the G7 Hiroshima reporting framework (https://oecd.ai/en/hiroshima). The principles are soft law: adherence is monitored via OECD reporting, not enforced. Looked for: sanctions or compliance findings - none exist by design.

Signed with negotiated amendments aligning it to California's SB 53, then finalised via chapter amendment on 2026-03-27; an oversight office sits in the NY Department of Financial Services (https://www.wiley.law/alert-New-York-Finalizes-RAISE-Act-for-Frontier-AI-Models-Law-Takes-Effect-January-1-2027). Like SB 53 it faces the federal preemption executive order. Not yet in force, so no compliance record.

The first US frontier-AI law, after SB 1047 was vetoed in 2024; it requires published frontier frameworks, transparency reports, and 15-day critical-incident reporting (https://en.wikipedia.org/wiki/Transparency_in_Frontier_Artificial_Intelligence_Act). It is a target of the Dec 2025 federal preemption order and DOJ litigation task force (https://www.lawfaremedia.org/article/trump-signs-executive-order-to-preempt-state-ai-laws). OpenAI subpoenaed SB 53 advocates Encode during the legislative fight (https://fortune.com/2025/10/10/a-3-person-policy-non-profit-that-worked-on-californias-ai-safety-law-is-publicly-accusing-openai-of-intimidation-tactics). Looked for: first enforcement actions or incident filings - none public yet.

EO 14179 revoked EO 14110's policies and ordered an action plan within 180 days; the plan was delivered 2025-07-23 (https://www.whitehouse.gov/wp-content/uploads/2025/07/Americas-AI-Action-Plan.pdf) with 90+ actions on deregulation, infrastructure and exports. The 2025-12-11 order created a DOJ AI Litigation Task Force to challenge state AI laws and conditioned BEAD broadband funds on states not enforcing 'onerous' AI laws (https://www.whitehouse.gov/presidential-actions/2025/12/eliminating-state-law-obstruction-of-national-artificial-intelligence-policy/; https://www.lawfaremedia.org/article/trump-signs-executive-order-to-preempt-state-ai-laws). The administration also refused to sign the Paris statement and withheld support for the 2026 International AI Safety Report. Meanwhile CAISI expanded voluntary pre-release testing to Google, Microsoft and xAI in May 2026 (https://www.cnbc.com/2026/05/05/ai-oversight-trump-google-microsoft-xai.html).

Milestones met: prohibited practices applied 2025-02-02; GPAI obligations applied 2025-08-02 with the Code of Practice signed by OpenAI, Anthropic, Google, Microsoft, Mistral (Meta refused; xAI signed only the safety chapter) (https://techinformed.com/meta-rejects-eu-ai-code-of-practice/). Milestone delayed: the Digital Omnibus (Regulation (EU) 2026/1744, OJ 2026-07-24, in force 2026-07-27; https://digital-strategy.ec.europa.eu/en/news/ai-omnibus-enters-force), pushed high-risk obligations from 2026-08-02 to 2027-12-02 (Annex III) and 2028-08-02 (Annex I), softened AI-literacy duties, and gave a watermarking grace period to 2026-12-02 (https://www.whitecase.com/insight-alert/eu-ai-omnibus-enters-force-amending-ai-act). The AI Office's powers over GPAI providers were expanded in the same package.

The EU deposited its ratification on 2026-05-15, with effects for the EU from 2026-09-01 (https://www.coe.int/en/web/artificial-intelligence/-/european-union-ratifies-the-council-of-europe-framework-convention-on-artificial-intelligence; https://www.nicfab.eu/en/posts/cets225-ratifica-ue/). The US signed under Biden (Sept 2024) but has not ratified, and the Trump administration has rejected 'global governance of AI' (https://en.wikipedia.org/wiki/Global_call_for_AI_red_lines). Looked for: the full ratification tally on the CoE treaty office page - the CoE page fetched lists parties but the count of ratifications beyond the EU could not be confirmed; treat number of ratifying states as unknown.

The declaration set seven 'pillars' and was paired with $250B+ in infrastructure pledges and ~$20B deep-tech funding announced at the summit (https://www.outlookbusiness.com/news/ai-impact-summit-2026-concludes-with-88-nations-organisations-adopting-new-delhi-declaration). It contains no safety thresholds or red lines, and India's minister declined to name non-signatories. Looked for: a monitoring mechanism for pledges - none announced.

Calls for a prohibition on developing superintelligence until there is scientific consensus on safety and public buy-in (https://www.cnbc.com/2025/10/22/800-petition-signatures-apple-steve-wozniak-and-virgin-richard-branson-superintelligence-race.html). No government has adopted a prohibition; the UK House of Lords debated superintelligence in Jan 2026 (https://hansard.parliament.uk/lords/2026-01-29/debates/848C6617-4CB6-463C-9957-82C53ACA2858/SuperintelligentAI) and 100+ UK parliamentarians signed ControlAI's statement for binding regulation (https://controlai.news/p/100-uk-parliamentarians-acknowledge). Frontier labs (Meta Superintelligence Labs, OpenAI, Anthropic) continued to pursue the goal explicitly.

US VP Vance rejected the text as risking 'ideological bias' and 'authoritarian censorship'; the UK cited national interest (https://techcrunch.com/2025/02/11/as-us-and-uk-refuse-to-sign-ai-action-summit-statement-countries-fail-to-agree-on-the-basics). Concrete outputs were the 'Current AI' public-interest foundation and a coalition on sustainable AI; safety was largely dropped from the agenda, drawing criticism from Anthropic's CEO. The next summit (New Delhi, Feb 2026) produced a broader declaration signed by 88 nations and organisations (https://www.outlookbusiness.com/news/ai-impact-summit-2026-concludes-with-88-nations-organisations-adopting-new-delhi-declaration).

Companies promised to publish safety frameworks with 'red line' thresholds by the Paris summit (2025-02-10). The Midas Project's Seoul Commitment Tracker graded 6 of 16 as fulfilled, 4 partial, 6 unfulfilled at the deadline (https://www.seoul-tracker.org/). xAI published only a draft, promised a final within three months, missed that too (May 2025), and finally released a framework on 2025-08-20 (https://www.themidasproject.com/article-list/xai-misses-a-second-self-imposed-deadline-to-implement-a-frontier-safety-policy; https://data.x.ai/2025-08-20-xai-risk-management-framework.pdf). No government has published its own compliance review.

A 2025 peer-reviewed audit (AIES) of all 16 signatories found average fulfilment of 53%, with OpenAI highest (83%) and Apple lowest (13%); 11 of 16 scored zero on model-weight security and 8 scored zero on third-party reporting (https://arxiv.org/abs/2508.08345). MIT Technology Review's one-year check (July 2024) found progress mostly on red-teaming and watermarking (https://www.technologyreview.com/2024/07/22/1095193/ai-companies-promised-the-white-house-to-self-regulate-one-year-ago-whats-changed/). No successor administration mechanism tracks them after EO 14110's rescission.

The statement asks only that extinction risk be 'a global priority'. Signatory CEOs subsequently signed the White House voluntary commitments (July 2023) and Seoul commitments (May 2024), yet by 2025-26 OpenAI loosened its Preparedness Framework and Anthropic dropped its pause language in RSP v3 (https://www.governance.ai/analysis/anthropics-rsp-v3-0-how-it-works-whats-changed-and-some-reflections). CAIS's Action Fund co-sponsored California SB 1047 (vetoed 2024) and supported SB 53 (signed 2025) (https://en.wikipedia.org/wiki/Center_for_AI_Safety). Looked for: measurable prioritisation by signatory governments - the UK/US AI safety institutes (Nov 2023) are the clearest downstream action.

The declaration influenced Canadian discourse (Montreal's Algora Lab and Mila cite it) but Canada's proposed AI and Data Act (Bill C-27) died when Parliament was prorogued in January 2025, so no Canadian statutory framework implementing its principles exists (https://www.parl.ca/legisinfo/en/bill/44-1/c-27). Looked for: a follow-up implementation report from the declaration's secretariat; the site remains a static signatory page.

No enforcement mechanism was ever attached; the principles are cited in later documents (e.g., they are referenced in California's 2018 ACR 215 legislative resolution endorsing them) but no signatory institution is bound by them. FLI itself later moved to harder asks: the 2023 'Pause' letter and the 2025 'Statement on Superintelligence' (https://superintelligence-statement.org/). Principle 22 ('Recursive Self-Improvement... strict safety and control measures') and Principle 23 ('Common Good') have no reporting or audit trail. Looked for: any signatory compliance reporting - none found.

The call asks for an international agreement on prohibited AI uses by end of 2026. The US (Michael Kratsios) rejected global AI governance days later (https://en.wikipedia.org/wiki/Global_call_for_AI_red_lines). Partial echoes: Anthropic refused to remove weapons/surveillance restrictions for the Pentagon (Feb 2026), and Senator Schiff proposed codifying those limits (Mar 2026). No treaty negotiation on red lines has started as of 2026-09-01. Looked for: UN or summit text adopting 'red lines' - the New Delhi Declaration did not.

The UK government's proposed text-and-data-mining opt-out was fought by the creative sector ('Make It Fair', Feb 2025) and the Data (Use and Access) Act 2025 passed without the copyright exception after repeated Lords amendments. Courts moved in different directions: Getty largely lost against Stability AI in the UK (Nov 2025, https://www.lw.com/en/insights/getty-images-v-stability-ai-english-high-court-rejects-secondary-copyright-claim) while GEMA won against OpenAI in Munich (Nov 2025, https://www.reedsmith.com/our-insights/blogs/viewpoints/102lrzg/copyright-ai-german-court-issues-landmark-decision-in-favor-of-rights-holders/) and Anthropic paid $1.5B to authors (2026). No legislature has mandated licensing for training.

No lab paused; Musk founded xAI in July 2023 and OpenAI, Google, Anthropic and Meta all trained larger models (GPT-5 shipped August 2025) (https://en.wikipedia.org/wiki/Pause_Giant_AI_Experiments:_An_Open_Letter). When the six months expired FLI pivoted to calling for regulation (https://futureoflife.org/ai/six-month-letter-expires/). The letter spawned the PauseAI movement (https://pauseai.info/) and FLI's 2025 Statement on Superintelligence. Looked for: any documented training pause by a signatory company - none.

Google's 2026 Environmental Report (released 2026-06-30) shows its total (ambition-based) 2025 footprint up 18% year on year and 81% above the 2019 baseline — Google's own blog headlines a 2% cut in operational emissions, so the two figures measure different scopes, with carbon-free energy flat at ~65% despite a 37% jump in electricity use; the company says the path 'will not be linear' (https://www.esgdive.com/news/googles-emissions-continue-to-climb-due-to-ai-buildout/824373/; https://blog.google/company-news/outreach-and-initiatives/sustainability/2026-environmental-report/). Looked for: a revised target - Google has not formally withdrawn the 2030 goals.

Microsoft's 2026 sustainability report (published 2026-07-09) shows emissions rose 25% in fiscal year 2025 (July 2024–June 2025; 20.29 vs 16.21 MtCO2e) as AI data centres expanded, and the company said 'sustainability solutions are not scaling fast enough to meet demand' while reaffirming the 2030 goal (https://www.engadget.com/2212136/microsoft-ai-drive-carbon-emissions-grow-25-percent-2025/; https://thehill.com/policy/technology/5966327-microsoft-emissions-jump-ai-data/). Emissions were already ~30% above the 2020 baseline by 2023 (https://www.theregister.com/2024/05/16/microsoft_co2_emissions/). Report hub: https://www.microsoft.com/en-us/corporate-responsibility/sustainability/report.

The April 2025 rewrite added a clause allowing OpenAI to 'adjust' safeguards if a rival ships a high-risk system without comparable protections, removed persuasion as a tracked category, and shifted toward automated evaluations; the Financial Times reported testers were given under a week for some models (https://techcrunch.com/2025/04/15/openai-says-it-may-adjust-its-safety-requirements-if-a-rival-lab-releases-high-risk-ai). OpenAI has published system cards with Preparedness scorecards for each frontier release and signed the EU GPAI Code of Practice (Aug 2025). The Midas Project's Watchtower logs framework changes (https://www.themidasproject.com/projects). Looked for: any instance of OpenAI delaying a launch under the framework - none publicly documented.

Anthropic activated ASL-3 protections for Claude Opus 4 (May 2025) under v2.x. RSP v3.0 (effective 2026-02-24, https://www.anthropic.com/responsible-scaling-policy/rsp-v3-0) replaced the unconditional commitment not to train or deploy without adequate safeguards with conditional delay commitments (Appendix A: 'We will delay AI development and deployment as needed' — but only if Anthropic has a clear lead, or until it matches competitors' risk posture), alongside periodic Risk Reports and a non-binding Frontier Safety Roadmap; GovAI's analysis calls this an honest but weakened commitment (https://www.governance.ai/analysis/anthropics-rsp-v3-0-how-it-works-whats-changed-and-some-reflections). Anthropic did hold a separate red line: it refused the Pentagon's Feb 2026 demand to drop weapons/surveillance restrictions and won an Aug 28 2026 ruling that the retaliatory blacklist was unlawful (https://techcrunch.com/2026/08/28/anthropic-gets-its-first-court-win-over-the-pentagons-supply-chain-risk-label/).

OpenAI completed a restructuring on 2025-10-28/29 into OpenAI Group PBC, with the nonprofit OpenAI Foundation holding an equity stake (~26%, ~$130B) and Microsoft ~27%; California and Delaware AGs attached conditions on safety and charitable assets (https://sfist.com/2025/10/29/openai-finalizes-restructuring-with-nonprofit-in-control-state-ag-oks-it/). The Midas Project's 'OpenAI Files' (June 2025) catalogues departures from the charter's commitments (https://www.openaifiles.org/). OpenAI also subpoenaed nonprofits (Encode, The Midas Project) that criticised the restructuring (https://fortune.com/2025/10/10/a-3-person-policy-non-profit-that-worked-on-californias-ai-safety-law-is-publicly-accusing-openai-of-intimidation-tactics). Looked for: any public 'assist' or merge-and-assist action under the charter - none found.

Rescinded on 2025-01-20 by EO 14148 and replaced by EO 14179 on 2025-01-23 (https://en.wikipedia.org/wiki/Executive_Order_14179). Before rescission it created the US AI Safety Institute (Nov 2023), OMB memo M-24-10, and NIST's dual-use guidance; several agency work products survive under new names (US AISI became CAISI in June 2025, https://www.nist.gov/caisi). Looked for: whether the 10^26 reporting requirement continued - it did not; CAISI now relies on voluntary agreements.

On 2025-02-04 Google deleted the 'applications we will not pursue' section (weapons, surveillance violating international norms, technologies causing overall harm) in a post by Demis Hassabis and James Manyika (https://www.vice.com/en/article/google-abandoned-core-parts-ai-pledge-weapons/; https://www.cnbc.com/2025/02/04/google-removes-pledge-to-not-use-ai-for-weapons-surveillance.html). Senator Markey wrote to Google demanding explanation (https://www.markey.senate.gov/download/letter-to-google-on-ai-principles-revisions?download=1). Google then deployed Gemini to 3 million Pentagon users (Mar 2026) and negotiated a classified 'all lawful uses' deal, prompting a 580+-employee letter (https://thenextweb.com/news/google-employees-classified-military-ai-pentagon).