AI development races ahead of governance. Frontier labs operate with voluntary and inconsistent safety commitments. Dangerous capabilities emerge before international agreement is possible.
AI development operates within internationally agreed limits. Safety thresholds are shared across jurisdictions. Export controls and compute governance align across allies. Dangerous capabilities are coordinated, not raced.
UN opened worldwide public submission portal (deadline 30 Apr 2026) for input into the first Global Dialogue on AI Governance, to be held in 2026 in Geneva back-to-back with the ITU AI for Good Summit. Dialogue + Independent International Scientific Panel on AI established by UNGA Resolution A/RES/79/325 (adopted by consensus 26 Aug 2025). Signals formal public-expectation channel for cross-border AI governance.
The European Parliament approved the European Union's conclusion of the Council of Europe Framework Convention on Artificial Intelligence.
On 11 Mar 2026 the European Parliament gave consent (455 in favour, 101 against, 74 abstentions) to EU conclusion of the Council of Europe Framework Convention on AI and Human Rights (CETS 225), the first legally binding international AI treaty. Advances ratification; treaty enters into force after 5 states ratify.
The OECD published a report analyzing trends in AI-related incidents and hazards as reported by the media.
On 24 Feb 2026 Anthropic released RSP v3.0, removing its previous implication that it would pause training/deployment if risks exceeded acceptable levels. Some mitigations (e.g. RAND SL4) reframed as 'industry-wide recommendations' rather than unilateral commitments. GovAI and multiple safety commentators flagged this as material weakening of the flagship voluntary framework.
At the India AI Impact Summit (16-21 Feb 2026; declaration adopted 18-19 Feb, announced 21 Feb 2026), 92 countries and international organisations — including US, UK, China, Russia, EU, Switzerland — endorsed the New Delhi Declaration committing to international cooperation on safe, inclusive AI. Substantially broader than Paris 2025 signatory base; US re-engaged. (Initial 88 on 21 Feb, grew to 91 by 24 Feb, 92 as of 5 Mar 2026.)
Future of Life Institute (UN SG's designated civil society co-champion for AI) released pre-summit recommendations calling for Global South participation in AI governance, national audit capacity, and commensurate safety guarantees from developers. Published 15 Feb 2026, days before the 16-21 Feb New Delhi summit.
European coalition of civil society, trade unions, and academics (60+ organisations incl. Access Now, EDRi, Amnesty Tech, BEUC) signed open letter to MEPs and Commission urging rejection of AI Omnibus amendments that would delete Art. 49(2) high-risk transparency safeguard.
Second annual International AI Safety Report chaired by Yoshua Bengio, backed by EU/OECD/UN and 30+ countries, released 3 Feb 2026 ahead of the India AI Impact Summit. 200-page report with 1,451 references. Confirms scientific consensus on escalating risks (cyber, bio, loss of control, eval-aware models). (Note: arxiv preprint 2602.21012 uploaded 24 Feb 2026; primary release was UK gov on 3 Feb 2026.)
On 18 Dec 2025 the UK AI Security Institute released its first Frontier AI Trends Report, synthesizing two years of testing 30+ frontier models. Key findings: cyber apprentice-level task success rose from <9% (2023) to ~50% (2025); first expert-level cyber task completed in 2025; models outperform PhD-level experts on chem/bio knowledge; hour-long software tasks completed >40% of the time. Shared publicly to inform the International Network's evaluation science.
On 11 Dec 2025 Trump signed EO 'Ensuring a National Policy Framework for AI' creating an AI Litigation Task Force at DOJ (established by AG memo 9 Jan 2026) to challenge state AI laws and conditioning federal grants on states not enforcing them. Colorado AI Act explicitly named. Preempts subnational compute/licensing experiments that could have aligned with international norms.
At the 4-5 Dec 2025 San Diego meeting, the network (launched at Seoul Summit May 2024; formally established Nov 2024) was renamed 'International Network for Advanced AI Measurement, Evaluation and Science' — widely read as a concession to keep the US (CAISI) engaged. UK took the Network Coordinator role. Australia joined after 25 Nov 2025 AU AISI announcement ($29.9M, operating early 2026). Members: AU, CA, EU, FR, JP, KE, KR, SG, UK, US.
On 2 Aug 2025 the EU AI Act's GPAI rules took effect, codifying a 10^25 FLOP training-compute threshold for 'systemic-risk' GPAI models subject to safety, security, and transparency obligations (10^23 FLOP threshold for baseline GPAI). First binding international compute-based safety threshold in force. Full enforcement from 2 Aug 2026.
The EU GPAI Code of Practice (Transparency, Copyright, Safety & Security chapters) was published on 10 Jul 2025 and endorsed as an 'adequate voluntary tool' by the Commission and AI Board. Signatories include Amazon, Anthropic, Google, Microsoft, OpenAI, Mistral, Cohere, IBM, Aleph Alpha, and others. Meta and Chinese providers declined. xAI signed only the Safety & Security chapter.
UK AISI renamed 'AI Security Institute' on 13 Feb 2025 (announced at Munich Security Conference by Peter Kyle), shifting emphasis from safety to security/national defense. US counterpart renamed 'Center for AI Standards and Innovation' (CAISI) on 3 Jun 2025 with explicit mandate to serve industry as NIST-housed primary point of contact. Signals reframing of voluntary evals posture in both lead jurisdictions.