AI Safety Summit
Last reviewed
Sources
32 citations
Review status
Source-backed
Revision
v5 · 7,858 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
32 citations
Review status
Source-backed
Revision
v5 · 7,858 words
Add missing citations, update stale details, or suggest a clearer explanation.
The AI Safety Summit is a recurring series of intergovernmental summits on the risks and governance of advanced artificial intelligence, launched by the United Kingdom at Bletchley Park in November 2023 and continued in Seoul (2024), Paris (2025), and New Delhi (2026), with Geneva confirmed for 2027 and the United Arab Emirates slated for 2028. Its founding text, the Bletchley Declaration, signed by 28 countries and the European Union on 1 November 2023, was the first international agreement to recognise the risks posed by frontier AI. The series is the most sustained multilateral effort to date to coordinate global responses to AI safety concerns, bringing together heads of state, technology executives, researchers, and civil society organisations.
The AI Safety Summit refers to a series of international summits convened by national governments to address the risks and governance challenges posed by advanced artificial intelligence systems. The series began with the Bletchley Park Summit in November 2023 and has continued through events in Seoul, Paris, and New Delhi, with a fifth summit confirmed for Geneva in 2027 and a sixth slated for the United Arab Emirates in 2028. These summits represent the most sustained multilateral effort to date to coordinate global responses to AI safety concerns, bringing together heads of state, technology executives, researchers, and civil society organisations.
The summit series emerged against a backdrop of rapid advances in large language models and growing alarm among researchers about the potential risks of frontier AI. In the months preceding the first summit, the Center for AI Safety released a one-sentence statement signed by hundreds of prominent figures warning that "mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war," and the Future of Life Institute's open letter calling for a six-month pause on giant AI experiments had drawn over 30,000 signatures [1][2]. The extinction statement was signed by, among others, Turing Award winners Geoffrey Hinton and Yoshua Bengio and the chief executives of OpenAI, Google DeepMind, and Anthropic, lending the catastrophic-risk framing the institutional weight that the British government would later build the summit around [1].
What began as a tightly focused gathering on catastrophic risk at Bletchley Park has, over four iterations, broadened into a wider conversation about innovation, public-good infrastructure, and inclusive deployment. The branding shift is itself part of the story: the 2023 and 2024 events were called "AI Safety Summit" and "AI Seoul Summit" respectively, while Paris in 2025 dropped "safety" entirely in favour of "AI Action Summit," and New Delhi in 2026 styled itself the "AI Impact Summit." Whether this evolution reflects a healthy maturation of the agenda or a quiet retreat from the frontier-risk frame remains contested among diplomats, civil society groups, and AI labs themselves.
The summit format was a British initiative. On 7 June 2023, then UK Prime Minister Rishi Sunak announced that the United Kingdom would host the world's first global summit on AI safety in autumn 2023. The announcement was the centrepiece of a wider UK pitch to lead international discussion of frontier AI, which Sunak's team had been assembling since early 2023 alongside the creation of the Frontier AI Taskforce in April of that year. The taskforce, chaired by Ian Hogarth, received an initial allocation of 100 million pounds and would later become the UK AI Security Institute [3].
The political logic behind the summit was a mix of opportunity and anxiety. Sunak's government calculated that London could play a convening role between Washington, Brussels, and Beijing on a topic where no incumbent forum existed. The G7 was running its parallel Hiroshima Process on AI; the OECD had its AI Principles; the United Nations was beginning to consider an AI advisory body. None of those tracks had a binding answer to the question of whether the world's most capable models posed catastrophic risks. By focusing the summit specifically on "frontier AI," Sunak's team set out to claim that gap.
At the same time, several governments wanted a venue to talk about AI risks without committing to specific rules. The European Union was finishing the EU AI Act. The United States had just finalised voluntary commitments from seven leading AI developers in July 2023 and was preparing the executive order on AI that President Joe Biden would sign on 30 October 2023, two days before the Bletchley Summit opened. China had introduced binding rules on generative AI in force from August 2023. A multilateral declaration that affirmed the importance of safety without imposing new obligations suited all of them.
The Bletchley Park venue was chosen for its symbolism. The country house in Buckinghamshire, about 80 kilometres north-west of London, was the wartime home of British signals intelligence work that broke German Enigma traffic. Alan Turing did much of his work there. The British government wanted the summit to evoke that history of national leadership in computing and to suggest that another moment of decisive technical and political coordination was at hand.
From the outset the format was conceived as a series rather than a one-off event. Pre-summit communications between the UK and other capitals included an agreement that the Republic of Korea would co-host a follow-up event roughly six months later, with France hosting a full successor summit one year after Bletchley. This rotating-host model has held since, with India taking over in 2026 and Switzerland slated for 2027.
The first AI Safety Summit took place on 1-2 November 2023 at Bletchley Park in Milton Keynes, United Kingdom. It was hosted by Prime Minister Rishi Sunak, who had positioned Britain as an aspiring global leader on AI governance, and ran across two days with a different format on each.
The summit brought together representatives from 28 countries and the European Union, with about 150 in-person delegates representing governments, AI labs, academia, and civil society. Notably, both the United States and China participated, marking a rare instance of cooperation on technology policy between the two rivals. China was represented by Vice Minister of Science and Technology Wu Zhaohui. Other state attendees included delegations from Brazil, France, Germany, India, Ireland, Israel, Italy, Japan, Kenya, the Kingdom of Saudi Arabia, Nigeria, the Philippines, Rwanda, Singapore, South Korea, Spain, Switzerland, Turkey, Ukraine, the United Arab Emirates, and others. The European Commission also signed alongside individual EU member states, with Commission President Ursula von der Leyen attending in person [4].
US Vice President Kamala Harris led the American delegation. Italian Prime Minister Giorgia Meloni was the only G7 leader other than Sunak to attend in person. UN Secretary-General António Guterres also attended, and King Charles III delivered a recorded video address. Russia was not invited.
Technology leaders in attendance included OpenAI CEO Sam Altman, Google DeepMind CEO Demis Hassabis, Anthropic CEO Dario Amodei, Meta AI chief Yann LeCun, Microsoft President Brad Smith, AWS CEO Adam Selipsky, Mustafa Suleyman of Inflection AI, Nick Clegg of Meta, and Elon Musk of xAI. On 2 November Sunak conducted a live conversation with Musk on X, which dominated UK media coverage of the summit. Academic researchers, civil society groups, and international organisations also attended [5].
The summit's central political outcome was the Bletchley Declaration, signed by all 28 countries and the EU on 1 November 2023. The declaration represented the first international agreement on the risks posed by frontier AI. It affirmed that AI should be "designed, developed, deployed, and used in a manner that is safe, in such a way as to be human-centric, trustworthy and responsible." Its central passage on catastrophic risk reads: "There is potential for serious, even catastrophic, harm, either deliberate or unintentional, stemming from the most significant capabilities of these AI models" [4].
Key elements of the declaration are summarised below:
| Element | Description |
|---|---|
| Shared risk recognition | Acknowledgement that frontier AI poses potentially catastrophic risks, including in cybersecurity, biotechnology, and disinformation |
| Need for cooperation | Commitment to international collaboration on understanding and mitigating AI risks |
| State-led safety testing | Agreement on the importance of government-led evaluation and testing of frontier AI systems |
| Developer transparency | Recognition that AI developers should be transparent about their safety practices and risk assessments |
| Research collaboration | Commitment to building shared scientific understanding through collaborative research |
The declaration was non-binding, meaning it carried no legal enforcement mechanism. Critics noted this limitation, arguing that voluntary commitments without enforcement teeth would be insufficient to address the pace of AI development [6]. New Zealand subsequently joined the declaration on 23 October 2024, bringing the total to 29 states plus the EU.
One of the most concrete outcomes of the Bletchley Summit was the announcement of the UK AI Safety Institute (later renamed the UK AI Security Institute), tasked with testing and evaluating frontier AI models. The institute received approximately 100 million pounds in initial public funding and quickly built one of the world's largest safety evaluation teams. It was the first government body specifically dedicated to evaluating the safety of advanced AI systems before and after deployment [7].
Vice President Kamala Harris also used the summit to announce the US AI Safety Institute, to be housed within NIST under the authority of Biden's executive order. The two institutes signed a partnership agreement in April 2024 and conducted joint pre-deployment evaluations of frontier models from OpenAI, Anthropic, and others over the following year.
The summit also commissioned Yoshua Bengio, a Turing Award-winning deep learning researcher, to lead the production of a "State of the Science" report on the capabilities and risks of frontier AI, to be delivered ahead of the next summit. That report eventually became the International AI Safety Report 2025 [5].
Participants agreed on a schedule for future summits: the Republic of Korea would co-host a follow-up event within six months, and France would host the next full in-person summit approximately one year later. Eight major developers (Anthropic, Google DeepMind, OpenAI, Meta, Microsoft, AWS, Mistral AI, and Inflection AI) signed a separate commitment to support pre-deployment evaluations of their forthcoming models. Several countries also announced bilateral AI safety agreements and research partnerships [5].
The AI Seoul Summit was held on 21-22 May 2024, co-hosted by the Republic of Korea and the United Kingdom. The event was structured as a virtual leaders' session on Day 1 followed by an in-person ministerial meeting in Seoul on Day 2, reflecting the compressed timeline between Bletchley and Seoul and the host governments' desire to keep heads-of-government attention on the file [8].
The leaders' session was co-chaired by Sunak (whose government would lose the UK general election on 4 July 2024) and ROK President Yoon Suk Yeol. Other leaders dialling in included Italian Prime Minister Giorgia Meloni, Canadian Prime Minister Justin Trudeau, German Chancellor Olaf Scholz, French President Emmanuel Macron, Japanese Prime Minister Fumio Kishida, Australian PM Anthony Albanese, Singaporean PM Lawrence Wong, and US Vice President Kamala Harris (representing President Biden). The European Commission was again represented by Ursula von der Leyen.
The most concrete outcome of the Seoul Summit was the Frontier AI Safety Commitments, a set of voluntary pledges signed on 21 May 2024 by 16 leading AI developers. Initial signatories were Amazon, Anthropic, Cohere, Google (covering DeepMind), G42, IBM, Inflection AI, Meta, Microsoft, Mistral AI, Naver, OpenAI, Samsung Electronics, Technology Innovation Institute, xAI, and Zhipu AI. Four further companies (01.AI, Magic, MiniMax, and NVIDIA) joined the commitments later, bringing the eventual total to 20 [9].
The commitments were structured around three numbered Outcomes covering risk identification, accountability, and transparency:
| Outcome | Key Requirement |
|---|---|
| I. Risk identification and mitigation | Companies must evaluate risks throughout the AI lifecycle, define thresholds at which severe risks would be deemed intolerable, articulate mitigations to keep risks below those thresholds, and commit not to develop or deploy a model if mitigations cannot keep risks below the thresholds |
| II. Accountability | Companies must develop internal governance frameworks, assign clear roles and resources for safety oversight, and continuously review those frameworks |
| III. Transparency | Companies must provide public transparency on the implementation of the above commitments and explain how external actors are involved in risk assessment; signatories were required to publish a safety framework focused on severe risks before the Paris Summit in February 2025 |
The "do not develop or deploy" clause in commitment I.iv was the moment the Seoul Summit imported the central idea of Anthropic's Responsible Scaling Policy into a multilateral text. The publish-by-Paris deadline turned the Seoul Commitments into a tracking exercise. By February 2025 the evaluation organisation METR counted 12 of the 20 signatories with published frameworks meeting the Seoul standard, including Anthropic's RSP v2.0, OpenAI's Preparedness Framework v2, Google DeepMind's Frontier Safety Framework, Microsoft's Frontier Governance Framework, and Meta's Frontier AI Framework [9].
The leaders also adopted the Seoul Declaration for Safe, Innovative and Inclusive AI (the Seoul Declaration), endorsed on 21 May by 10 countries plus the EU. On 22 May, technology and digital ministers from 27 countries plus the EU adopted a more concrete Seoul Ministerial Statement, which named specific severe-risk categories: AI's potential to lower barriers to chemical and biological weapons by helping non-state actors, and AI's capacity to evade human oversight through manipulation, deception, or autonomous replication and adaptation. Ministers also committed for the first time to develop "shared risk thresholds" for frontier AI [8].
A notable absence at Seoul was China, which had signed the Bletchley Declaration but was not invited to the Seoul leaders' session and did not sign the ministerial statement. The decision to omit China was a deliberate choice by the co-hosts on the rationale that Seoul would focus on "like-minded" democracies. Brazil and Ireland, both Bletchley signatories, also did not appear in either Seoul document.
Alongside the declaration, the same 11 leaders endorsed the Seoul Statement of Intent toward International Cooperation on AI Safety Science. This shorter document committed signatories to support the creation of, and cooperation among, national AI safety institutes, and articulated a "shared ambition to develop an international network among key partners to accelerate the advancement of the science of AI safety."
This commitment became, six months later, the International Network of AI Safety Institutes, formally launched in San Francisco on 20-21 November 2024 with 10 founding members: Australia, Canada, the European Union, France, Japan, Kenya, the Republic of Korea, Singapore, the United Kingdom, and the United States. The Network announced more than 11 million US dollars in funding for synthetic content research, the findings of its first multilateral testing exercise, and a joint statement on risk assessments of advanced AI systems [10].
The UK AI Safety Institute also announced 8.5 million pounds in research funding for systemic AI safety work as part of the summit outcomes.
An interim version of the International Scientific Report on the Safety of Advanced AI, chaired by Yoshua Bengio and mandated at Bletchley, was published on 17 May 2024 and presented to the Seoul leaders' session as the principal evidence base for their discussion. The interim report was produced with input from 30 nations plus the UN, OECD, and EU through an Expert Advisory Panel and laid the groundwork for the full International AI Safety Report 2025 delivered at Paris.
Between the Seoul and Paris summits, a parallel scientific track was developed. On 26 April 2025, the Singapore Conference on AI (SCAI): International Scientific Exchange on AI Safety convened on the sidelines of the International Conference on Learning Representations (ICLR) 2025 in Singapore. The event was organised by Singapore's Infocomm Media Development Authority (IMDA) and brought together over 100 participants from 11 countries (Australia, Canada, Chile, China, France, Japan, Republic of Korea, Netherlands, Singapore, the United Kingdom, and the United States), including academics, industry researchers, government representatives, and policy specialists. Yoshua Bengio delivered the keynote address and Max Tegmark gave the closing address [29].
The conference produced The Singapore Consensus on Global AI Safety Research Priorities, a technical research agenda organised around three core areas:
| Area | Focus |
|---|---|
| Risk Assessment | Tools to measure potential harm from AI systems, improve measurement accuracy, and enable third-party audits |
| Development | Approaches for designing AI systems that behave as intended through specification, system design, and verification |
| Control | Monitoring and intervention mechanisms to manage AI behaviour in real time and enhance societal resilience |
The Singapore Consensus was framed by its organisers as a research-side counterpart to the political declarations of the summit series: a document on what AI safety researchers should work on, rather than what governments should declare. It was subsequently presented at the Asia Tech x Singapore Summit on 28-29 May 2025 and incorporated into the briefing materials for the India AI Impact Summit [29].
The AI Action Summit ("Sommet pour l'action sur l'intelligence artificielle") was held on 10-11 February 2025 at the Grand Palais in Paris, France. The Grand Palais summit followed a week of preparatory events: "Science Days" on 6-7 February at the Institut Polytechnique de Paris, a "Cultural Weekend" on 8-9 February, and the high-level political segment on 10-11 February. The summit was hosted by President Emmanuel Macron and co-chaired with Indian Prime Minister Narendra Modi, an arrangement that signalled the gradual broadening of the summit series toward the Global South [11].
The rebranding from "AI Safety Summit" to "AI Action Summit" was deliberate. French officials briefed in the run-up to the event that they wanted to broaden the agenda beyond catastrophic risk to encompass innovation, sustainability, public-good infrastructure, and labour market impacts. The change of name was negotiated between the French and outgoing UK governments and reflected, among other things, Macron's view that an exclusive focus on existential risk was politically self-limiting and could push France's domestic AI champions, especially Mistral AI, to the margins of policy discussions.
The Paris summit was substantially larger than its predecessors. Participants from over 100 countries attended, with more than 1,000 in-person delegates including heads of state, technology executives, researchers, and civil society leaders. Beyond Macron and Modi, attendees included US Vice President JD Vance, Chinese Vice Premier Zhang Guoqing, EU Commission President Ursula von der Leyen, German Chancellor Olaf Scholz, Italian PM Giorgia Meloni, Canadian PM Justin Trudeau, and UN Secretary-General António Guterres. Sundar Pichai of Google, Sam Altman of OpenAI, Dario Amodei of Anthropic, Demis Hassabis of Google DeepMind, and Arthur Mensch of Mistral AI led the corporate contingent [11].
The political flashpoint of the summit was the speech delivered by US Vice President JD Vance on 11 February 2025, which marked the first major foreign policy statement of the second Donald Trump administration on AI. Vance opened by saying "I'm not here this morning to talk about AI safety, which was the title of the conference a couple of years ago. I'm here to talk about AI opportunity" [12].
Vance laid out four priorities for the Trump administration's approach to AI:
The speech was understood as a direct rebuke of European regulatory frameworks, particularly the EU AI Act and the burden it placed on general-purpose AI providers. Framing his case against caution, Vance told delegates: "When conferences like this convene to discuss a cutting-edge technology, oftentimes, I think our response is to be too self-conscious, too risk-averse. But never have I encountered a breakthrough in tech that so clearly calls us to do precisely the opposite" [12]. Vance also warned against partnerships with "authoritarian regimes," a thinly veiled reference to China and to companies dependent on Chinese AI infrastructure [13].
The summit's headline diplomatic outcome was the Statement on Inclusive and Sustainable Artificial Intelligence for People and the Planet. The statement called for AI policies that are "open, inclusive, transparent, ethical, safe, secure and trustworthy," and structured itself around six objectives: AI accessibility, ethical and trustworthy systems, innovation without market concentration, positive labour market impacts, environmental sustainability, and strengthened international cooperation [14].
The text was endorsed by 64 countries plus the European Union and the African Union Commission, for a total of 66 signatories. Signatories spanned Africa (Côte d'Ivoire, Djibouti, Egypt, Kenya, Morocco, Nigeria, Rwanda, Senegal), the Americas (Brazil, Canada, Chile, Colombia, Mexico, Peru, Uruguay), Asia-Pacific (Armenia, Australia, Cambodia, China, India, Indonesia, Japan, Kazakhstan, New Zealand, Republic of Korea, Singapore, Thailand), and most of Europe (Austria, Belgium, Bulgaria, Croatia, Cyprus, Czech Republic, Denmark, Estonia, Finland, France, Germany, Greece, Hungary, Iceland, Ireland, Italy, Latvia, Liechtenstein, Lithuania, Luxembourg, Malta, Moldova, Monaco, Netherlands, Norway, Poland, Portugal, Romania, Serbia, Slovakia, Slovenia, Spain, Sweden, Switzerland, Ukraine) [14].
Two notable refusals drew significant attention: the United States and the United Kingdom both declined to sign. A spokesperson for UK Prime Minister Keir Starmer's government stated that the declaration "didn't provide enough practical clarity on global governance" and did not "sufficiently address harder questions around national security." Vance and other US officials cited the same governance concerns and the framing around market concentration as reasons not to sign [13][15].
The absence of US and UK signatures did not prevent China from signing, which left observers reading the Paris outcome as a partial inversion of Bletchley: at the founding summit, US, UK, and China all signed; at Paris, only one of those three did.
The International AI Safety Report 2025, commissioned at Bletchley Park and chaired by Yoshua Bengio, was published on 29 January 2025, just ahead of the summit. The report was produced by 96 contributing AI experts, with an Expert Advisory Panel of representatives nominated by 30 countries plus the UN, OECD, and EU. It was formally presented at the Paris summit as part of the "Trust in AI" pillar [16].
The report assessed three main risk categories: malicious use, malfunctions, and systemic risks. Its headline finding was that since Bletchley the capabilities of general-purpose AI had increased markedly, with new models showing better performance on programming and scientific reasoning, and many companies investing in autonomous AI agents. The report stopped short of recommending specific policies, treating itself as an evidence base rather than a position paper.
The Paris summit produced a wide range of investment and infrastructure announcements:
| Initiative | Description |
|---|---|
| Current AI Foundation | A 400 million US dollar foundation, backed by nine governments and a mix of philanthropic and private partners, dedicated to creating AI "public goods" including datasets and open-source tools |
| InvestAI | European Commission President Ursula von der Leyen launched a 200 billion euro initiative, including 20 billion euro for four AI "gigafactories" to train large models |
| ROOST Coalition | Google, Discord, OpenAI, and Roblox launched the Robust Open Online Safety Tools initiative to develop free, open-source tools for detecting child sexual abuse material |
| Coalition for Sustainable AI | Led by France, the UN Environment Programme, and the International Telecommunication Union, with support from 11 countries and 37 technology companies |
| Private investment | Macron announced approximately 110 billion euro in private investment pledges for France's AI sector, including 30-50 billion euro from the UAE and 20 billion euro from Brookfield Corporation |
| India-France AI Roadmap | Modi and Macron jointly launched a bilateral roadmap focused on safe, open, and trustworthy AI |
The "Trust in AI" pillar of the summit included sessions on the International Network of AI Safety Institutes, but the institutional centre of gravity at Paris was clearly investment rather than evaluation. Anthropic CEO Dario Amodei publicly criticised the summit afterwards as a "missed opportunity" for meaningful safety commitments, writing that "at the next international summit, we should not repeat this missed opportunity" and that on the risks of AI "we must move faster and with greater clarity to confront them." His view was echoed by several civil society organisations [11][32].
A secondary outcome of the Paris summit was Modi's announcement that India would host the next summit. The announcement formalised what had been discussed since the Seoul Summit and gave the series, for the first time, a Global South host nation outside the original UK-Korea-France lineage.
The AI Impact Summit, the fourth event in the series, was held at Bharat Mandapam in New Delhi from 16 to 21 February 2026 (originally scheduled to close on 20 February but extended by one day). It was hosted by India under Prime Minister Narendra Modi, who had pre-announced the event during his co-chairing role at Paris [17].
The summit drew over 20 heads of government and 60 ministers, with delegations from more than 100 countries and 20 international organisations. In-person attendance reached approximately 600,000 across the week, with over 900,000 cumulative views through live virtual streaming. Attendees included French President Emmanuel Macron, UN Secretary-General António Guterres, Brazilian President Luiz Inácio Lula da Silva, and Swiss President Guy Parmelin. Technology leaders included Sundar Pichai of Google, Sam Altman of OpenAI, Dario Amodei of Anthropic, Demis Hassabis of DeepMind, and Mukesh Ambani of Reliance Industries [18].
The summit's central political outcome was the New Delhi Declaration on AI Impact, adopted on 19 February 2026. As of 21 February the declaration had 88 endorsements, and Bangladesh, Costa Rica, and Guatemala subsequently joined in the week after the summit, bringing the total to 91 countries and international organisations, the largest number of signatories in the summit series. The United States joined the declaration despite the Trump administration's general scepticism of multilateral AI governance; White House Office of Science and Technology Policy director Michael Kratsios represented the US at the summit and stated that Washington would "reject global governance of AI" even as it signed the non-binding declaration [28].
The declaration's content centred on inclusive, human-centric AI development. Its main themes included democratising access to AI infrastructure, expanding AI's role in health care and education, promoting ethical safeguards and transparency, narrowing the digital divide between wealthy and developing economies, and ensuring sustainability of AI infrastructure. The text broadened the agenda further from Paris by emphasising AI for development and downplaying frontier-risk language [17].
The New Delhi summit produced several substantive initiatives:
| Outcome | Details |
|---|---|
| New Delhi Frontier AI Impact Commitments | 13 leading global and Indian frontier model developers pledged to promote trustworthy and inclusive AI deployment |
| Global AI Impact Commons | A voluntary initiative featuring more than 80 impact stories across 30+ countries, designed to share and replicate successful AI use cases |
| Indian sovereign compute | India announced an additional 20,000 GPUs for its sovereign compute capacity, adding to the 38,000+ GPUs already provisioned under the IndiaAI Mission |
| Indian indigenous models | Sarvam AI launched its language and vision models; BharatGen released its Param2 model |
| Microsoft commitment | A 50 billion US dollar investment commitment for AI infrastructure in lower-income countries |
| Reliance commitment | Reliance Industries pledged 110 billion US dollars over seven years for AI infrastructure |
| Aggregate investment | More than 200 billion US dollars in AI-related investment commitments across infrastructure, foundation models, hardware, and applications |
| Guinness World Record | India achieved a record for the most pledges received for an AI responsibility campaign in 24 hours, with 250,946 validated pledges between 16-17 February 2026 |
The Indian government framed the summit as the moment when the AI governance conversation broke definitively out of the Atlantic-North Asian core that had defined Bletchley, Seoul, and Paris. Critics, including some Indian civil society organisations, argued that the heavy emphasis on investment and sovereign infrastructure came at the expense of human rights and labour protections [19].
On 19 February 2026, during the high-level segment in New Delhi, Swiss President Guy Parmelin formally announced that Switzerland would host the next summit in Geneva in 2027. Parmelin framed the choice as a return to multilateral neutrality, calling Geneva "the epicentre of multilateralism." The Swiss Federal Council had signalled its interest several months earlier and confirmed that financing was secured and organisational preparations were already in place [20].
Parmelin also announced that the United Arab Emirates would host the 2028 summit. The UAE's selection completes a sequence of host countries from western Europe (UK, France), East Asia (Korea), South Asia (India), continental Europe (Switzerland), and the Gulf, suggesting a deliberate effort to rotate the venue across regions and political traditions.
The following table summarises the key characteristics and outcomes of each summit in the series.
| Feature | Bletchley Park (Nov 2023) | Seoul (May 2024) | Paris (Feb 2025) | New Delhi (Feb 2026) |
|---|---|---|---|---|
| Branded name | AI Safety Summit | AI Seoul Summit | AI Action Summit | AI Impact Summit |
| Host country | United Kingdom | South Korea (with UK) | France | India |
| Host leader | PM Rishi Sunak | President Yoon Suk Yeol; PM Sunak | President Emmanuel Macron | PM Narendra Modi |
| Co-chair / co-host | EU | UK | India (Modi) | None |
| Format | In-person, two-day | Day 1 virtual leaders, Day 2 in-person ministerial | One-week programme, two-day high-level segment | Six-day programme |
| Countries participating | 28 plus EU | 27 plus EU at minister level (10 plus EU at leader level) | 100+ countries | 100+ countries |
| Primary declaration | Bletchley Declaration | Seoul Declaration and Ministerial Statement | Statement on Inclusive and Sustainable AI | New Delhi Declaration on AI Impact |
| Declaration signatories | 28 plus EU (29 plus EU after NZ joined Oct 2024) | 10 plus EU (leaders); 27 plus EU (ministers) | 64 plus EU and African Union Commission (66 total) | 88 at adoption; 91 (final count, after Bangladesh, Costa Rica, Guatemala joined) |
| US participation | Signed (Harris) | Signed (Harris) | Refused to sign (Vance) | Signed (Kratsios) |
| China participation | Signed (Vice Minister Wu Zhaohui) | Not invited to leaders' session | Signed (Vice Premier Zhang Guoqing) | Signed |
| UK participation | Host; signed | Co-host; signed | Refused to sign (Starmer government) | Signed |
| Primary focus | Frontier AI catastrophic risk | Operationalising Bletchley with corporate commitments | AI innovation, sustainability, public goods | Inclusive deployment, Global South, sovereign compute |
| Corporate commitments | Eight developers pledged pre-deployment evaluation access | 16 (later 20) signed Frontier AI Safety Commitments | None | 13 signed Frontier AI Impact Commitments |
| Institutional outcomes | UK AI Safety Institute announced | International Network of AI Safety Institutes seeded | Current AI Foundation; InvestAI; Coalition for Sustainable AI | Global AI Impact Commons; sovereign compute expansion |
| Investment announcements | None | None | ~110 billion euro (France); 200 billion euro (InvestAI) | 200+ billion US dollars |
| Binding commitments | None (voluntary) | None (voluntary) | None (voluntary) | None (voluntary) |
The summit series has produced one durable institutional output: the International Network of AI Safety Institutes, sometimes called the AISI network. Seeded by the Seoul Statement of Intent in May 2024 and formally launched in San Francisco in November 2024, the network has expanded over time and remains active as of 2026.
| Member | Country / region | Joined | Notes |
|---|---|---|---|
| AI Security Institute | United Kingdom | November 2023 (founding) | Originally AI Safety Institute; renamed February 2025 |
| AI Safety Institute (later CAISI) | United States | February 2024 (founding) | Reorganised June 2025 as the Center for AI Standards and Innovation |
| AI Safety Institute | Japan | February 2024 | Housed at IPA |
| AI Safety Institute | Singapore | May 2024 | Operates the AI Verify framework |
| AI Safety Institute | Republic of Korea | November 2024 | Announced at Seoul Summit |
| AI Safety Institute | Canada | November 2024 | |
| AI Safety Institute | Australia | November 2024 | |
| AI Safety Office | European Union | June 2024 | Within the European Commission |
| AI Safety Institute | France | November 2024 | INESIA |
| AI Safety Institute | Kenya | November 2024 | First African member |
The network's first joint exercise, in October-November 2024, was a coordinated multilateral testing of frontier models across the founding institutes. At the inaugural convening in San Francisco the network agreed a joint mission statement, more than 11 million US dollars in funding for synthetic content research, and a joint statement on risk assessments of advanced AI systems. The UK and US institutes are the largest members; the AI Security Institute employs more than 100 technical staff, has annual funding of approximately 66 million pounds, and has access to over 1.5 billion pounds in compute through the UK AI Research Resource [10].
The network's future has been complicated by the US reorganisation. In June 2025 Commerce Secretary Howard Lutnick announced that the US AI Safety Institute would be renamed the Center for AI Standards and Innovation, dropping "safety" from its title and refocusing on national security and "pro-innovation" standards. The CAISI continues to participate in the network but under a narrower mandate that emphasises export controls, supply-chain security, and competitive standards rather than catastrophic-risk evaluation.
The AI Safety Summit series represents the first sustained multilateral effort to coordinate international responses to advanced AI risks. Its significance can be assessed across several dimensions.
Before the Bletchley Park Summit, there was no consensus international language for discussing AI risks. The Bletchley Declaration established shared terminology around "frontier AI," "severe risks," and the responsibilities of developers and governments. This common vocabulary has since been adopted in national policy documents, corporate safety frameworks, and academic research worldwide [4]. The Seoul package then linked that vocabulary to specific developer practices through the Frontier AI Safety Commitments. The Paris and New Delhi summits further integrated the language of "public goods," "inclusive AI," and "sovereign compute."
The summits catalysed the creation of new institutions. The UK AI Safety Institute became the first government body dedicated to evaluating frontier AI systems. The International Network of AI Safety Institutes, launched following Seoul, has expanded to include bodies in Japan, Singapore, Canada, France, Kenya, and others. The Current AI Foundation, established at Paris, gives the AI public-goods agenda a dedicated funding vehicle. The Global AI Impact Commons, established at New Delhi, gives the inclusive-deployment agenda a parallel one [21].
A notable trend across the summit series has been the gradual broadening of focus from pure AI safety concerns to wider questions about AI's economic and social impact. The Bletchley Park Summit was narrowly focused on existential and catastrophic risks from frontier AI. By Paris, the agenda had expanded to encompass AI for public good, sustainability, digital divides, and market concentration. The New Delhi summit continued this trajectory with its emphasis on inclusive development and sovereign AI infrastructure [11][17].
The rebranding is itself a signal. "AI Safety Summit" foregrounds risk; "AI Action Summit" foregrounds activity; "AI Impact Summit" foregrounds outcomes. Each shift moves the centre of gravity further from the founding concern. Some observers, particularly in safety-focused civil society and in Anthropic's leadership, have read the trajectory as a quiet retreat. Others, including the French and Indian governments, have framed it as a necessary maturation that brings developing economies and innovation-focused stakeholders into a process previously dominated by frontier-risk specialists.
Reception of the summit series has been mixed across academic, civil society, and industry observers.
The Ada Lovelace Institute argued that Bletchley was a "missed opportunity" to involve civil society at scale and that its narrow focus on frontier risk left documented present harms (algorithmic discrimination, surveillance, labour displacement) outside the agenda. More than 100 civil society organisations and trade unions signed an open letter ahead of Seoul echoing this point. The TUC, the UK trade union confederation, called Bletchley "dominated by Big Tech." Mozilla, Access Now, and Article 19 made similar criticisms in subsequent commentary on Seoul [22].
The Centre for Strategic and International Studies (CSIS) characterised Seoul as a "mini summit" whose principal achievement was operationalising Bletchley by scaling up the AISI network. CSIS warned that without Chinese participation, the series risked sliding from a global to a Western-aligned forum. The Centre for Emerging Technology and Security at the Alan Turing Institute agreed that civil society participation had been "limited" [23].
Reception of Paris was sharper. The European Policy Centre published an essay titled The Paris Summit: Au Revoir, global AI Safety?, arguing that the event had effectively ended the Bletchley settlement and replaced it with a looser industrial-policy frame. Brookings argued that Paris had "lost" the safety thread without replacing it with anything as substantive. Civil society groups noted that the multistakeholder steering committee at Paris had "variable impact and restricted access to the Main Summit," with civil society sometimes physically excluded while major tech CEOs were present [24][25].
Reception of New Delhi was more positive among Global South commentators. Brookings argued that the event had successfully reframed AI governance around development priorities and that the New Delhi Declaration's broad signatory base gave the summit unmatched legitimacy with developing economies. Critics, including the Centre for Strategic Studies on Hate, argued that the emphasis on infrastructure deals and the absence of explicit human-rights protections made it vulnerable to capture by both sovereign-AI nationalism and Big Tech [26][27].
A recurring critique across all four summits has been the gap between voluntary commitments and actual implementation. AILabWatch and METR have produced regular comparative analyses of corporate frameworks, but such third-party tracking is not part of the summit architecture itself [9].
The summit series has faced several other criticisms. All declarations and commitments have been voluntary and non-binding, lacking enforcement mechanisms. The refusal of the US and UK to sign the Paris declaration raised questions about the durability of the consensus achieved at Bletchley. Critics have also noted that the summits have disproportionately featured perspectives from wealthy nations and large technology companies, with historically insufficient representation from the Global South, civil society, and communities most affected by AI deployment. The New Delhi summit partly addressed the geographic bias but the structural dominance of major AI firms in the corporate roundtables has continued.
The scheduling of the Paris summit during Chinese Lunar New Year and the start of Ramadan was widely cited as evidence of insufficient sensitivity to non-Western religious calendars and as a practical barrier to meaningful participation by stakeholders from those communities [25].
The China question has remained unresolved across the series. China was a founding signatory at Bletchley, omitted at Seoul, returned at Paris (signing despite the US-UK absence), and signed at New Delhi. The summit series has not produced the kind of US-China safety dialogue that some at Bletchley imagined, and no Chinese AI lab has signed the Frontier AI Safety Commitments other than Zhipu AI.
Switzerland will host the next summit in Geneva in 2027, formally announced by President Guy Parmelin on 19 February 2026 during the New Delhi event. The Swiss summit is expected to lean heavily on Geneva's existing multilateral infrastructure, including the International Telecommunication Union, the World Intellectual Property Organisation, and the UN Office at Geneva. Swiss officials have framed the summit as an opportunity to bridge the political divides that opened at Paris [20].
The United Arab Emirates will host the 2028 summit. The selection consolidates the Gulf's role in the AI governance conversation, building on the UAE's signature on Bletchley, its corporate participation through G42 and the Technology Innovation Institute, and its substantial private investment in French AI infrastructure announced at Paris.
The United Nations is also planning its first global forum on AI for July 2026, scheduled within the framework of the UN Office for Digital and Emerging Technologies. The relationship between this UN process and the Bletchley series is not formalised; some observers expect the two tracks to merge over time, others expect them to remain parallel.
As of 2026, the AI Safety Summit series has established itself as the primary venue for international AI governance discussions. The series has grown from 28 country signatories at Bletchley Park to 91 at New Delhi, reflecting expanding global engagement with AI governance issues. However, the series faces headwinds. The US withdrawal from the Paris declaration, combined with the Trump administration's deregulatory stance on AI, the rescission of Biden's Executive Order 14110 on the first day of the second Trump term (20 January 2025), the release of Winning the AI Race: America's AI Action Plan on 23 July 2025 (which set out an explicitly innovation-first international diplomacy posture) [30], and the renaming of the US AI Safety Institute to the Center for AI Standards and Innovation, has introduced uncertainty about American participation in future multilateral AI safety efforts. The UK's parallel refusal to sign in Paris further complicated the picture, given that the UK originated the summit series. The decision of the United States to nevertheless sign the New Delhi Declaration (which Trump administration officials publicly characterised as a non-binding statement of cooperation rather than a regulatory commitment) tempered some of those concerns but did not fully resolve them.
The institutional infrastructure created by the summits, particularly the International Network of AI Safety Institutes, continues to operate and expand. The International AI Safety Report published ahead of the Paris summit has become a reference document for policymakers worldwide, and a first key update was released in October 2025. The corporate safety frameworks prompted by the Seoul commitments have introduced a degree of standardisation to how frontier AI companies communicate about risk, even if external verification remains limited.
At the same time, the trajectory from "safety" to "action" to "impact" has unsettled some of the original constituency for the summit series. Whether voluntary, non-binding mechanisms can keep pace with the rapid advancement of AI capabilities, and whether the broadening agenda has come at the expense of frontier-risk focus, remain the central questions facing the summit process. The Geneva summit in 2027 will be the first test of whether a return to a more neutral multilateral setting can re-anchor the conversation, or whether the centrifugal forces visible since Paris will continue to pull it apart.