The Global Landscape of AI Safety Institutes
Authors: Juhi Kore, Ebani Dhawan
This article provides a snapshot of key insights from a forthcoming comprehensive report.
I. Introduction
The recent Paris AI Action Summit, held in February 2025, was a pivotal moment in the international governance of artificial intelligence (AI). This series of summits began with the groundbreaking UK AI Safety Summit at Bletchley Park in November 2023, where the concept of AI Safety Institutes (AISIs) was first introduced and formalised through the Bletchley Declaration, reflecting the global community’s sustained commitment to address AI safety.
As the summit series evolved, France and other international stakeholders appeared to move its focus away from safety. This shift raises important questions: is the safety-focused approach of AISIs still relevant, and if so, why does it continue to matter amid the rapidly evolving AI landscape?
This report aims to provide a comprehensive examination of AI Safety Institutes as a novel governance model. We will explore what these institutes do, present a catalogue of current AISIs worldwide, analyse the challenges and opportunities they face, and examine the recent shift from 'safety' to 'security' in some jurisdictions. As the first global institutional model for AI governance, AISIs represent a significant innovation in how nations approach the complex challenges posed by increasingly capable AI systems.
II. What do AISIs do?
What is an AISI?
An AI Safety Institute is a state-backed specialised entity with three primary objectives: (1) evaluating AI systems, particularly frontier and advanced models; (2) conducting safety research to advance the science of AI risk assessment; and (3) facilitating information exchange among diverse stakeholders including governments, industry, and academia (Araujo et al., 2024). These institutes represent the first global-scale institutional model specifically focused on AI governance, providing technical expertise to inform policy decisions while remaining independent from direct regulatory power. It is important to note AISIs are not the catch-all government agency for any AI-related activity.
First-wave AISIs share several defining characteristics: they are safety-focused, government-affiliated, emphasise technical expertise, and operate as specialised organisations with mission-driven mandates. These institutes typically maintain a startup-like culture, bringing together experts from diverse fields including computer science, policy, and various domains of AI safety (Hobbhahn, 2023).
It is equally important to understand what AISIs are not. They generally lack direct regulatory power (although this varies by jurisdiction) and are not designed to be comprehensive authorities on all AI-related activities within government. According to the International AI Policy Secretariat (IAPS), the core functions of first-wave AISIs can be classified into three main areas: research, international cooperation, and standards development (Fort, 2024).
History of AISIs
The AISI model can be traced back to April 2023, when the United Kingdom established the Foundational Model Taskforce (later renamed the AI Safety Institute). This initiative gained significant momentum when then-Prime Minister Rishi Sunak secured commitments from OpenAI's Sam Altman, DeepMind's Demis Hassabis, and Anthropic's Dario Amodei to provide pre-release access to their frontier AI models for safety evaluation (UK Government, 2023).
The United States followed in November 2023 with the announcement of its own AI Safety Institute, housed within the National Institute of Standards and Technology (NIST). Japan joined this growing network in early 2024, and we are now witnessing a second wave of institutes emerging across nations including India, Canada, Germany, Brazil, Israel, and other jurisdictions. Some countries have established entities with different names but similar functions, such as Singapore's Digital Trust Centre and the European Union's AI Office (Allen & Adamson, 2024).
What Makes AISIs a Novel Governance Model?
AI Safety Institutes represent a significant innovation in technology governance for several key reasons:
Specialised technical capacity within government: AISIs provide governments with in-house technical expertise specifically focused on frontier AI evaluation, addressing the significant knowledge gap that exists between rapidly advancing AI technology and public sector understanding (Variengien & Martinet, 2024).
Hybrid positioning: These institutes operate in a unique space between government, industry, and academia. Unlike traditional regulatory bodies, they focus on scientific evaluation rather than enforcement, allowing for collaborative relationships with AI developers while maintaining independence (OECD, 2024).
Pre-deployment focus: AISIs emphasise proactive evaluation before deployment rather than reactive regulation after problems emerge—a crucial shift for technologies that could pose significant societal risks (Kelly, 2024).
Public interest mandate: While working closely with industry, AISIs maintain a distinct public interest mandate, creating an independent counterweight to private sector safety claims (Yeung, 2024).
International coordination by design: Unlike previous technology governance approaches that began nationally and later attempted international alignment, AISIs were conceived with international collaboration as a fundamental component, recognising AI's inherently global nature (Lang, 2024).
Bridge between technical assessment and policy: AISIs translate complex technical evaluations into insights that can inform policy decisions, addressing a critical gap in effective AI governance (Davies et al., 2024).
This novel institutional approach represents a significant departure from traditional regulatory frameworks and could become a template for governance of other emerging technologies where rapid innovation outpaces conventional regulatory approaches.
What Does Evaluation Mean?
The evaluation functions of AISIs typically encompass several key activities:
Automated capability assessments: Systematic testing of AI models to understand their abilities across various domains, with particular attention to capabilities that might pose safety risks.
Red-teaming: Adversarial testing where experts attempt to make AI systems produce harmful outputs or demonstrate unsafe behaviours.
Human uplift evaluations: Assessing how AI systems might enhance human capabilities, including potential misuse scenarios.
AI agent evaluations: Testing AI systems that can operate with varying degrees of autonomy to understand potential risks from goal-seeking behaviour.
The UK's AISI has emphasised that their evaluations focus on advanced AI safety for the public interest. They maintain independence as evaluators and keep certain methodological details confidential to prevent gaming or manipulation of their assessment procedures. Importantly, AISI evaluations do not aim to provide comprehensive safety certifications or designate any system as definitively "safe." Rather, they contribute to the nascent science of AI safety testing, which lacks established standards of best practice (UK Government, 2024).
However, as Davies, Strait and Birtwistle (2024) from the Ada Lovelace Institute argue, current evaluation practices face significant limitations. They note that existing methods like red-teaming and benchmarking are "easy to manipulate or game by training models on the evaluation dataset," can be strategically chosen to present models in a favourable light, and are "particularly poorly suited to assessing the impacts of AI systems on minoritised communities." Moreover, they highlight that "small changes to an AI product built on a foundation model can cause unpredictable changes in its behaviour," making evaluations of one version potentially irrelevant for later versions.
What is the Role of International Collaboration?
The AI Safety Summit in Seoul, South Korea (May 2024) saw the launch of the International Network of AI Safety Institutes, formalising collaboration between these national entities. This network aims to coordinate research efforts, share evaluation methodologies, and develop common approaches to AI risk assessment (UK Government, 2024).
The network has launched several collaborative projects since its inception:
Joint Evaluation Protocol (JEP): A standardised framework for evaluating frontier AI models
Global AI Incident Database: Collaborative effort to track and analyse AI system failures
Open Safety Benchmarks Initiative: Development of open-source testing suites
Capacity Building Programme: Training programme for establishing new AISIs in developing regions (Allen & Adamson, 2024)
The inaugural meeting of this network was held in November 2024 in San Francisco, bringing together representatives from various national institutes to establish shared priorities and working methods. According to Allen and Adamson (2024), this meeting aimed to "align priority work areas" and begin collaboration on AI safety, with outcomes to be presented at the Paris AI Action Summit in February 2025. However, as elaborated upon later in this report, safety was unfortunately reduced to a side event during this year’s summit.
III. Catalogue of Existing AISIs
Established Institutes (Fully Operational) - These are institutes that are already functioning with defined leadership, funding, and operational frameworks.
Announced/In Development Institutes - These are institutes that have been formally announced but are still in the planning or early implementation stages.
Network Members Without Formal Institutes - These are countries that participate in the international AI safety network but haven't yet established dedicated institutes.
Established Institutes (Fully Operational)
Announced/In Development Institutes
Network Members Without Formal Institutes
IV. Challenges, Opportunities and Evaluating AISIs
Shifting Priorities: From Safety to Security
A significant development in the evolution of AI governance is the UK's rebranding of its AI Safety Institute to the "AI Security Institute" in February 2025. This shift represents more than a mere name change; it signals a fundamental pivot in focus from broader safety concerns to narrower short-term/immediate security priorities.
Speaking at the Munich Security Conference, Technology Secretary Peter Kyle described this rebranding as "the logical next step" in the UK's approach to responsible AI development. The Institute will now concentrate on "serious AI risks with security implications," such as how AI can be used to develop chemical and biological weapons, enable cyber-attacks, and facilitate crimes including fraud and child sexual abuse. Notably, the Institute will explicitly exclude bias and freedom of speech concerns from its remit (Kyle, 2025).
This strategic pivot appears aligned with changing geopolitical priorities, particularly influenced by the United States' hardening stance on AI. At the Paris AI Action Summit, US Vice President JD Vance articulated this position starkly, stating, "The AI future is not going to be won by hand-wringing about safety" (Bristow, 2025). This sentiment reflects a growing tension between security-focused and broader safety-focused approaches to AI governance.
The UK Institute has already made subtle but significant changes to its public-facing materials. References to "societal impacts" have been changed to "societal resilience," while mentions of AI creating "unequal outcomes" and "harming individual welfare" have been removed entirely. "Public accountability" as an evaluation criterion has been replaced with keeping the "public safe and secure" (Bristow, 2025).
These changes have drawn criticism from some quarters. Elizabeth Seger, director of digital policy at think tank Demos, expressed concern that "any attention to bias in AI applications has been explicitly cut out of the new AISI's scope," while Michael Birtwistle of the Ada Lovelace Institute warned this risks leaving "a whole range of harms to people and society unaddressed" (Bristow, 2025). Instead of continuing their intended focus on setting the blueprint for the development of safe and ethical AI, the Paris AI Action Summit saw governments trapped in a supposed binary: innovation or regulation. “Attention on safety, including long-term fears [of] the emerging technology eventually destroying humanity, has given way to short-term needs around galvanising AI for economic gain.”
The pivot has sparked international debate about whether this represents the future direction for other AISIs. The US position appears supportive of the security focus, while EU representatives have emphasised maintaining broader safety considerations. There are concerns about creating a two-tier system where some AISIs focus exclusively on security while others maintain broader safety mandates (McKeon, 2025). The new remit emphasises immediate security concerns, but does not offer room to look towards the long-term impact of AI, which is the primary concern for the public.
This shift raises fundamental questions about the future direction of AI governance globally. Is the original AISI model, with its broader focus on both safety and social impacts, being superseded by a narrower security-focused approach? Does this represent a maturation of AI governance or a concerning narrowing of scope? With major AI labs suggesting AGI is not far off as countries develop their own institutes, governments will need to consider where on this spectrum their own priorities lie.
Tensions Between AISIs and Other Regulatory Approaches
There has been significant debate about whether the AISI model complements or competes with more traditional regulatory approaches like the EU AI Act. Some critics argue that AISIs may be used to delay or replace more stringent regulation, while proponents see them as complementary technical infrastructure that enables more informed policy decisions (Davies et al., 2024). With the rapid ‘start-up’-like nature of AISIs, there may be friction with the deliberative processes and fixed legislative cycles that traditional regulatory approaches undergo.
In addition, AISIs benefit from adaptive, iterative methodologies that allow rapid responses to emerging threats. However, this agility can clash with the need for consistency and long-term regulatory standards, which are essential for providing certainty and legal enforceability.
The EU has attempted to bridge this divide by incorporating its AI Safety Unit within the broader EU AI Office responsible for implementing the AI Act. This creates a direct link between technical evaluation and regulatory enforcement. Other jurisdictions have maintained clearer separation, potentially creating challenges for coordinated governance.
These tensions raise important questions about the future evolution of AI governance models. Will AISIs eventually develop regulatory teeth, or will they remain primarily focused on research and evaluation? How will they interface with emerging regulatory frameworks as they develop novel technical assessments? The resolution of these questions will significantly influence the effectiveness of global AI governance.
Diversity & Leadership
A critical examination of the current landscape of AI Safety Institutes reveals notable gaps in diversity among leadership teams, both in terms of geographical representation and demographic diversity. The concentration of AISIs in Western and high-income Asian nations raises questions about global inclusivity and the universality of safety standards being developed. Additionally, the technical focus of these institutes has resulted in leadership teams heavily weighted toward computer science and technical AI expertise, potentially at the expense of broader disciplinary diversity including social sciences, ethics, and regional expertise.
This limited diversity may impact the effectiveness of AISIs in identifying and addressing the full spectrum of AI risks, as cultural contexts and societal impacts vary significantly across different regions and populations. Addressing these gaps will be crucial to ensuring that safety standards and evaluation methodologies are robust and applicable across diverse global contexts.
Global South Representation in the AISI Network
Beyond Kenya's inclusion in the AISI International Network, there are growing efforts to improve Global South representation. The network's Capacity Building Programme specifically targets establishing mini-AISIs or regional hubs in Africa, Southeast Asia, and Latin America. This initiative aims to provide technical training, resources, and mentorship for emerging institutes (Allen & Adamson, 2024).
Brazil's announcement at the Paris AI Action Summit regarding the establishment of an institute focused on environmental applications represents a promising step toward broader geographical representation. However, critics note that meaningful participation requires both representation and resources. Without adequate funding and technical capacity, Global South institutes risk becoming token members rather than equal partners in shaping global AI governance (Lang, 2024).
Measuring AISI Success
Evaluating the effectiveness of AISIs presents significant challenges, particularly given their relatively recent establishment. Several key metrics might be considered:
Impact on policy and regulations: Are AISI findings and recommendations being incorporated into national and international regulatory frameworks?
Changes in public perception: Has there been a demonstrable increase in public trust regarding AI development and safety measures?
Influence on industry practices: Are there observable changes in how AI companies build their models, particularly regarding the implementation of robust ethical safeguards?
Scientific contributions: Have AISIs advanced the science of AI safety through publication of novel research, evaluation methodologies, or technical tools?
International coordination: How effectively have AISIs collaborated across borders to develop shared standards and approaches?
The International Network is developing shared metrics for assessing AISI effectiveness, and there are discussions about establishing external review mechanisms to ensure AISIs remain accountable to their public interest mandates. These emerging frameworks for evaluating impact could help demonstrate the value of the AISI model and guide future investments (OECD, 2024).
The UK AI Security Institute, which celebrated its first anniversary in June 2024 as the "AI Safety Institute," has made notable progress in building a team of over 30 technical researchers, recruiting an advisory board, defining its approach to evaluating AI systems, releasing an open-source software library to assess specific capabilities of models, and announcing a Silicon Valley office. However, the long-term impact of these activities on broader AI governance remains to be assessed, particularly in light of its recent shift toward security concerns (Davies et al., 2024).
Civil Society Concerns
Various civil society organisations have expressed concerns about AISIs' governance structures and transparency. Some advocate for formal civil society representation in AISI governance structures to ensure broader societal perspectives inform technical evaluation methodologies. There are also calls for all evaluation methodologies developed by AISIs to be open-sourced where possible, enabling external scrutiny and validation (Davies et al., 2024).
These concerns highlight the tension between the technical focus of AISIs and the need for broader societal engagement in AI governance. While technical expertise is crucial for effective evaluation, narrowly defined "safety" or "security" considerations may miss important societal dimensions of AI impacts. Finding ways to integrate diverse perspectives while maintaining technical rigour remains a key challenge.
Funding Challenges and Sustainability
Several AISIs have reported challenges in attracting and retaining technical talent given competition from industry. Governments face difficult trade-offs in allocating sufficient resources to make their institutes effective while balancing other priorities. There are concerns about long-term funding sustainability, particularly for smaller nations' institutes, and questions remain about whether current funding levels are sufficient for the ambitious scope of work (Hobbhahn, 2023).
Talent acquisition remains particularly challenging. With frontier AI companies offering substantially higher compensation packages, AISIs must appeal to researchers' desire for public service and impact. Some institutes have explored novel approaches, including joint appointments with universities, research sabbaticals from industry, and rotating fellowships to attract top talent without competing directly on compensation (Allen & Adamson, 2024).
Long-term sustainability will likely require demonstrating clear value to policymakers while developing diversified funding models that might include industry contributions, research grants, and international support mechanisms for institutes in resource-constrained settings.
Industry Partnerships and Safety Claims
While public-private collaboration is central to the AISI model, managing these relationships without compromising independence presents ongoing challenges. The UK's new partnership with Anthropic demonstrates the evolving nature of these relationships, with a focus on both evaluation and collaborative innovation.
However, there are serious concerns about the gap between AI companies' safety rhetoric and actual practices. Yeung (2024) argues that despite claims about prioritising safety, AI companies often create cultures that "dismiss warnings and hide evidence about unsafe practices, whether to preserve profits, avoid slowing progress, or simply to spare the feelings of leaders." According to whistleblowers at companies like OpenAI and Microsoft, internal safety concerns are frequently suppressed, suggesting that AISIs may need to develop stronger independent verification mechanisms rather than relying on companies' self-reported safety measures.
AI Standards Development
An important area where AISIs can contribute to AI governance is in the development of international standards for AI safety. According to Fort (2024), AISIs are "particularly well-positioned to contribute to the international standard-setting processes for AI safety" due to their in-house technical expertise, mandate for international engagement, and convening power as government-backed institutions.
Standards can be categorised into five types: foundational and terminology standards; process and management standards; measurement standards; product testing and performance standards; and interface and networking standards. Different institutional arrangements may be better suited to developing different types of standards. For example, process and management standards might be more effectively developed by like-minded countries through the AISI network, while interface and networking standards might benefit from the broader participation possible through established bodies like the ISO/IEC JTC 1/SC 42 (Fort, 2024).
As AISIs are uniquely positioned to translate highly technical materials into policy-speak, their intermediary role in the development of standards would ensure that they are both rigorous and implementable. The ability to convene government, international players, and industry in a pre-competitive environment allows a holistic discussion as to what current best practices are and what current public needs are in order to find the common good. These can be tested out in regulatory sandboxes, where standards can be tested and refined before being widely adopted.
Importantly, AISIs can lead the creation of standards that are not static but designed to evolve as AI technologies and associated risks change. By implementing continuous review cycles and rapid feedback loops based on real-world incidents, AISIs ensure that standards remain current and responsive.
Historical Parallels
The emergence of AISIs as specialised technical entities with international coordination can be compared to previous global technology governance efforts, including nuclear safety agencies, climate change monitoring bodies, and internet governance organisations. These historical parallels offer valuable lessons:
Nuclear safety: International coordination through agencies like the International Atomic Energy Agency demonstrates the importance of verification mechanisms and trust-building between nations on matters of safety.
Climate science: The Intergovernmental Panel on Climate Change shows how technical consensus-building can inform policy without direct regulatory authority.
Internet governance: Organisations like ICANN illustrate the challenges of balancing multi-stakeholder governance with national sovereignty concerns.
These examples suggest that successful technical governance depends on balancing rigorous, independent assessment with inclusive stakeholder representation and clear pathways for technical findings to influence policy decisions (Lang, 2024).
V. Comparing the AISI Model and the AI Safety Network Approach
An important consideration in the evolution of AI safety governance is whether the current model of individual national AI Safety Institutes (AISIs) is more effective than a more centralised AI Safety Network approach. This comparison merits examination, particularly in light of the UK's pivot from safety to security.
What are the Differences in the Models?
The AISI Model refers to the current approach where individual nations establish their own dedicated safety institutes with national funding, leadership, and priorities. These institutes then collaborate through voluntary coordination mechanisms.
The AI Safety Network Approach would represent a more centralised model where, rather than building separate national institutes, resources and expertise would be pooled into a unified international entity or a more tightly integrated network with shared governance, similar to organisations like the International Atomic Energy Agency (IAEA).
Strengths of the AISI Model:
National Sovereignty and Control: Individual institutes allow nations to maintain control over their approach to AI safety, ensuring alignment with national priorities and regulatory frameworks.
Contextual Relevance: Each institute can focus on issues most relevant to its national context, economy, and cultural values.
Implementation Speed: Nations can establish institutes without waiting for international consensus, allowing faster response to emerging technologies.
Regulatory Alignment: National institutes can more easily integrate with existing national regulatory structures.
Competition and Innovation: Multiple institutes may drive innovation in evaluation methodologies through healthy competition.
Strengths of the AI Safety Network Approach:
Resource Efficiency: Pooling resources internationally would reduce duplication of efforts and enable more comprehensive evaluations.
Standardisation: A unified network could more effectively establish globally consistent standards and evaluation methodologies.
Equity and Access: A truly international entity could better ensure participation from resource-limited countries.
Independence: An international body might maintain greater independence from any single nation's political priorities.
Bargaining Power: A unified network would have greater leverage when negotiating with global AI developers for access and cooperation.
The Geopolitical Dimension
The UK's rebranding of its institute to focus on security rather than safety highlights a significant challenge for the AISI model: institutes are inherently vulnerable to shifting national priorities and geopolitical pressures. The apparent influence of US perspectives on the UK's repositioning suggests that even established institutes may diverge in their approaches as geopolitical winds change (Bristow, 2025).
This raises questions about whether a more centralised network approach might better insulate AI safety governance from political fluctuations. However, such a network would require broad international buy-in that may be increasingly difficult to secure in a fractured geopolitical landscape.
The current hybrid model—individual AISIs with international coordination through the Network of AI Safety Institutes—attempts to balance these concerns. However, this arrangement faces several challenges:
Coordination Costs: Significant time and resources must be devoted to alignment between institutes.
Uneven Resource Distribution: The concentration of technical expertise and funding in wealthy nations creates imbalances in global representation.
Potential Contradictions: Different institutes may reach conflicting conclusions about the same AI systems.
Divergent Priorities: As demonstrated by the UK's pivot to security, national institutes may evolve in different directions based on domestic political considerations.
VI. Conclusion
Implications for AI Stakeholders
The emergence of AI Safety Institutes represents a significant development in the global governance landscape for artificial intelligence, with distinct implications for various stakeholders:
For the private sector, AISIs offer both opportunities and challenges. Companies developing frontier AI systems must now navigate an additional layer of scrutiny while potentially benefiting from independent validation of their safety claims. The UK's new partnership with Anthropic, Google DeepMind's staff exchange programme with the US AISI, and Microsoft's joint research lab with Japan's AISI demonstrate how these relationships might evolve, with a focus on both evaluation and collaborative innovation.
For policymakers, AISIs provide crucial technical expertise to inform regulatory decisions. The independence of these institutes from direct regulatory authority allows them to focus on scientific and technical assessment without becoming entangled in regulatory politics, potentially enabling more evidence-based policy development. However, the UK's pivot toward security signals that these institutes may increasingly reflect national security priorities.
For academic and research institutions, AISIs represent new partners in advancing the science of AI safety while raising questions about the relationship between public and private research efforts. The emphasis on safety evaluation may influence research priorities and funding landscapes across AI development more broadly.
Future Expansion
Looking ahead, several considerations should guide the potential expansion of the AISI model to additional countries:
Regional representation: The inclusion of Brazil and India, along with Kenya's participation in the AISI Network, represents progress in broadening geographical representation. The Capacity Building Programme's focus on establishing regional hubs in Africa, Southeast Asia, and Latin America could further enhance global inclusivity and bring diverse perspectives into safety evaluation.
Complementary expertise: As seen with Brazil's focus on environmental applications, new institutes might focus on specialised domains where they have particular expertise, such as healthcare AI safety, democracy and information integrity, or particular application contexts relevant to their regional economies.
Coordination capacity: The Joint Evaluation Protocol and other collaborative projects launched by the International Network demonstrate progress in establishing coherent approaches without unnecessary duplication of efforts. These mechanisms will need to scale as the network expands.
Balancing safety and security: New institutes will need to determine where they stand on the spectrum between broad safety concerns and narrower security priorities, an issue highlighted by the UK's recent rebranding and the ensuing international debate.
Evolution of the Governance Model
The shift from safety to security in the UK signals a potential evolution in how frontier AI is governed globally. Countries developing new institutes will need to consider whether to embrace this narrower security focus or maintain a broader safety mandate that includes societal impacts such as bias and fairness.
The most effective approach will likely be an evolved version of the current hybrid model, with strengthened coordination mechanisms that preserve national autonomy while enhancing resource sharing and methodology standardisation. The establishment of shared metrics for assessing AISI effectiveness and discussions about external review mechanisms suggest growing attention to accountability and impact measurement.
Addressing funding challenges and talent acquisition will be crucial for long-term sustainability, particularly for institutes in resource-constrained settings. The AI for Global Challenges initiative's €2.5 billion funding offers promising support, but ensuring equitable distribution and meaningful participation remains an ongoing challenge.
The AI Safety Institute model represents a promising innovation in global technology governance, offering technical depth while maintaining independence from direct regulatory authority. Its continued evolution—now challenged by the UK's pivot to security—will play a crucial role in shaping how societies worldwide manage the unprecedented opportunities and challenges presented by increasingly capable AI systems. The success of this model ultimately depends on balancing technical rigour with inclusive representation, scientific independence with policy relevance, and security priorities with broader societal concerns as we navigate the frontier of artificial intelligence development.
VII. References
Allen, G. C., & Adamson, G. (2024). The AI Safety Institute International Network: Next Steps and Recommendations. Center for Strategic and International Studies (CSIS). https://www.csis.org/analysis/ai-safety-institute-international-network-next-steps-and-recommendations
Araujo, R., Fort, K., & Guest, O. (2024). Understanding AI Safety Institutes: Core Characteristics, Functions, and Challenges. Institute for AI Policy and Strategy.
https://static1.squarespace.com/static/64edf8e7f2b10d716b5ba0e1/t/6705a4a7fd9bf94a706d10ce/1728423090604/IAPS+-+Understanding+the+First+Wave+of+AI+Safety+Institutes.pdf
Bristow, T. (2025, February 14). Britain dances to JD Vance's tune as it renames AI institute. Politico. https://www.politico.eu/article/jd-vance-britain-ai-safety-institute-aisi-security/
Davies, M., Strait, A., & Birtwistle, M. (2024, May 17). Safety first? Reimagining the role of the UK AI Safety Institute in a wider UK governance framework. Ada Lovelace Institute. https://www.adalovelaceinstitute.org/blog/safety-first/
Fort, K. (2024, September 17). The Role of AI Safety Institutes in Contributing to International Standards for Frontier AI Safety. arXiv. https://arxiv.org/html/2409.11314v1
Hobbhahn, M. (2023, September 21). There should be more AI safety orgs. LessWrong. https://www.lesswrong.com/posts/MhudbfBNQcMxBBvj8/there-should-be-more-ai-safety-orgs
Kelly, E. (2024, July 26). 5 questions for the AI Safety Institute's Elizabeth Kelly. Politico. https://www.politico.com/newsletters/digital-future-daily/2024/07/26/5-questions-for-the-ai-safety-institutes-elizabeth-kelly-00171436
Kyle, P. (2025, February 14). Tackling AI security risks to unleash growth and deliver Plan for Change. GOV.UK. https://www.gov.uk/government/news/tackling-ai-security-risks-to-unleash-growth-and-deliver-plan-for-change
Lang, C. (2024, June 21). Advancing AI safety requires international collaboration. Here's what should happen next. Atlantic Council. https://www.atlanticcouncil.org/blogs/new-atlanticist/advancing-ai-safety-requires-international-collaboration-heres-what-should-happen-next/
McKeon, C. (2025, February 14). Rebranded AI Security Institute to drop focus on bias and free speech. The Independent. https://www.independent.co.uk/news/uk/politics/peter-kyle-keir-starmer-rishi-sunak-government-san-francisco-b2698033.html
OECD. (2024, July 29). AI Safety Institutes: Can countries meet the challenge? The AI Wonk. https://oecd.ai/en/wonk/ai-safety-institutes-challenge
UK Government. (2023, November 1). UK AI Safety Summit 2023: The Bletchley Declaration. GOV.UK. https://www.gov.uk/government/publications/ai-safety-summit-2023-the-bletchley-declaration
UK Government. (2024, May 21). Seoul Statement of Intent toward International Cooperation on AI Safety Science, AI Seoul Summit 2024 (Annex). GOV.UK. https://www.gov.uk/government/publications/seoul-declaration-for-safe-innovative-and-inclusive-ai-ai-seoul-summit-2024/seoul-statement-of-intent-toward-international-cooperation-on-ai-safety-science-ai-seoul-summit-2024-annex
Variengien, A., & Martinet, C. (2024, July 29). AI Safety Institutes: Can countries meet the challenge? OECD.AI. https://oecd.ai/en/wonk/ai-safety-institutes-challenge
Yeung, D. (2024, July 9). AI Companies Say Safety Is a Priority. It's Not. RAND Corporation. https://www.rand.org/blog/2024/07/ai-companies-say-safety-is-a-priority-its-not.html
Would you like to be involved in writing our forthcoming report on AISIs?
Please fill out the form below to express interest in contributing to our forthcoming report on AISIs:
