The growing use of artificial intelligence (AI) opens up both new opportunities and challenges in various fields. According to the Report of the International Economic Forum 2024, the negative consequences of AI could be one of the most serious global risks in the long run. In this regard, the governments of leading countries are actively considering the prospect of forming official policies in the field of artificial intelligence. 

One of the potential measures for managing AI risks in international practice is AI Safety Institutes, created to test and evaluate the latest systems. If possible challenges are taken into account and implemented correctly, such institutes can play a leading role in the management of AI technologies, being especially important in the Ukrainian context.

Tasks and Activities of AI Safety Institutes 

The first AI Safety Institutes were established by the UK and US Governments following the AI Safety Summit held in 2023. 

The UK AI Safety Institute (AISI) has become an independent governmental body and the successor to the UK's first AI safety taskforce, the Frontier AI Taskforce. When announcing the creation of the Institute, UK Secretary of State for Science, Innovation and Technology Michelle Donelan called the Institute an "international standard setter" that will "help policymakers around the world to manage the risks associated with cutting-edge AI capabilities so that we can maximize the enormous benefits".

The main goal of the AI Safety Institute, according to the British Government, is to minimize the unexpected and negative consequences of using these technologies for "UK and humanity". Although the Institute's mandate does not include government regulation of AI, it aims at raising public awareness of the risks of the technology based on the belief that governments are the guarantor of accountability of AI systems and support research in this area. 

The British AI Safety Institute primarily aims at fulfilling the following three functions:

1. Conducting evaluations of advanced artificial intelligence technologies. Representatives of the British Department of Science, Innovation and Technology note that AI companies are responsible for testing their systems for risks, but the Government should play a "key role" in conducting external evaluations "independent of commercial pressure". In particular, in May of this year, the British Government announced the launch of Inspect, an open-source AI testing platform that allows evaluating systems from various fields. The key areas of evaluation identified in the UK are the possibility of malicious interference, the impact on society, system security and the availability of controls. 
2. Facilitation of AI research. Technology research activities will include the development of tools for managing artificial intelligence systems, improving the methodology for evaluating AI and conducting research on the safer use of technology.
3. Information exchange. The Institute plans to promote the dissemination of research results and AI safety evaluations between industry, academia and the public.

In addition, the British Institute envisages cooperation with leading AI developers and companies that evaluate AI systems. 

In the USA, the AI Safety Institute was established based on the National Institute of Standards and Technology (NIST). The Institute will use NIST's previous work on standards for AI technologies and create new AI safety projects based on the body's key goals. 

The National Institute of Standards and Technology in the USA defines the main strategic goals of the AI Safety Institute. 

First, the Institute aims at improving AI safety through research, testing and system evaluation. To achieve this goal, the Institute plans to create auxiliary tools and guidelines for the safe use of artificial intelligence, test AI models before operation and check existing systems if a number of risks are identified. 

Secondly, the Institute will develop and disseminate the practice of safe AI application. This includes creating and publishing metrics, tools and methods for evaluating the risks of AI systems and disseminating recommendations for responsible design, development and use of advanced AI models.

In addition, the Institute's activities will focus on supporting institutes, communities and initiatives aimed at improving AI safety. The support measures include both facilitating a dialogue between government developers, experts and users and developing international cooperation in the area of AI safety.

The AI Safety Institute Consortium was created to ensure effective cooperation between the US Government and the private sector. According to NIST, the association consists of more than 200 organizations that will work with the US Government to develop AI governance principles. The Consortium already includes such leading companies as OpenAI, Google Alphabet, Anthropic and Microsoft.

International Cooperation

Today, the Artificial Intelligence Safety Institutes established in the USA and UK are the only organizations of their kind on a global scale. However, it is expected that the number of institutes will continue to grow. At the AI Safety Summit held in May of this year, 10 leading countries and the EU signed an agreement to establish a network of AI Safety Institutes to ensure compatibility of approaches to AI testing and evaluation. Japan, South Korea and Canada have also announced their intentions to establish institutes.

In addition, the institutes are already starting to cooperate to conduct joint research and evaluations of AI safety. This year, in April, the USA and UK signed a Memorandum of Understanding that provides for the joint development of tests and evaluations of advanced artificial intelligence systems. According to the agreement, governments will be obliged to develop a joint work program and approach to model evaluation, including the use of a common methodology and tools.

"This partnership will accelerate the work of both our institutes across the full spectrum of risks to both our national security and our wider community. Our partnership makes it clear that we are not running away from these problems – we are running after them. Through our collaboration, our institutes will gain a better understanding of artificial intelligence systems, conduct more robust evaluations and provide more rigorous guidance", said U.S. Secretary of Commerce Gina Raimondo.

At the same time, international cooperation is not limited to safety institutes. Cooperation initiatives are also emerging at the level of other institutes that will be involved in AI safety research. For example, the US AI Safety Institute will work on tools for evaluating AI models together with the European Union's AI Office.

For its part, the AI Office will become the European body responsible for compliance with the EU AI Act. In addition to regulating and coordinating AI policy, one of the priority areas of the Office's activities will be to improve the safety of artificial intelligence, including the development of methodology and tools for technology evaluation. The AI Office is expected to serve as an AI safety institute for the EU.

Margarethe Vestager, Executive Vice President of the European Commission, considers the mutual commitments between the EU and the US "more ambitious" than those of the Memorandum of Understanding between the USA and UK. "It's not just testing, it's everything related to it: benchmarks, methodologies, understanding and interpretation of regulatory approaches in both jurisdictions", she said.

Prospects and Challenges

The activities of AI safety institutes can have a significant impact on further research on artificial intelligence safety. In particular, international cooperation based on the institutes can deepen a common understanding of the benefits and risks of AI, facilitate the exchange of information and experience and accelerate the development of international standards and approaches to technology testing. 

AI safety institutes are also gaining support among government representatives. "Safety fosters innovation, so it is paramount that we get this right and that we do so in concert with our partners around the world to ensure the rules of the road on AI are written by societies that uphold human rights, safety and trust", said Gina Raimondo.

In addition to the significant advantages of creating and operating such initiatives, it is worth considering the challenges that may hinder the development of safety institutes.

First, current approaches to AI safety testing run the risk of being ineffective or insufficient to determine whether these technologies are safe for society. Experts from the UK Ada Lovelace Institute emphasize that existing methods for evaluating AI safety institutes, such as red teams and benchmarking, may have a number of limitations when it comes to AI safety research. Such methods can prioritize some types of risks over others, which significantly reduces the effectiveness of testing. In addition, experts believe that it is important that AI technologies should be evaluated not in a vacuum but in the environment for which they are intended. Testing AI without taking into account the context does not allow drawing sufficiently reasonable conclusions about the safety of using AI in a certain environment.

Second, as AI technologies are evolving very rapidly, existing approaches to evaluation and testing may become outdated in the long run. Representatives of the UK AI Safety Institute emphasized the impossibility of testing all available models, instead focusing on advanced AI technologies that may have "potentially harmful" capabilities. However, even previous studies by the UK Department of Science, Innovation and Technology have noted that it is currently impossible to reliably predict what new capabilities an advanced AI model can gain in the course of further training. For example, when evaluating large language models, such as Claude or GPT-4, researchers from Model Evaluation and Threat Research (METR) noted that it is possible to test only "a small part of what is on the surface" of such systems. The dynamics of advanced AI technologies emphasizes that current evaluations may be vulnerable to technological changes.

Third, the current approaches of safety institues mostly focus on the technical aspects of AI safety. Previous research shows that most AI safety evaluations focus on the model's capabilities, while evaluations of the social impact of AI are rare. This can create gaps in understanding AI safety, as a purely technical focus is insufficient to identify the risk and negative impact of technology on society. The situation emphasizes the need for interdisciplinary expertise and the implementation of approaches that focus on the sociotechnical, ethical and legal aspects of AI safety.

Possibility of Establishing an AI Safety Institute in Ukraine

As for Ukraine, there are currently no plans to create a corresponding Institute.

According to the "White Paper on AI Regulation in Ukraine" prepared by the Ministry of Digital Transformation, Ukraine is considering the establishment of an authorized executive body as an institutional mechanism for regulating this area. This body will be vested with control and regulatory functions, since, as the book says, "without the creation of a regulatory body... we risk not only getting... situations with multiple and repeated violations of a particular human right, but also risks with a much greater and deeper negative impact and consequences".

The need to create a regulatory body is obvious. At the same time, the Institute, by its nature, is an institution that is called upon to summarize current practices and develop proposals for AI policy that will be subsequently implemented as part of specific solutions. On the one hand, it can become an analytical and discussion platform for various stakeholders, and on the other hand, it can develop institutional cooperation with foreign institutes and centers.

The format of the Institute would be preferable to the practice of creating Working Groups, which depend on the level of interest and involvement of participants, political decisions of the Government and the Ministry of Digital Transformation and are not permanent. The creation of such an institute would ensure the sustainability and continuity of the work on developing AI regulatory standards and would be in line with international best practices.

In addition, the Ministry of Digital Transformation itself writes in its White Paper that due to a lack of resources, "in times of war, despite the importance of the AI sphere, it is obviously extremely unlikely". In fact, this postpones the creation of a regulatory body until the war is over and Ukraine introduces legislation similar to the European AI Act.

This situation creates additional arguments for the creation of the Institute, based on the model already in place in the USA, UK and other countries. In terms of resources, for the functioning of such an intellectual and analytical center, rather than a government body, we can try to attract donor funds for the period of war, which will not require additional budgetary resources.

This Institute could be entrusted with both some of the decisions envisaged for the interim period in the White Paper, such as developing draft methodologies or coordinating cooperation with civil society, and the functions inherent in the Institutes currently being created abroad, in particular:

• development of a common methodology and tools for evaluating and researching AI safety;
• testing AI models;
• participation in the development of international standards and approaches for further AI safety research.

Conclusions and Recommendations

AI Safety Institutes are an important step towards a better understanding of AI safety. Such initiatives, especially at the international level, can further play a key role in researching and developing strategies to address potential risks associated with artificial intelligence, promoting the safe and responsible deployment of new technologies.

At the same time, the effectiveness and efficiency of safety institutes will depend on the competence and relevance of the approaches used to study AI safety. The current activities of the institutes, while contributing to building better standards for AI safety, require significant improvements in model evaluation and testing. To improve the capacity of institutes, it is necessary to:

1. Improve the methodology of AI safety evaluation and testing. Measures that include evaluation and testing of artificial intelligence should be aimed at improving the understanding of the context in which certain technologies are used and take into account the rapid development of the AI industry. For this purpose, it is important to develop a context-oriented approach with the involvement of industry regulators to improve the understanding of AI technologies in different environments.
2. Introduce approaches that take into account the broader social aspects and risks of AI technologies. AI safety research should take into account ethical risks, negative consequences for society and potential biases in AI systems. It is important to improve AI safety research by developing a broader evaluation methodology and engaging experts and stakeholders from various industries.

In the context of Ukraine, it is important to consider the possibility of establishing a Ukrainian AI Safety institute following the example of existing institutes. The main advantage of creating such an institute in Ukraine is cooperation with international partners to share best practices and ensure a comprehensive approach to AI safety evaluation and research. With limited resources and the impossibility of creating a separate regulatory body, the institute could provide the necessary standards and expertise on AI safety.