Small Index – Great responsibility
OECD worked out an index to compare countries’ measures on AI development
Из выпуска мониторинга No. 2 (26), February 2026 · выпуск целиком, PDF · на сайте Института Гайдара
The main issue for most countries today is how to provide conditions for the AI development, without sacrificing safety. In February 2026, the OECD launched the OECD.AI IndexI, a tool to assess the implementation of an earlier approved Recommendation 2019 on AI.
The index shows how a country is doing in terms of creating a "comfortable environment" for AI development. The index allows countries to assess how well they are doing relative to other countries in developing AI, where there are weaknesses, and which policies should be addressed. In their turn, companies will be able to determine where they are potentially more comfortable designing and training their systems.
The Index currently includes only OECD countries (Russia is not included). The "most advanced" of the 38 countries were: the US and the United Kingdom with the best performance in AI research; the US on infrastructure and policies on AI; Switzerland on employment and skills; and Luxembourg on international cooperation.
Of all venture investments in 2025 ($258.7 bn) invested globally in AI companies
Index figures also include, for example, the number of high-quality AI scientific publications in international databases, the number of patent applications, fiber optic connections, data center clusters with AI accelerators (graphics processors), use of AI by authorities, availability of AI-"sandboxes",1 influx of specialists in the field of AI, etc.
Russia has launched a federal project "Artificial Intelligence," setting goals for scientific publications on AI, the use of AI by government agencies, and participation in international AI standardization groups. By June 2026, the Ministry for Digital Technology must submit a plan for development of data center infrastructure.
In February 2026, the OECD has also published Due Diligence Guide for Responsible AI. These are recommendations for companies on how to identify and prevent risks and negative impacts of AI on employees, consumers, competition, etc.
The OECD cites "bad" company practices, such as monitoring employees and relying on emotion-recognition algorithms to decide who to fire, as examples of how to avoid such behavior.
Suppose a company develops AI for skin cancer diagnosis and buys data from a contractor whose employees manually label patient photos. However, the contractor does not possess measures to protect mental health of employees. For the latter, this means the risks of labor disputes and that data purchasing might be rejected for unethical reasons. For the developer, this also poses a commercial risk: if it turns out that the AI was trained on data obtained in violation of employees’ rights, companies demanding responsible supply chains in AI development may refuse to cooperate with such a supplier.
Another question arises about attraction of investment. Currently, only 38% large technological companies (out of 200) publish principles of responsible AI, and only 10% disclose their internal mechanisms for AI management.
What needs to be done? Start with yourself: assess whether there are practices in the process of creating or operating an AI system that could harm consumers, employees and other stakeholders, distort competition, make the AI's operation opaque, etc. Introduce verification procedures throughout the entire life cycle of AI. If harm cannot be prevented, it is necessary to understand how to mitigate negative consequences and gradually eliminate bad usage practices. Next, work with suppliers and demand to adhere to responsible AI practices.
In Russia, principles of responsible AI are not commonly implemented; there is "soft regulation." The Bank of Russia developed Recommendations on responsible AI in the financial sector, while the Alliance formulated a Code of AI ethics in the AI sphere in 2021.
What is next?
The OECD plans to expand geography of the AI Index, which could include China and Singapore.
Some of the Index's figures can be calculated for Russia as well. For example, the "Graphics Processing Units" indicator is calculated based on Epoch.AI open data . Thus, in Russia, a number of companies, such as Yandex, set up computing centers equipped with graphics processors designed for training and operating AI systems. AI can be trained without such processors, however, using them, training goes faster 10-100 times . Thus, the OECD estimates the development of infrastructure for AI in the country.
If we calculate the cluster score for Russia and the 21 OECD countries possessing confirmed clusters, Russia will be in the 21st place (out of 22 with clusters), ahead of the UK.
- The information refers to official statistics and administrative data, surveys, etc.: for example, data on publications is from OpenAlex and Elsevier/Scopus, on patents is from OECD arrays, on large models and clusters of graphic processors is from Epoch AI, on supercomputers is from the TOP500 list, on venture investments is from Preqin, on specialists - from LinkedIn, on software projects is from GitHub, etc. ↑
- To measure the relative power of these processors in different countries, their computing power has to be compared to the Nvidia H100 processor. The fact is that Nvidia H100 is one of the most common server accelerators for AI training, therefore, capacities of different systems are converted into H100 equivalents so that they can be compared in understandable terms. There's a specific methodology for measuring this indicator. It involves using Epoch AI data on the capacity of confirmed data center clusters commissioned no later than 2024 (converted to H100 equivalent). For example, in Russia, 7 clusters with a total capacity of 1570.338409 NVidia H100 equivalents are included in the calculation. Next, the resulting equivalent must be divided by the population aged 15–64. In Russia, the working-age population, according to the World Bank, is 94.078,090 in 2024: ↑
- 570.338409 ÷ 94 078 090 = 0,0000166919 (around 16.7 NVidia H100 equivalents for 1 million people). Then, it is necessary to perform min-max normalization: take the result obtained for the NVidia H100 equivalent (Russia) and divide it by the difference between the best (according to Epoch AI – Switzerland) and the worst (a number of countries have a zero score) results. The rating ranges from 0 to 1, with Switzerland having the highest score among countries and being assigned a weight of 1. To calculate Russia's weight, follow these steps: 16.7 equivalents ÷ 2 943 equivalents NVidia H100 (by Switzerland result) = 0,0057. ↑
- United Kingdom has also low ranking (10.9 equivalents NVidia H100 for 1 million people): 10,9 ÷ 2 943 = 0,0037. ↑
From the monitoring issue No. 2 (26), February 2026. Download the full issue (PDF) · issue page at the Gaidar Institute
