Responsible AI in business and government

The OECD experience

In September 2025, the OECD published a review of the first reports from companies as part of the Hiroshima Process on AI, a G7 initiative to establish universal rules for the development and use of AI systems. The review examined companies' practices in seven¹ areas:

1) Identification and assessment of² organizational risks. Companies like Microsoft, Google, OpenAI, IBM, and others conduct AI “red teaming” exercises, where teams use simulated attacks to “trick” the model to identify risks of system malfunction.

2) Risk management through procedures and technical measures. First, AI is tested within the company, then for a limited circle of trusted users, and only then is it made widely available. Technical measures are implemented: cleaning and selecting training data, “fine-tuning” the model for specific tasks, and checking the results of queries before they are seen by humans.

3) Disclosure of information about the transparency of AI systems. For consumerfacing AI products, companies publish model “passports” and transparency reports: what the system can do, where its weaknesses lie, and how it was tested. Providers of B2B solutions include disclosure obligations in their contracts.

4) Incident management. There are prescribed scenarios for incidents: who monitors, who records, who responds— companies have teams of specialists. Hotlines are set up for reporting AI problems (at KYP.ai and Rakuten), and special procedures are in place for reviewing the operation of high-risk AI systems.

5) Creation of mechanisms for authenticating and tracking AI content. Companies are implementing labeling of AI- generated content (such as watermarks) and other ways to inform users that they are interacting with AI.

6) investing in AI security. Most companies are investing in cybersecurity, increasing trust in information (for example, Google is developing tools to detect fakes), and identifying discriminatory behavior in AI (for example, Fujitsu and OpenAI are testing AI models for discrimination).

7) the contribution of AI systems to achieving socially important goals. Large companies are conducting research on sustainability, fairness, transparency of models, and confirmation of content origin (Microsoft has an AI & Society network and an AI Frontiers lab). Many projects are focused on healthcare, education, accessibility, and climate, and support the UN SDGs.

The OECD has also published a review³ of the use of AI in public administration. Like companies, public authorities are guided by the OECD Recommendation on AI, the Hiroshima Process rules, etc. For example, Canada has a mandatory algorithm impact assessment for all automated decisions in public administration. Typically, (in 45 out of 200 cases reviewed), AI is implemented into public services. For example, in Greece, AI “reads” and analyzes documents for real estate registration, which has accelerated the assessment of such transactions from several months to 10 minutes. In second place in terms of popularity is AI in open government and interaction between the state and civil society. For example, the European Parliament Archives have implemented AI to assist in searching and analyzing documents from the archives. AI is also often used in judicial process. For example, in Brazil, the Supreme Court has implemented AI for the initial review of applications to determine whether they meet the requirements and whether the content of the applications can be reviewed (e.g., whether a sufficient set of documents has been submitted). As a result, the time required to review applications has been reduced from more than 40 minutes to a few seconds.

Russia’s experience

In Russia, ethical principles for the development and implementation of AI systems are based on “soft” regulation for specific industries. In Monitoring No. 7 (19), we already discussed the Bank of Russia's Code of Ethics for AI in the Financial Market. In addition, the AI Alliance (which includes Yandex, VK, Sber, and others) has developed voluntary codes: a general AI Code of Ethics, a Code of Ethics in Medicine, a Declaration on Generative AI, and others. These documents partially replicate the 2019 OECD Guidelines on AI.

In 2025, the Russian government also initiated an experiment on the use of generative AI in public administration: the Ministry of Digital Development, Communications and Mass Media will develop methodological recommendations and rules for its implementation (including criteria for services and restrictions on application scenarios), with participation envisaged for federal and regional⁴ authorities. It appears that in the next 1-2 years, Russia will move from soft regulation to basic mandatory requirements for AI in public administration and in certain sectors (e.g., transport, finance, medicine). Based on the experiment with generative AI, the Ministry of Digital Development, Communications and Mass Media may develop standard methodologies: testing AI systems before they are released to the market, incident management procedures, and complaint review mechanisms. Industry recommendations may also be introduced (checklists for managing AI system risks, similar to financial services, for example, in the transport sector).

From the monitoring issue No. 9 (21), September 2025. Download the full issue (PDF) · issue page at the Gaidar Institute