Responsible AI governance

The UK experience

In July 2025, a report was released on¹ practices for countering deepfakes including methods for labeling content so that users can recognize and flag deepfakes: invisible² watermarks file origin metadata, content labels (“made with AI”), etc.

Rather than shifting the burden of recognition onto the audience, platforms should use watermarks and metadata to prioritize their moderation efforts. It is recommended to explicitly establish distinctions between fully and partially generated content, where AI has only edited the original content (applied filters, changed elements in the image, etc.).

The EU experience

In July 2025, the European Commission published a Code of Practice for General³ Purpose AI. The Code provides for:

1) Transparency. The Code contains “model documentation” — a universal reporting form that companies can use to disclose information about AI: what the system does, what it was trained on, how many resources it consumes.

2) Safety and systemic risks (for models⁴ with systemic risk. Responsibility may lie with company management, teams that develop and maintain AI, and auditors. Shifting responsibility to the user is not advised.

3) Protection of intellectual property rights. For example, it is recommended to implement rules for “crawling” — the automatic collection of data for training. Data in this process should only be collected on a legal basis. General-purpose AI systems should recognize prohibitions on the use of content for training.⁵

Meta has refused to join. However, OpenAI, Amazon, Google, IBM, and others (more than 25 companies) have announced their⁶ intention to join the Code.

The US experience

In July 2025, the American AI Plan was⁷ released. The plan outlines measures to:

1) Support startups developing open AI models and models with open “weights”. A National AI Research Resource is being created to give startups access to computing power, models, and data without expensive contracts with private companies.

2) Combating deepfakes in the judicial system. False videos created by AI can end up in court as fake evidence and deprive people of their right to justice. Therefore, it is proposed to develop mandatory standards for the recognition of deepfakes in courts.

3) State testing of foreign models for “propaganda and bookmarks.” The Ministry of Industry and Trade will check foreign AI models for censorship, the ability to secretly transfer data or control the AI system without the user's knowledge, and threats to critical infrastructure. At the moment, the US is the only country where such rules are planned to be introduced.

Russia’s experience

In July 2025, the Bank of Russia published a Code of Ethics for AI in the financial⁸ market.

The Code identifies five guiding principles: human-centricity, fairness, transparency, security, and responsible risk management. For example, customers must be given the right to refuse to communicate with a bot and request that decisions made by AI be reviewed by a human (e.g., in the event of a loan refusal). To prevent bias, AI algorithms must exclude nationality and religion from customer assessments, and their data sets must be proactively screened for these attributes. Companies will be required to indicate that recommendations made by robot advisors are generated by AI. The measures provided for in the Code should make AI more transparent and understandable to bank customers and increase public confidence in AI technology as a whole.

From the monitoring issue No. 7 (19), July 2025. Download the full issue (PDF) · issue page at the Gaidar Institute