// wiki · navigator · September 2026

Map of AI regulation in Russia

Draft article. Facts and references — as of September 2026; they are verified against official texts before publication.

Four layers of requirements

Russia has no single “AI law” that would answer every product question in one text. Regulation is built from four layers, and for each product function a different layer turns out to be the main one:

The basic law on supporting AI

In summer 2026 the federal law on supporting the development of artificial intelligence technologies was adopted — Federal Law No. 243-FZ of 26.07.2026 (introduced as draft law No. 1271570-8; its core provisions take effect on 1 September 2026). What matters for product teams:

• Labelling of generations is voluntary. There are no mandatory requirements for users: the decision to label audio, photo and video content is made by the user. Amendments on mandatory labelling were rejected by the profile committee on 10 June 2026.
• The platforms' duty. Platforms built around large foundation models, together with social networks, must give users a tool for applying the label.
• Training on works. Without rightsholders' consent, developers are allowed to work with works — extraction, comparison and analysis — under three conditions: the copy was obtained legally, the material is in open access, and the result is needed only to train sovereign AI systems (public discussion cited large Russian models as examples). Details — in the article “Training a model on someone else's works”.

The format and procedure of labelling (where it is applied) are set by the agreement between the model provider and the user — so the rules must be looked for in the service terms.

Sectoral rules: where AI content meets advertising and data

Even the voluntary nature of generation labelling does not cancel sectoral rules. Most often three apply:

• Advertising. If AI content is internet advertising, ad labelling applies: tokens, transferring data to ERIR through an advertising data operator. The “generated by AI” mark is a second layer here, not a replacement.
• Personal data. Processing data in an AI service is subject to 152-FZ: purposes, grounds, localisation, cross-border transfer. Passing client data to a public service can also breach a secrecy regime — see “AI and trade secrets”.
• Copyright. Rights to generations and to model training are a separate practice topic: “Who owns an AI generation”.

Court practice as a source of rules

Where the law is silent, the courts are already shaping the rules. Four reference points of 2025–2026:

• Plenum of the Supreme Court No. 15 of 21.05.2026, para. 42 — a party to proceedings must inform the court about the use of AI in preparing documents.
• Case No. А27-7831/2025 — a 50,000 ₽ fine under part 5 of Article 119 of the Arbitrazh Procedure Code for references to non-existent practice “selected” by AI.
• Case No. 02-1545/2026 — uploading information to a public AI service was recognised as disclosure of a trade secret.
• Case No. 02-4220/2025 — the first dispute over rights to neural network images: without a human's creative contribution there is no protection.

All six cases analysed — in the review “Court practice on AI in Russia”.

AI in the work of the courts themselves

Courts do not only create practice about AI — they are themselves becoming operators of AI tools. Order of the Chairman of the Supreme Court No. 59-P of 03.09.2026 approved the Concept for the introduction of artificial intelligence technologies in court proceedings:

• What AI will do. Transcribing hearings (speech recognition with speaker identification, real-time translation), preparing drafts of procedural documents — the structure and text of a draft ruling, individual blocks built from case materials and audio recordings — and real-time hints of relevant norms and case law for the judge.
• Targets for 2030. Over 95% of judges are regular users of judicial AI services; 30–50% of procedural document drafts are generated by AI; at least 80% of drafts are approved by judges without substantive edits; the hearing time for routine cases shrinks by 20–30%.
• The key principle. AI is an aid to the judge, not a replacement: the final decision always stays with a human. The first risk in the Concept’s own risk matrix is model hallucination; the safeguards are generation only with references to the source (RAG architecture), verification of the result by the judge, and a ban on public AI services. The duty to check generated output now applies not only to the parties (Plenum of the Supreme Court No. 15) but to the judiciary itself.

Implementation is overseen by the Centre for Judicial Competencies in AI — an expert group under the Supreme Court.

Experimental regimes

Federal Law No. 123-FZ (2020) introduced experimental legal regimes for AI in Moscow: testing products without complying with certain requirements under the regulator's supervision. For a startup this is a way to check a product in a “sandbox” before a full market launch; the regimes do not exempt from the basic rules on data and secrets.

Content export: the EU

If the content reaches the European Union market, the European layer applies: from 2 August 2026 Article 50 of the AI Act requires marking content generated or altered by AI. For Russian products with a foreign audience labelling is mandatory, not voluntary — both jurisdictions must be checked separately.

What to check in your product

• System category. Are you a model developer, a platform, or only a user of a service? Which duties of the basic law concern you at all depends on this.
• Service terms. The labelling format is set by the agreement with the model provider — read the section on generations before launching a feature.
• Advertising footprint. If generations take part in promotion — set up ad labelling and transfer to an ad data operator right away, not after a regulator's letter.
• An AI policy. Which tools are allowed, what data may be passed into them, who answers — see the measures in the note “AI and trade secrets”.
• Evidence base. If AI takes part in preparing documents for a court or the state — record it and verify the result: responsibility always lies with the person, not the model.

Analysing the regulatory requirements of a specific product is part of product rights and risks practice; the map in this section is updated as the norms change.