Russia’s Sovereign-AI Law: What Federal Law 243-FZ Reveals About Regulating Foundation Models

A Foundation-Model Law Built to Promote, Not Restrict

While the European Union spent August 2026 switching on binding transparency obligations, Russia enacted a foundation-model law aimed at the opposite goal: accelerating domestic AI development. According to analysis of Draft Law No. 1271570-8, the State Duma adopted Draft Law No. 1271570-8 on July 8, 2026. The law was subsequently signed by the president and takes effect September 1, 2026, with certain provisions deferred to March 1, 2027 and a transition period running until September 1, 2032 for previously operational systems.

The scope is deliberately narrow. The law applies exclusively to “big foundation models” — defined as those containing at least 1 billion parameters — leaving traditional machine learning, computer vision and specialized AI solutions outside its reach. That threshold is a design choice: it targets the general-purpose models that governments worldwide have identified as systemically important, while sparing the far larger population of narrow AI tools from a new compliance layer.

Voluntary Labeling: The Sharpest Contrast With the EU

The single most revealing provision is what the law does not require. Where the EU AI Act’s transparency rules that took effect on August 2, 2026 make labeling of AI-generated content a binding obligation, Russia’s law makes labeling voluntary. Users must merely be “given the opportunity” to include a warning that AI technology was used, with the format left to agreement between the parties. It is a shift from obligation to option — the same subject matter, an opposite legal posture.

For a global content ecosystem, this divergence matters. A multinational deploying the same generative system in Brussels and Moscow now faces mandatory machine-readable labels in one jurisdiction and a purely optional warning in the other. Reporting on the law frames it as designed to boost domestic models in sectors like healthcare and government, and the light-touch transparency regime is consistent with that promotional intent: fewer obligations on developers, more room to scale. The trade-off is that end users receive weaker guarantees about when they are interacting with, or consuming content from, an AI system.

The Copyright Exemption That Favors Domestic Models

The law’s second defining feature is an intellectual-property carve-out for training data. It permits reproduction — via short-term storage in computer memory — of lawfully obtained copies and publicly accessible works, without circumventing technical protections, solely for training sovereign and national big foundation models, with no remuneration required. In plain terms, Russian foundation-model developers get a text-and-data-mining exemption that many jurisdictions are still litigating, and they get it specifically to build models the state designates as strategic.

That designation carries real weight. A “sovereign” model, per the analysis, is one built entirely in Russia with reproducible training and original weights, while a “national” model may include foreign components. Both statuses unlock government support and simplified dataset access, but qualification is conditional: developers must be Russian entities controlling more than 50% of voting rights, with development conducted domestically and processing occurring in Russian-owned data centers. The copyright exemption is thus not a general liberalization — it is an industrial-policy tool that rewards local, state-aligned development while excluding foreign builders.

What This Means for Global AI Compliance Teams

For legal, policy and product teams operating across borders, Russia’s foundation-model law is a data point in a fragmenting global map. The prescriptions below apply to any organization tracking how foundation-model rules diverge.

1. Map jurisdictional divergence on labeling before you ship one global model

Do not assume a single content-labeling policy satisfies every market. The EU now mandates machine-readable labels while Russia makes them optional; other jurisdictions sit in between. Build a per-market labeling matrix into your release process so a product decision made for Brussels is not silently misapplied — or over-applied — elsewhere.

2. Track the parameter threshold, because scope is defined by model size

Russia’s law triggers only at 1 billion parameters, and other regimes use different technical thresholds (compute, capability, risk tier). Maintain a live inventory of which of your models cross which jurisdiction’s definition of “foundation” or “systemic,” and re-run that classification whenever you scale a model up.

3. Treat data-provenance rules as jurisdiction-specific, not universal

Russia’s text-mining exemption applies specifically to state-designated sovereign and national models, not to all developers. Copyright and training-data rules vary sharply by market and by who is building the model. Document the legal basis for your training data per jurisdiction rather than assuming one exemption travels everywhere.

4. Read localization and ownership conditions as market-access gates

Sovereign-model status in Russia requires Russian control, domestic development and in-country processing. These are not neutral technical rules — they are conditions that determine which developers can access the benefits. When entering any market, separate the transparency obligations (which bind everyone) from the promotional benefits (which may be gated to local entities).

The Bigger Picture: Two Philosophies of Foundation-Model Governance

Placed beside the EU’s August 2026 enforcement, Russia’s new law crystallizes a genuine divide in how states are choosing to govern the most capable AI. Brussels leads with binding transparency and the threat of fines; Moscow leads with promotion, voluntary disclosure and a copyright exemption tied to state-designated national champions. Neither model is simply “more” or “less” regulation — they optimize for different ends. The EU prioritizes user protection and market trust; Russia prioritizes rapid, sovereign capability and industrial self-sufficiency, with the “traditional values” alignment its own leadership has publicly required layered on top. For any country still drafting its own foundation-model rules — including across the Global South and emerging markets weighing how to balance innovation against oversight — these two 2026 laws now serve as the reference poles. The practical lesson is that “AI regulation” is not one thing: the same technology can be governed to restrain it or to accelerate it, and where a jurisdiction lands on that spectrum increasingly shapes where models get built and how freely they circulate.

Frequently Asked Questions

Sources & Further Reading

• State Duma adopts draft law on regulation of Artificial Intelligence — Denuo
• Law supporting the development of AI technologies in Russia — President of Russia
• New Russian law aims to boost domestic AI models in healthcare and government — GxP News
• Russia adopts AI regulations, lags behind in the race — bne IntelliNews
• Commission starts enforcing AI Act rules and new transparency requirements on 2 August — European Commission