// wiki · analysis · September 2026
The legal nature of AI models and AI products
Draft article. The qualifications are as of September 2026; doctrine and practice on the status of models continue to form.
A draft for discussion. The law does not name a model as an independent object of rights — the article shows which known regimes it is assembled from in practice and what depends on that.
The short answer
The Civil Code does not single out an “artificial intelligence model” as an object of rights. In practice the product is assembled from known regimes: the code is a program, the dataset is a database, the weights and architecture are most often a know-how under the trade secret regime. Qualification is not theory: the ability to ban the use of a copy, the tax regime, the procedure for transfer in a deal and the investor's check of the asset all depend on it.
What an AI product consists of
| Layer | Legal regime | What it gives |
|---|---|---|
| Training and inference code | A computer program (Art. 1261 of the Civil Code); registration with Rospatent and the software register | Copyright protection, evidence, the register's VAT benefit |
| Architecture and trained weights | Not definitively determined; the working regime is a know-how (Art. 1466 of the Civil Code) under trade secret | Protection against an employee or contractor carrying the weights out |
| Dataset | A database: the maker's related right (Arts. 1333–1334 of the Civil Code) + rights to the elements | Control over extractions, the base for the model's next versions |
| Prompts, instructions, pipelines | Know-how; partly works | Preserving the team's knowledge through turnover |
| The product as a whole (service, API) | A complex; the user is granted access under a contract | Monetisation without transferring the weights themselves |
What the court said
In case No. 02-4220/2025 the court described the neural network as a “computer program” and its result as created “by means of technical devices”. For a dispute about authorship of generations this passed painlessly, but it shows the risk: if the status of the weights is not formalised, the court has nothing to protect except the code. Details — in the article “Who owns an AI generation”.
The working stack of protection
- The trade secret regime on weights and datasets: a list of protected information, need-to-know access, a marking, logs — Article 10 of Law No. 98-FZ; without a regime there is nothing to protect.
- The contractual chain: employees' rights (works made for hire, Art. 1370 of the Civil Code), contractors' (Art. 1296), component licences — all layers must converge on the company.
- Registration of the program and the database with Rospatent and inclusion in the unified register of Russian software: evidence in a dispute plus tax consequences (IT regime rates).
- A patent — surgically: an algorithm “as such” is not protected, but a technical solution implemented by the system can be an invention; for a startup this is usually more expensive and slower than the secrecy regime.
What depends on it
- Deals: transferring “the model by an acceptance act” without describing the layers (weights, code, dataset, documentation) leaves the buyer a box without a key — see “Deals with AI products”.
- Taxes: a register licence and ordinary services differ in rates — the product's qualification directly changes the economics (the tax regime of an IT company).
- Investments: the investor checks the chain layer by layer — a hole in any of them is a red flag (due diligence of an AI asset).
Matching protection to a specific product and formalising the chain of rights is the task of technology rights practice.