// wiki · September 2026
The diff.legal wiki
A reference on AI and technology law: short answer-articles to the questions technology companies ask most often — from labelling AI content to checking an AI asset in a deal. General information, not legal opinions; the team assesses a specific task after an enquiry.
This is a project section — structure and depth of the articles under discussion. Twenty-two articles are written in full (marked “for discussion”), the rest are planned. Before publication, facts in each article go through editing and verification against primary sources.
Map of AI regulation in Russia
Layers of requirements: the basic law on supporting AI technologies, sectoral rules, court practice, experimental regimes — and which of them concerns a specific product.
Court practice on AI
Six cases of 2025–2026: hallucinations, trade secrets, authorship, disclosure of AI use. An editorial deep dive by diff.legal.
Glossary of terms
AI system, model, dataset, prompt, RIM, personal data operator — short definitions in one place.
01 / regulation
AI regulation
Public requirements for AI products: what is already mandatory, what is voluntary and what is being prepared. Norms, registers, labelling, liability.
Map of AI regulation in Russia
Regulation is assembled from layers, not one law: the law on supporting AI technologies, sectoral rules, court practice and experiments. We take apart which layer asks about what.
AI regulation abroad
Three models: the EU's horizontal AI Act, China's mandatory double labelling, the fragmented laws of US states. What a product for export should do.
AI content labelling
In Russia labelling of generations is voluntary; in the EU it is mandatory from 2 August 2026. Whom and which mark it concerns.
AI in judicial decisions
What the courts have already answered: liability for AI texts, disclosure under Plenum No. 15, an AI opinion is not evidence (No. 22-1201/2026). A table of positions.
Liability for a model's error
Who answers for harm from an AI system's recommendation: the developer, the owner or the user. The link to the general tort rules.
02 / content rights
Rights to AI content
Who owns what is generated: authorship, prompts, training models on others' works, protection of the results.
The legal nature of AI models and AI products
The model is not named in the law: code is a program, weights are a know-how, a dataset is a database. How protection is assembled and what depends on the qualification.
Who owns an AI generation
The first Russian decision in a dispute over neural network images: the program is not the author, a prompt is not a creative contribution. When protection is still possible and how to secure it.
Training a model on someone else's works
What the Civil Code allows, what the law on supporting AI added, and what to write into the data provenance log so the dataset does not become a problem in a dispute or a deal.
AI hallucinations in court
A 50,000 ₽ fine for invented practice and the CIR's position: “the service selected it” is no excuse. In the review of court practice on AI.
A prompt as an object of rights
Whether a prompt can be protected as a work or a know-how, and what to record so the company's prompt library stays its asset.
03 / data and secrets
Data and secrets
Trade secrets, personal data and their boundaries: what may be passed to AI services and beyond the company's contour.
AI and trade secrets
Uploading reports to a public AI service is disclosure: case No. 02-1545/2026 and four measures the company should formalise.
Training AI on secrets and personal data
Whether a model may be trained on corporate secrets and personal data: the closed contour, grounds under 152-FZ, data under NDA and the model's memorisation risk.
AI at work: what an employee may do
What may and may not be done when using AI at work: case No. 02-1545/2026, disclosure of AI before the court under Plenum No. 15, a “which content into which service” checklist.
Personal data in AI services
152-FZ as applied to models: who is the operator, when consent is needed, what to write in the policy and how to formalise processing in a closed contour.
De-identification of data
When data stops being personal under Roskomnadzor and why removing names does not always save the day.
Cross-border transfer
Localisation and Roskomnadzor's “unfriendly” registers as applied to cloud AI services and teams with relocated staff.
04 / software and licences
Software and licences
Open source in a product, licence maps, program registration, choosing between a know-how and a patent.
Open source in a company product
Permissive and copyleft licences on one page: what may go into a proprietary product, why AGPL is dangerous for SaaS and how to keep a component list.
Vibe coding and lawyers
Code from a dialogue with a model: whose rights to the result, licensing risks of generations, liability for bugs — and what vibe coding gives the lawyer personally.
The software licence map
Commercial, open and proprietary licences: rights, restrictions, typical mistakes in development contracts.
A know-how instead of a patent
What to protect by the trade secret regime and what by a patent: selection criteria for algorithms and models.
Registration of programs with Rospatent
Why register code and databases, what it gives in a dispute and a deal, how long it takes.
05 / deals
Deals and investments
Checking and formalising AI assets in M&A and rounds: rights to the model, datasets, warranties, ESOP tables.
Due diligence of an AI asset
Six blocks of checks when buying an AI company: rights to the model and code, dataset provenance, licences, regulatory requirements, the team, client contracts.
Deals with AI products
Four questions of a deal: the check, forms of transfer (licence, assignment, SaaS, white-label), completeness of the technology, monetisation and taxes.
ESOP and team options
Option plans for Russian companies: instruments, tax consequences, typical paperwork mistakes.
Warranties in technology deals
Representations and indemnities: how warranties on datasets, licences and freedom from claims are drafted.
06 / taxes
Taxes
The tax side of an AI business: the IT regime and rates, VAT on buying foreign services, coefficients for AI expenses, taxes in deals.
The tax regime of an IT company: 2026 rates
Accreditation and the software register as the two keys to the benefits; a 2025→2026 rate table: profit 5%, contributions 15%, the VAT benefit with the general rate at 22%.
Paying for foreign AI services: VAT and expenses
A subscription as an import of an electronic service: agency VAT 22/122, invoices, the unified tax payment, the deduction — and the risks of paying through intermediaries.
AI expenses: coefficients and deductions
How the state shares AI expenses: coefficient 2 for R&D and register software, depreciation of AI assets with coefficient 3, regional deductions.
Taxes when selling an AI asset
Assigning the model versus licensing, selling a share versus selling the business: what is more advantageous and which expenses are capitalised.
Payments to foreign contractors
Personal income tax and contributions under contracts with relocated staff and foreign freelancers, currency control and the risk of employment being found.
Intra-group licences: transfer pricing
Payments for software and models between related companies: transfer pricing and proving the arm's-length nature of royalty rates.
07 / legal work
AI in legal work
Training materials on a lawyer's work with AI — verification protocols, assistants, permissible data — are collected in the diff.academy wiki; legal topics and practice reviews remain here.
AI for lawyers: tasks, agents, chatbots
How the model predicts the next word, RAG, the difference between an agent and a chatbot, gates and the 2026 penalty practice, a chatbot without programming and three data shelves.
The diff.academy wiki
The training base of diff.legal's educational project: the protocol for verifying AI answers, building an assistant for the team, permissible data. Enrollment in the programmes is not open yet.
Verifying AI answers: a lawyer's protocol
Five verification steps, signs of hallucination and recording AI use — inspired by the cases of fines for invented practice.
A lawyer's assistant: where to start
From a recurring task to a training prototype: the role, permissible materials, test assignments and mandatory human verification.
AI in legal departments
Deloitte UK and RSGI data (June 2026): adoption dynamics, budgets and the gap between expectations and the maturity of agent governance.
The future of the legal profession
What compresses, what grows and what does not change: Deloitte/RSGI data, court practice and new skills — from verifying AI answers to managing agents.
Vibe-lawyering: law with AI for a non-specialist
Whether a non-specialist can work with law together with AI: what is permissible to do alone, the three-question rule and when a lawyer is indispensable.
An AI agent as a representative
Whether an agent can make deals and sign documents: a power of attorney, an electronic signature, the limits of authority.
08 / deep dives
Deep dives and reviews
In-depth editorial materials with verifiable references: a review of AI court practice with case numbers, notes on specific rulings and research digests. Updated as new practice arrives.
Digital law in 2026
There is no single definition — there is a test for a special rule. The digital ruble (2026–2028 stages), Federal Law No. 282-FZ on digital currency and digital rights, No. 243-FZ on AI, No. 289-FZ on platforms.
Content on the internet: website, domain, platform
The website as a bundle of rights, domains without exclusive rights (Court for Intellectual Rights practice), the anti-piracy mechanism, marketplaces after Constitutional Court ruling No. 39-P, disputes over trademarks, photos, databases and software.
Court practice on AI in Russia
Six cases of 2025–2026: fines for AI hallucinations, uploads to an AI service as trade-secret disclosure, the status of AI documents, disclosure of AI use. Practice stated as of 21 September 2026.
AI and trade secrets
When uploading data to a public AI service is a disclosure: case No. 02-1545/2026, the logic of the Constitutional Court and four measures a company should formalise.
AI in legal departments
Findings of the Deloitte UK and RSGI study (June 2026): 121 legal-function leaders on AI adoption — dynamics, budgets and the gap between expectations and maturity.
rules
Section rules
- diff.legal prepares and publishes the materials; changes are dated.
- The materials are protected by copyright; cite with attribution to the source and an active link to the page. The rules for AI systems are in llms.txt.
- The facts are verifiable: case numbers, dates and norm references are named; case statuses may change — verify current texts in the court registers and official sources.
- Articles in the “for discussion” status are a draft: wording and conclusions may change after editing.
- Quotations from judicial acts are marked as quotations; a summary does not replace the text of the act.
- Personal data and client projects are not published in the materials.
A question on the topic of a material — hello@diff.legal. The section map is updated as articles come out.