AI and IP study note · 2026
Whose Model Is It After Fine-Tuning?
A claim that a company owns 'the model' hides a stack of assets, licences, trade secrets, data rights and unresolved legal questions.
- Reading time
- 3 minutes
- Published through
- LinkedIn and Astana Hub
- Topic
- AI Governance and IP
- Series
- AI IP architecture
About this edition. This English site edition combines the published research note and the deeper Russian Astana Hub article. Sources were checked for the original publication in August 2026; model terms and legal positions can change.
Imagine a team one week before an investment round. The product works: an open model fine-tuned on the company's support data resolves requests faster than people. The deck contains one confident line: "Proprietary AI model".
The investor asks one question: what exactly is yours?
The pause does not necessarily mean the team is hiding anything. The word model compresses a stack of very different assets, and the rights around each layer operate differently.
Treat the model as a stack of assets
Public descriptions of major AI partnerships separate research methods, architecture, weights, inference code and fine-tuning code. WIPO guidance likewise describes a combination of copyright, patents, trade secrets, contracts and licences.
| Layer | Examples | Typical protection or control |
|---|---|---|
| Architecture and methods | Network design, training and inference methods | Patent where available, secrecy or deliberate publication |
| Source code | Training, inference, SDK and safety code | Copyright, trade secret and software licences |
| Weights and checkpoints | Parameters, states and adapters | Secrecy, access control and contract; copyright status remains uncertain |
| Training recipe | Data mix, hyperparameters and synthetic generation | Primarily secrecy and operational control |
| Data and evaluations | Training sets, annotations, benchmarks and red-team sets | Contracts, licences, provenance and secrecy |
| Prompts and alignment | System prompts, reward models, policies and skills | Copyright where applicable, licences and secrecy |
| Infrastructure and brand | Serving, APIs, protocols, names and compatibility marks | Copyright, patents, open licences and trade marks |
The practical rule is uncomfortable: a company cannot own the model more completely than it controls the rights and restrictions around each material layer.
Do not make copyright carry the whole structure
Copyright protects particular expression such as code, text and documentation. It does not automatically protect a method or settle the status of model weights.
For many closed providers, practical control over weights depends on secrecy, access restriction, contractual limits and an API architecture that delivers outputs without delivering the core files or training recipe.
The most valuable file can also be the layer with the least settled copyright answer. The operating perimeter matters because doctrine alone may not carry it.
Read the licence underneath the fine-tune
A fine-tuned model contains layered rights. Base weights remain subject to the upstream licence. New adapters or delta weights may be created by the company. Data needs its own authority. Names may be constrained by trade marks or licence conditions.
Licences can differ on commercial use, derivatives, naming, acceptable use, distillation and the use of outputs for training another model.
"We fine-tuned it, therefore it is ours" does not survive a careful reading of the dependency chain.
Do not treat open weights as one legal regime
Several different models are often described simply as open:
- a proprietary model available only through an API;
- an open-weight model with downloadable weights but incomplete code or data information;
- source-available code without the freedoms of an open licence;
- open-source AI with the materials required to study, use, modify and share it under the applicable definition;
- an open standard or protocol around implementations that may remain proprietary.
"Our model is open" is not useful until the team names what is open, under which licence and with which restrictions.
Collect five answers before due diligence
- Whose architecture and methods are used?
- Who owns the training and inference code?
- Whose weights are present, and what does the base-model licence require?
- Whose data was used, and on what authority?
- Who owns the adapters and fine-tuned layers, and are they treated as derivatives under the upstream terms?
A rights matrix is normal for a modern AI product. The risk is not that several parties and regimes appear in it. The risk is that the matrix is being assembled for the first time in someone else's meeting.
The "Proprietary AI model" slide may remain. The difference is that a traceable asset and rights map now sits behind it.