Prompts, datasets and fine-tuned models: what is protectable as an IP asset?

Prompts, datasets and fine-tuned models: what is protectable as an IP asset?

by

Are prompts protected by copyright? Only when the prompt is itself an original piece of writing, and even then that protection does not make you the author of what the model generates. Datasets are protected through database rights over their structure or the investment behind them, not over the data itself. A fine-tuned model is usually best protected as a trade secret backed by contracts, with patents available for some training methods. This guide is for technology companies in Europe and Latin America that want to know which AI assets they actually own.

Key takeaways

  • A short instruction is an unprotectable idea. A long, original system prompt may be a literary work, but its value usually lies in keeping it secret.
  • The US Copyright Office concluded in January 2025 that prompts alone do not make a user the author of the output.
  • In the EU, a dataset can attract the sui generis database right for 15 years if obtaining, verifying or presenting it required substantial investment. The Andean regime has no equivalent right.
  • Model weights are not clearly copyright works; trade secrets and licences do most of the work.
  • The EPO accepts AI inventions when the model serves a technical purpose or a specific technical implementation; patenting means disclosing.

Are prompts protected by copyright?

Copyright protects original expression, not ideas. Spain’s Intellectual Property Act (TRLPI) protects “original literary, artistic or scientific creations” expressed in any medium (article 10 TRLPI, BOE). In the Andean Community (Bolivia, Colombia, Ecuador and Peru), Decision 351 states that only the form in which ideas are expressed is protected, never the ideas themselves (article 7).

Applied to prompts, the line is clear. “Write a product description in a friendly tone” is an instruction nobody owns. A system prompt of several pages, with an original structure, examples and drafting choices, can be a literary work. Copying it word for word could infringe; rewriting the same instructions generally would not.

The US Copyright Office report on copyrightability (January 2025) follows the same logic. It accepts that prompts themselves may be copyrightable if sufficiently creative, but concludes that, with current technology, prompts alone do not give the user enough control to be the author of the output: they “function as instructions that convey unprotectible ideas”. Human material that the user feeds in and that remains perceptible in the output is a different matter.

The prompts that matter commercially encode business logic: tone rules, workflows, guardrails, product knowledge. Copyright in that text is thin protection, because a competitor can rewrite it. That is why we usually advise treating system prompts as confidential know-how.

Can prompts and know-how be protected as trade secrets?

Yes, if they meet the three conditions in Spain’s Trade Secrets Act (Law 1/2019, article 1), which implements Directive (EU) 2016/943: the information is not generally known or readily accessible, it has business value because it is secret, and its holder has taken reasonable measures to keep it secret. Prompts, evaluation sets, data-cleaning rules and hyperparameters fit when handled as confidential.

The weak point is the third condition: a system prompt exposed in a public interface, or shared with contractors without a confidentiality clause, is hard to defend. The law also treats as lawful the observation or testing of a product available to the public, unless a valid obligation prevents it (article 2(1)(b) of Law 1/2019). Your terms of use are what turn prompt extraction or model distillation from lawful reverse engineering into a breach.

Can a training dataset be protected?

A dataset can be protected in two different ways, and neither covers the individual data points.

  • Copyright in the structure: a database whose selection or arrangement of contents is an intellectual creation is protected as such, but only as to that structure (article 12 TRLPI; article 28 of Decision 351).
  • The sui generis database right: in the EU, the maker of a database who has made a substantial investment in obtaining, verifying or presenting its contents can prohibit the extraction or reuse of all or a substantial part of it (article 133 TRLPI). It lasts 15 years from completion, and a substantial new investment gives the updated database its own term (article 136 TRLPI).

For AI teams, the second right matters most: a curated, cleaned and annotated corpus often reflects exactly that investment. Keep evidence of it: budgets, labelling hours, verification processes and version histories.

Two cautions. Decision 351 contains no investment-based database right, so in Andean markets rely on contracts and trade secrets and check national law. And if you provide a general-purpose AI model in the EU, recital 107 of the AI Act requires a public summary of training content that names the main datasets, while taking due account of trade secrets. Decide early what will be disclosed and what stays confidential.

Who owns a fine-tuned model, and what protects it?

A model is a bundle of assets, and each one has a different form of protection.

Asset Main protection What to check
Short prompts and parameters None as such; confidentiality if secret Terms of use, access controls
Long system prompts and guidelines Copyright in the text if original; trade secret Who wrote them; confidentiality clauses
Curated datasets Sui generis right (EU); copyright in structure; trade secret Licences for source data; proof of investment
Training and inference code Copyright as a computer program Employee and contractor assignments; open-source licences
Model weights Trade secret and contract; copyright status unclear Base model licence; distribution terms
Training or fine-tuning method Patent, if technical; otherwise trade secret No disclosure before filing

Code is the most solid asset. Computer programs are protected if they are their author’s own intellectual creation, but the ideas and principles behind them are not (article 96 TRLPI). Where an employee writes the program as part of their duties, the economic rights belong to the employer unless agreed otherwise (article 97.4 TRLPI). The law gives no equivalent rule for freelancers and agencies, so their work needs a written assignment.

Weights are less certain. In Getty Images v Stability AI (High Court of England and Wales, 4 November 2025), the court found that the model weights “have never contained or stored” the training images and are the product of the patterns learnt during training; Getty was given permission to appeal, according to CMS (21 January 2026). That reasoning helps developers against infringement claims, but it also shows that weights are the output of a computation rather than human expression. Our recommendation is to protect them through confidentiality, licence terms and access controls, not copyright.

If you fine-tuned someone else’s model, the base licence comes first: some restrict commercial use or impose terms that pass on to derivative models.

Can you patent an AI model or a training method?

At the European Patent Office, yes, under conditions. The EPO Guidelines for Examination (April 2026), G-II 3.3.1, treat neural networks and similar models as abstract mathematical methods in themselves. They contribute to an invention’s technical character when applied in a field of technology or adapted to a specific technical implementation. A neural network in a heart monitor that detects irregular heartbeats qualifies; classifying text documents purely by their content does not. Generating the training set and training can also count when they serve that technical purpose.

The trade-off is disclosure: the application must reveal the dataset characteristics needed to reproduce the technical effect, though not normally the dataset itself. Publishing a paper or demo before filing can destroy novelty.

What this means for your business

  1. Build an AI asset register: prompts, datasets, code, weights and methods, with owner and protection route for each.
  2. Close the chain of title: employment contracts, written assignments from contractors and annotators, and licences for every data source.
  3. Apply trade secret measures that you can prove: access controls, logging, confidentiality clauses and terms of use that prohibit extraction and distillation.
  4. Decide patent versus secret before any publication, with an EPO filing strategy that can be extended to Latin America through the PCT (Patent Cooperation Treaty).
  5. Map differences across the corridor: the EU database right does not exist in the Andean regime, so contracts carry more weight there.

If you need this mapped across several jurisdictions, our team for AI and digital asset protection can review your portfolio and prepare the filings and contracts, together with our European and international patent team where a method may be patentable.

Where companies get AI asset protection wrong

  • Relying on copyright for prompts. A rewritten prompt escapes copyright; only confidentiality and contract protect the know-how.
  • Exposing the asset without terms. A public interface without terms that prohibit extraction weakens any later trade secret claim.
  • Missing assignments. Code, prompts or labels produced by freelancers or agencies remain theirs unless assigned in writing.
  • Publishing before filing. Conference papers and launch posts can make a patentable method unpatentable.
  • Ignoring upstream licences. A base model or dataset licence can limit commercial use of everything built on it; our IP licensing and enforcement team reviews those terms before a launch or a deal.

Frequently asked questions

Can I copyright a prompt?

Only if the prompt is an original piece of writing in its own right, such as a long system prompt with a creative structure and drafting choices. Short instructions and style requests are unprotected ideas, and even a protected prompt does not make you the author of the output. For valuable prompts, confidentiality and contractual terms are usually stronger protection.

Is a training dataset protected in the European Union?

The data points themselves are not, but the dataset can be. If its selection or arrangement is an intellectual creation, copyright protects that structure. If obtaining, verifying or presenting the contents required substantial investment, the sui generis database right lets the maker prevent extraction or reuse of a substantial part for 15 years, renewed by substantial new investment.

Who owns a model fine-tuned by an external contractor?

It depends on the contract and on the base model’s licence. Under Spanish law, software written by employees in the course of their duties belongs to the employer, but there is no such rule for contractors, so their code, prompts and weights need a written assignment. The base licence must also allow you to create and exploit the fine-tuned version.

Can IP Global Guard audit and protect our AI assets?

Yes. We map your prompts, datasets, code, weights and methods, identify the right protection for each, draft assignments, confidentiality terms and licences, and prepare patent and copyright filings. We act directly where our professionals are qualified and otherwise through qualified representatives, coordinating local correspondents across Europe, Latin America and Africa from a single point of contact.

How IP Global Guard can protect your AI assets

The value of an AI company sits in assets that the law protects in very different ways, and getting the mix right is what makes them defensible in a dispute and credible in a due diligence. IP Global Guard, the IP services line of META Channel Corporation Limited, handles this with one strategy and one billing relationship across more than 25 jurisdictions; see our coverage in Europe, Latin America and Africa. Within the same group, our regulatory team covers the AI Act obligations that interact with these assets.

Send us a list of your models, datasets and the markets where you operate. We will tell you what you own, what is exposed and what to file or sign first. Ask our AI and IP team for a review.

This article is general information, not legal advice, and reflects the position on its publication date.

Sources

Share

IP Global Guard

A specialist industrial and intellectual property service line of META Channel Corporation.

IP Global Guard is a service line of META Channel Corporation Limited, a Private Company Limited by Shares incorporated in Ireland. CRO 801240. Registered office: 34 Holywell Avenue, Feltrim Hall, Swords, Dublin, K67 KF22, Ireland.

© 2026 META Channel Corporation Limited. All rights reserved.

Secret Link