Training or fine-tuning a model on your own work may appear straightforward, yet ‘yours’ may refer to the physical copy, the authorship or the rights. The process also incorporates the base model, code, weights, metadata and services with their own conditions.
Fine-tuning does not automatically render the artist a provider of a general-purpose AI model. The model, the scope of the modification and the manner in which it is offered are significant factors that must be analysed in accordance with the text of the Regulation and the Commission's guidelines.
Four distinct operations
- Train from scratch: fit an architecture on a large dataset.
- Fine-tune a model: modify weights through fine-tuning, LoRA or another technique.
- Use a service: upload files for a third party to customise the model.
- Generate without adapting: use prompts or references without producing new weights.
Identify the actual process. Do not call every generation training.
Rights regarding the dataset
For each file, record the work, authorship, rights holder, photograph or digitisation, source, licence, restrictions, personal data and inclusion.
Being the author of a painting does not imply being the rights holder of a photograph taken by another person. Purchasing a work does not transfer exploitation rights. Review commissions, collaborations and licensed resources. Consult how to authorise the reproduction of a work and how to reserve your rights against data mining.
Personal data and confidentiality
Faces, voices, private text or location data may remain in files, logs or weights. Document the purpose, legal basis, minimisation, retention, processors, international transfers and memorisation tests. An authorisation to exhibit does not necessarily authorise training.
If the model receives data from the public, consult the guide If your work collects data from the public.
Base model licence
Identify the exact version, file, hash and licence. Separate training and inference code, architecture, weights, base, adapter, usage policy, hosting and output conditions.
Families do not always share the same licence. According to its licensing page, Stability AI offers a Community Licence, free for organisations with less than one million dollars in annual revenue, and an Enterprise Licence for others; the list of models covered by each appears on its site. Black Forest Labs publishes commercial licence tiers according to the model and links non-commercial terms separately. Conditions depend on the specific model and version: do not extrapolate one version to the entire brand or label a model as open simply because its weights can be downloaded.
Retain a dated copy of the licence and terms. Note:
- provider, model and version;
- file and hash;
- date of download;
- licence and usage policy;
- code licence;
- commercial and redistribution limits;
- conditions applicable to outputs.
Dataset manifest
Do not merely save a folder. Maintain an identifier, path, rights, transformations and partition. Separate the training dataset, validation dataset, test dataset, exclusions, duplicates, originals, crops, labels and synthetic data.
Retain the original and the transformation recipe. Describe the selection criteria: 'all my work' may conceal biases related to period, preservation or image availability.
Documenting the fine-tuning
Record the objective, base model, code and commit, method, hardware, steps, epochs, batch size, resolution, learning rate, optimiser, seeds, precision, compute, checkpoints, metrics, errors and decisions.
Not everything must be published; however, it must be retained for repetition and auditing.
Weights, adapters and versions
Distinguish between base weights, fine-tuned weights, LoRA, checkpoints, configuration, merge and deployed model. Add version, size and checksum. An adapter does not replace the base model nor grant the right to redistribute it.
Evaluation before use
Test for excessive similarity, memorisation, trademarks, signatures, data, biases, stability and out-of-scope capability. Retain both successful and failed results.
AI Act
Article 53 of the AI Act requires providers of general-purpose AI models (GPAI) to maintain technical documentation, provide information to those integrating the model, adopt a copyright compliance policy, and publish a sufficiently detailed summary of the content used for training. The first two obligations do not apply to certain models published under a free and open-source licence, except for those with systemic risk.
According to the European Commission, the obligations for GPAI model providers apply from 2 August 2025, and their enforcement powers from 2 August 2026. The AI Act was amended by Regulation (EU) 2026/1744, the so-called Digital Omnibus on AI, dated 8 July 2026 and in force since 27 July 2026 according to EUR-Lex. The official explorer of the Regulation provides the consolidated text as of that date and does not indicate changes to Article 53.
The Commission's guidelines are not binding: they reflect its interpretation and will guide application. They indicate that only those who perform significant modifications must comply, not those who introduce minor changes. Their detailed criteria for distinguishing between them have not been read here.
To analyse your case, ask yourself:
- Is it a model capable of a wide variety of tasks?
- Does the modification significantly change its capabilities or its generality?
- Is it offered in the Union under your name?
- Is it used solely internally to create artwork?
- Are weights or API access distributed?
- What cumulative compute was used?
Do not assume that a minor adjustment exempts you. If you distribute a broad model, request a specific analysis.
Responsible documentation and legal obligation
Even if you are not a provider of a GPAI model, retain the purpose, rights, model, method, dataset, code, parameters, compute, weights, evaluation, discarded uses, responsible party and date. Do not present this best practice as a universal obligation.
Human authorship of outputs
In Spain, Article 5.1 of the Intellectual Property Act identifies the natural person who creates the work as the author, and Article 10.1 protects original creations. No specific Spanish or European Union rule has been identified regarding the protection of outputs from a model fine-tuned on your own work, nor is there a criterion that automatically protects every result generated by AI.
The U.S. Copyright Office concluded in January 2025 that the use of AI as a tool to support human creativity does not reduce the protection of the work, but that, with the technology examined, prompts alone do not provide sufficient control over the expression. According to the report, the creative selection, coordination, or arrangement of material and the creative modifications made by a person may be protected, each case separately.
This is comparative law, not a Spanish rule. Retain source works, dataset selection, fine-tuning decisions, iterations, discards, subsequent editing, and human combination. Documentation does not establish authorship, but it assists in its analysis.
Publishing, selling, or entering competitions
Review the terms and conditions. They may require AI disclosure, prohibit certain models, or request traceability. Do not state "trained solely with my work" if a base model trained on other materials exists. Use the following formulation: "base model X fine-tuned with an identified set of the artist’s own works".
Delivery and preservation
Determine whether the manifest, base, adapter, weights, code, configuration, documentation, outputs, licence, and environment form part of the work. Do not distribute data or weights without verifying rights. For the technical package, consult How to prepare creative code so that it can run again and Software licences and artistic works: what each licence covers.
Checklist
- Training, fine-tuning, service, and inference are distinguished.
- Authorship, ownership, and photography have been reviewed.
- Personal and confidential data are controlled.
- Model, version, hash, and licence are recorded.
- Code, weights, policy, and outputs are separated.
- The dataset includes a manifest and partitions.
- Parameters, hardware, seeds, and controls are recorded.
- Memorisation, signatures, biases, and failures were evaluated.
- An analysis was conducted on whether a GPAI model is being distributed or placed on the market.
- The guidelines are not presented as binding law.
- Human decisions are preserved.
Further reading
- How to reserve your rights against data mining.
- If your artwork collects data from the public.
- Registering a work with the Intellectual Property Registry.
- How to authorise the reproduction of a work.
- Software licences and artistic works: what each licence covers.
- How to prepare creative code so that it can run again.
Sources
- Official AI Act Explorer. Consolidated text as of 27 July 2026; refers to the Official Journal texts as the authentic version.
- AI Act, Article 53, in the official explorer. Obligations for providers of general-purpose models; consolidated as of 27 July 2026.
- European Commission, guidelines for GPAI model providers. Non-binding nature, application dates; last page update: 28 April 2026. The full document of the guidelines has not been read.
- Regulation (EU) 2026/1744, in EUR-Lex. Digital Omnibus on AI, dated 8 July 2026; only the page metadata has been read, not its articles.
- Intellectual Property Act, consolidated BOE text. Articles 5.1 and 10.1; last update published on 30 March 2022.
- U.S. Copyright Office, report on copyright and artificial intelligence, part 2. January 2025; comparative law.
- Stability AI, licences. Community Licence and Enterprise Licence; the page only shows the 2026 copyright notice, without an update date.
- Black Forest Labs, licences. Commercial licence tiers by model; the page only shows the 2026 copyright notice, without an update date.
Status: Verified. Review: 11 October 2026. General information, not legal or data protection advice. If you find an error, write to me.