AI training data licensing agreement: key clauses

An AI training data licensing agreement should state exactly which uses are allowed (training, fine-tuning, retrieval or outputs), who guarantees the origin of the content, how opt-outs and removals work, how the licensor is paid and how compliance is checked. In Spain, Portugal and the main Latin American markets, copyright licences are read restrictively, so a vague clause or an old licence that never mentioned AI will rarely cover model training. This guide is for publishers, labels, image libraries and AI developers negotiating deals across Iberia and Latin America.

Key takeaways

The European Parliament’s resolution of 10 March 2026 calls for a functioning licensing market and opposes flat-rate global licences for AI training.
Spanish, Portuguese, Mexican and Brazilian law limit a licence to the uses it expressly names; anything not written down is presumed excluded.
The five clauses that matter most are scope of use, warranties of origin, opt-outs, remuneration and audit.
What happens to an already trained model when the licence ends must be agreed upfront; it cannot be fixed later.
A choice-of-law clause governs the contract, but infringement claims remain subject to the law of each country where protection is claimed.

Why licensing is becoming the main route to AI training data

Under Article 4 of the DSM Directive, implemented in Spain by Article 67 of Royal Decree-law 24/2021, commercial text and data mining is allowed unless the rights holder has reserved its works by machine-readable or other appropriate means. Once a reservation is in place, a developer needs permission. German courts have also held that works memorised in a model are reproduced without cover from that exception (Munich, GEMA v OpenAI, 2025, and GEMA v Suno, 2026).

Policy is moving the same way. In its resolution of 10 March 2026, adopted by 460 votes to 71 with 88 abstentions (Eunews), the European Parliament called for a licensing framework that restores rights holders’ bargaining power, voluntary collective licensing per sector, and remuneration negotiated in good faith; it opposed any global licence for training in exchange for a flat-rate payment (paragraphs 2, 9 and 21). In its follow-up to the resolution, the Commission said it will explore measures to facilitate licences that respect the market value of content. The copyright chapter of the General-Purpose AI Code of Practice (10 July 2025) states that it does not affect agreements between signatories and rights holders authorising the use of works, so a licence sits alongside, not under, the Code.

How Iberian and Latin American law reads a licence

The common thread in the corridor is restrictive interpretation: if a use is not named, it is not licensed. That matters for AI, because most existing licences were drafted before model training existed.

Country	Rule	Effect on AI training
Spain	Intellectual Property Act, arts. 43, 45 and 46: transfer limited to the modes of exploitation expressly provided; five years and the country of signature if not stated; no unknown modes; written form; proportional remuneration as a rule	A licence silent on AI is unlikely to cover training; uses unknown when signed are excluded
Portugal	Copyright Code, arts. 41, 43 and 44-A: written authorisation specifying the form of use, time, place and price; adequate and proportionate remuneration	The AI use must be described specifically; authors and performers can claim proportionate pay
Mexico	Federal Copyright Law, arts. 30, 31 and 33: licences in writing or void; transfers always for payment and for a limited time; five years if not stated	Unwritten or perpetual “all uses” deals are vulnerable
Brazil	Law 9.610/1998, arts. 4, 49 and 50: copyright deals read restrictively; only existing modes of use; country of signature unless agreed; five years without a written term	Same logic as Spain: name the AI uses and the territory

These rules protect authors in their dealings with publishers and producers. A licensor that sublicenses to an AI developer can only pass on what it received, so the first check is always the chain of title.

Key clauses in an AI training data licensing agreement

1. Scope of use

Define each permitted act: pre-training, fine-tuning, evaluation, retrieval-augmented generation (RAG, where the system consults the content at the time of answering) and any display in outputs. Name the models or model families covered, whether successor models are included, the territory and the term. The Parliament’s resolution treats uses beyond training, such as inference and RAG, as requiring express consent (paragraph 6), so do not assume a training licence covers them.

2. Warranties of origin and chain of title

The licensor should warrant that it holds the rights it grants, including from authors, translators, photographers and performers, and disclose any content it cannot license. In Mexico, the 2026 reform of Article 87 of the Federal Copyright Law requires performers’ express consent for use of their image, including voice, and extends that protection to AI-generated results; content featuring performers needs separate clearance.

3. Opt-outs, exclusions and removals

Agree how excluded works are flagged, how quickly removal requests are applied to future training runs, and whether the developer will respect machine-readable reservations on content outside the licence. The Code’s copyright chapter commits signatories to follow robots.txt (RFC 9309) and other appropriate machine-readable protocols when crawling (Measure 1.3).

4. Remuneration

Options include a fixed fee, per-use or per-token payments, revenue share or a mix. Under Spanish and Portuguese law, authors are generally entitled to proportionate remuneration, so licensors should check that their upstream contracts allow the chosen model.

5. Transparency and audit

Providers of general-purpose AI models must publish a training content summary under Article 53(1)(d) of the AI Act. A licence can go further: usage reports, records of which datasets fed which model, and an audit right exercised by an independent expert, with confidentiality protection for trade secrets.

6. Outputs, termination and liability

Require safeguards against outputs that reproduce licensed content, mirroring Measure 1.4 of the Code. Decide what happens at termination: whether models already trained may continue to be used, and on what terms. Allocate liability with indemnities and caps that reflect who controls the data and the model. Remember that, under Article 8(3) of the Rome II Regulation, the law applicable to an infringement cannot be changed by agreement.

What this means for your business

Licensors: audit your catalogue and upstream contracts before negotiating; you cannot license AI uses you never acquired.
Developers: map where each dataset comes from and which national rules govern it; a Brazilian archive and a Spanish publisher may need different drafting.
Both: write the AI uses out in full, by name, rather than relying on “all media now known or later developed” language.
Plan registration and evidence: in Mexico, transfers must be recorded in the Public Copyright Register to have effect against third parties (LFDA, art. 32).

When a deal spans several countries, our AI and digital assets licensing team can align the drafting with each jurisdiction, and our IP licensing and transactions practice handles the wider contract.

Where cross-border AI licensing deals go wrong

Relying on legacy licences. In Spain and Brazil, a contract does not extend to modes of use unknown or non-existent when it was signed.
No chain-of-title review. A licensor warranting rights it never cleared exposes both parties.
Undefined model scope. “The licensee’s AI models” can later be read to include products the licensor never intended.
Ignoring termination. Without a clause on trained models, the end of the licence becomes a dispute.
One template for every market. Formalities such as written form, notarised signatures for partial transfers in Portugal (art. 43) or registration in Mexico differ.

Frequently asked questions

Does an existing content licence cover AI training?

Usually not, unless it names AI training expressly. Spanish law limits a transfer to the modes of exploitation expressly provided and excludes modes unknown when signed; Brazilian law reads copyright deals restrictively and only covers existing modes of use. A licence signed before generative AI, or silent on it, should be renegotiated or confirmed in writing.

Is a flat fee acceptable for licensing content to train AI?

It can be agreed between companies, but check upstream rights. Spanish and Portuguese law generally give authors proportionate remuneration, and the European Parliament opposed global flat-rate training licences in March 2026. Many deals combine an upfront fee with usage-based payments and reporting.

What happens to a trained model when the licence ends?

Only what the contract says. Removing specific content from a trained model is often impractical, so parties should agree in advance whether existing models may keep being used, whether new training must exclude the content and whether a payment applies after termination.

Can IP Global Guard negotiate an AI licensing deal across Spain, Portugal and Latin America?

Yes. We review the chain of title, draft and negotiate the licence and align it with each country’s formalities, coordinating local correspondents in Latin America from a single point of contact. Within the same group, META Channel also advises on the AI Act and GDPR issues these deals raise.

How IP Global Guard can help you license content for AI

Licensing for AI training turns on precise drafting and a clean chain of title in every market where the content and the model operate. IP Global Guard, the IP services line of META Channel Corporation Limited, structures licences, assignments and portfolio strategy across more than 25 jurisdictions in Europe, Latin America and Africa, with one strategy and one billing relationship.

Send us the term sheet or draft, the catalogue involved and the countries concerned. We will flag the gaps in scope, warranties and remuneration and propose clauses that work in each market. Talk to our licensing team about your deal.

This article is general information, not legal advice, and does not replace an assessment of your specific contract.

AI training data licensing: key clauses for Iberia and Latin America

Key takeaways

Why licensing is becoming the main route to AI training data

How Iberian and Latin American law reads a licence