By Cesar Augusto Taborda Lima
A company that commissions an artificial intelligence system usually signs a software development contract. The trouble is that an AI system is not just software. There is the code, which Brazilian law handles well. There is the trained model, with its weights, about which the law says nothing. There are the data the client handed over and the data the developer prepared and labelled from them. And there is the question nobody asks at signing: where the material used in training came from. When the contract is silent on those layers, the client pays for the whole project and finds out, at the end, that only the code is theirs.
Art. 4. Unless otherwise agreed, the rights in a computer program developed and produced during a contract or statutory relationship expressly intended for research and development, or in which the activity of the employee, contractor or civil servant is provided for, or which arises from the very nature of the duties of those relationships, belong exclusively to the employer, the hirer of the services or the public body.
Law 9,609/1998 (Brazil’s Software Act), art. 4, caput (free translation).
What the law settles
Software is protected as a literary work (Law 9,609/1998, art. 2). Art. 4 does the rest: a program developed during a research and development contract belongs to whoever commissioned it, unless otherwise agreed. Up to this point the client is reasonably protected even under a poor contract.
The surprise is in art. 5. Authorised derivative versions belong to whoever makes them, again unless the contract says otherwise. Picture a developer who delivers a document classifier to a freight company and, months later, adapts the same system for a customs broker. If the freight company’s contract did not bar reuse, the law is on the developer’s side.
What the law does not settle
The weights of a neural network are millions of numbers adjusted by an automated process. They are not instructions someone wrote, so they are not a program. Nor are they a work, because Law 9,610/1998, the Copyright Act, requires human creation (arts. 7 and 11). A database is protected only for the selection or arrangement of its contents, never for the data themselves (art. 7, XIII and § 2). A training set assembled with the client’s effort and money may, in the end, have no owner in the copyright sense.
What is left is trade secrecy. Law 9,279/1996, the Industrial Property Act, makes it a crime of unfair competition to disclose or exploit, without authorisation, confidential information obtained through a contractual relationship, even after the contract has ended (art. 195, XI). But a secret does not arise on its own. It exists to the extent the contract defines it: what is confidential, who has access, for how long, under what controls. A contract that never says the weights are confidential will struggle to prove later that they were.
And where copyright does apply, it is suspicious of contracts. Dealings in copyright are construed restrictively (art. 4). An assignment must be in writing and is presumed to be for value (art. 50). Without a written term it lasts five years, applies only in the country where the contract was signed and only to the forms of use that existed on that date (art. 49, III to V). Future works, five years at most (art. 51). Some development contracts assign “all rights in the system” in a single line. Under the law, that line is worth much less than it looks.
An allocation that works
I prefer contracts that list what is being delivered, layer by layer, and say who owns each one. There is no single formula, but a reasonable starting point is this.
The project-specific code goes to the client, with the source code, the comments and the training scripts. Without the scripts, incidentally, the model cannot be reproduced and the client is a hostage. The developer’s general tools, the ones used on every project, stay with the developer; the client gets a perpetual licence for whatever was embedded in the deliverable. Data supplied by the client remain the client’s. Data labelled from those data too, because the labelling was paid for. The trained model and its weights are assigned, or licensed exclusively, and the developer is barred from reusing them for another client, whole or in part, including by distillation or by fine-tuning on top of them.
There is a limit the client needs to understand. If the developer started from a third-party model, open or licensed, the upstream licence rules. Open models come with use restrictions and attribution requirements that follow the fine-tuned model. The contract should list each component and its licence, and the client receives no more than the developer obtained.
Where the data came from
This is the clause whose value has changed most in the last two years. The developer declares where the training data came from, that it was authorised to use them, that it circumvented no technical protection and processed no personal data outside the cases the LGPD, Brazil’s data protection law, allows. And it takes on the defence and the costs if anyone claims that the training, or what the system produces, infringes rights.
The reason is concrete. On 31 July 2026 the Munich Regional Court I found against Suno for training its music generator on protected works and returning them recognisably in the outputs (the court’s press release, case 42 O 763/25, not yet final). Brazil does not need a new statute to reach a similar result: Law 9,610 already makes inclusion of a work in a database and its storage in a computer subject to the owner’s authorisation (art. 29, IX). The AI bill PL 2,338/2023, stalled in the Chamber of Deputies since 2025, would go further and require developers to publish a summary of the protected content used in development (art. 62). Whoever commissioned the system is the one who puts it on the market. The notice lands on their desk, not the developer’s.
With personal data, a developer who processes the client’s data to train the model is a processor and follows the controller’s instructions (LGPD, art. 39). The contract should say what the legal basis is for each use. If it is legitimate interest, the law asks for a concrete purpose, transparency and records of processing (arts. 10 and 37). Anonymised data fall outside the law, but only while the anonymisation cannot be reversed with reasonable effort (art. 12), and a model that memorises training records can give them back to anyone who asks the right way. It is worth requiring memorisation tests before accepting delivery. International transfers and incidents follow the rules I covered in the article on contracts with AI model providers.
Accepting something that does not always work
Ordinary software works or it does not. A model is right a percentage of the time, and that percentage falls as the world changes and the model does not. The contract needs to fix the metrics, the test set (kept apart from the training set), the acceptance threshold and what happens when performance drops after delivery. It also needs to provide for model documentation: data used, known limitations, bias evaluations, conditions of use.
That documentation stops being good practice when the system goes to the European Union. The AI Act requires anyone supplying components of a high-risk system to specify, by written agreement, the information and technical access the provider needs to comply with the regulation (art. 25(4)), on the 2027 and 2028 dates set by Regulation (EU) 2026/1744. In March 2025 the European Commission published model contractual clauses for the procurement of AI, one version for high-risk systems and a lighter one. They were written for public buyers, but they serve as a yardstick for any contract.
Liability and exit
The limitation of liability deserves separate caps for confidentiality, personal data and intellectual property indemnity. Those are the real risks; the rest is detail. The Civil Code requires the allocation of risks the parties agreed to be respected (art. 421-A, II), and what they did not agree the judge fills in as best he can.
On exit, the client takes the source code, the weights, the labelled data, the documentation and the training scripts, in a format that allows retraining elsewhere, with written confirmation that the developer has deleted its copies, save what the law requires it to keep.
As I see it, the most common mistake in these contracts is treating the model as if it were the software. “The system” is assigned and nobody asks what the system is. A list of deliverables with the owner of each one settles more than ten pages of general clauses.