Licenses and Intellectual Property When Using Open-Source AI Models in Public Administration

Why the license matters before deploying an AI model

Open-source AI models are attractive for cost, control and speed of deployment. But the license that accompanies the model (and its training data) defines rights and obligations that can directly affect a public administration: commercial use, obligations to publish code or weights, liabilities for derivative content and even the need to include specific clauses in public contracts (Law 9/2017). Ignoring these conditions creates legal, operational and continuity risks.

Below we explain the key elements every municipal team should review and translate the conclusions into concrete operational steps.

Key legal and technical points to assess

• Model license (software/weights)

  • Permissive (Apache 2.0, MIT): allow commercial use and modification; Apache adds a patent grant and requires preservation of copyright notices.
  • Strong copyleft (GPL, AGPL): require publishing source code when distributing; AGPL extends that obligation to network use (SaaS). If you plan to offer a public service built on the model, AGPL could force you to release code or derived works.
  • Model or data licenses with use clauses (for example, prohibitions on commercial use or attribution requirements) may limit or prohibit use by a public entity.

• Dataset licenses

  • Data under CC0/CC-BY are generally usable, but CC-BY-NC or clauses that limit commercial exploitation can prevent administrative uses or create incompatibilities with other licenses.
  • Verify the origin and rights over texts, images or databases included in training: lack of a clean-room process in the data increases the risk of copyright claims.

• Derivatives and fine-tuning

  • Fine-tuning a model can create a “derivative” that is subject to the same license. With strong copyleft this may require sharing weights or generated code.

• Provider contracts and warranties

  • Many open-source models are distributed “as is” without warranties. For public use, require contractual guarantees about the provenance of training data, absence of infringements and mechanisms for indemnification.

• Data protection and GDPR

  • If the model was trained with personal data or processes personal data in production, GDPR applies: legal bases, impact assessment (DPIA), retention and data subject rights.

• Regulatory and security requirements

  • ENS applies to the deployment of solutions in public entities. In addition, the AI Act imposes obligations for documentation and risk management for higher-risk systems.

Operational checklist for municipal teams

1. Preliminary inventory

   • Identify the model, version, license URL, and the dataset license. Record it in the digital services catalog.

2. Initial legal review

   • Check whether the license is permissive or copyleft. Flag models with AGPL/GPL as “require legal and technical review” before use.

3. Provider due diligence

   • Request in writing the provenance of the training data, a list of sources and certification of rights/licenses. Require indemnity clauses against third-party claims.

4. DPIA and risk classification

   • If personal data or impacts on citizens’ rights are involved, perform a Data Protection Impact Assessment (DPIA) and register the system under the AI Act if applicable.

5. Contract and mandatory clauses

   • Explicit use rights for the administration.
   • Delivery of technical documentation and model cards.
   • Delivery of weights and artifacts into escrow or with the right to export them at contract termination.
   • Confidentiality, security (ENS) and business continuity clauses.
   • Warranties about data origin and liability for IP infringements.

6. Testing in a controlled environment

   • Before production, test the model in a closed environment to detect problematic content and validate compliance.

7. Exit plan

   • Define how to replace the model if a claim arises or if the license changes. Ensure portability of data and artifacts.

Practical contractual clauses (high level)

• Warranty of data origin: the provider declares and warrants that the data used to train the model do not infringe third-party rights and that it holds the necessary licenses.
• Delivery of artifacts: the provider will deliver weights, code and technical documentation (model card, risk sheets) in an interoperable format and to a municipal-accessible repository.
• Right of audit: the administration may audit the data custody chain and the training process under a confidentiality agreement.
• Indemnity for IP claims: the provider assumes responsibility and costs for third-party IP claims.

Note: these are operational guidelines. Every clause must be reviewed by the legal department.

Practical decisions: host vs use via API

• Using an API (MaaS):

  • Pros: less technical responsibility; updates managed by the provider.
  • Cons: reduced sovereignty, obligation to review terms of service and possible restrictions from incompatible licenses.

• Hosting locally:

  • Pros: control, easier ENS compliance, ability to audit weights and fine-tunes.
  • Cons: higher operational burden and need for security and ML experts.

The choice should be based on internal capacity, sovereignty requirements and the type of license.

Takeaway — Recommended action for the next 30 days

Complete the checklist: identify all models/datasets in use, classify their licenses, and request minimum guarantees from the provider (data provenance, delivery of artifacts and indemnity clauses). If you plan to offer public services based on a model with a copyleft license (AGPL/GPL), stop the move to production until legal and technical review is completed.

For modular projects and a services catalog, integrating this license review into procurement templates avoids legal surprises and helps ensure compliance with ENS, GDPR and AI Act obligations. OptimGov can be integrated into the catalog and help maintain the inventory and traceability of licenses as part of the service lifecycle.