The Model Is Only One Layer: Fine-Tuning and Integration

A base model can satisfy the conditions for narrow Article 53 relief and still enter a new legal and technical relationship when someone changes it or embeds it in an application. This chapter focuses on those two transitions: when a fine-tuner may become a provider of a new model, and how an integrator’s intended-use decision creates system-level responsibility. The practical judgment is simple but important: the treatment of an openly released base model does not automatically travel unchanged through every derivative or application.

The word “provider” can become confusing here because the AI value chain contains more than one provider role. One actor provides the general-purpose model. Another may provide a modified model. A third may provide an AI system that uses the model. The legal analysis has to identify which object each actor has made available and which decisions they control.

Fine-tuning changes the object under discussion

Fine-tuning starts with an existing model and trains it further. The additional training may adapt the model to a domain, a narrower task, a particular style of output, or a new set of constraints. The downstream actor is not merely downloading and operating the base model; they are making a design choice that can alter how the model behaves and how it is presented to others.

That difference explains why a fine-tuner may become a provider of a new model. The scholarly analysis describes the Act as recognising that a GPAI model can be further modified or fine-tuned into a new model, with the downstream actor taking on corresponding responsibilities. The legal consequence is not that every small adjustment automatically creates a new provider. It is that the base model’s provider status and the modifier’s status are separate questions.

Consider a general model released for broad language tasks. A company trains it further on a specialised corpus and presents the result as a model for legal advice. Another company adapts a base model for use in medical diagnostics. These examples do not by themselves establish the legal classification of the resulting systems. They show why the downstream actor cannot simply point to the base model’s release conditions and assume the analysis is complete. The modification may introduce new capabilities, new limitations, or a new intended role.

The distinction also matters for documentation. The source analysis describes a proportional approach in which a fine-tuner’s documentation would focus on the changes made while relying on the original model’s documentation for the remainder. That makes practical sense as an information arrangement: a modifier may not have recreated the base model, but they do know what additional data, objectives, or constraints they introduced.

Yet proportionality is not the same as automatic inheritance. If the fine-tuned model is made available as a new model, the downstream actor may be responsible for explaining the part of the model they changed and the consequences of that change. The open-source treatment of the original model does not answer whether the modified model satisfies the same conditions, whether it is released under the same terms, or whether new risks have appeared.