Definition of general-purpose AI, transparency obligations and "Model Cards"
The details of the new European AI Act are still pending following the provisional political agreement reached by the EU institutions on December 8 and are expected to be finalized by February 2024. Nevertheless, it is worth taking a first look at some key points, such as the legal definition of general-purpose AI models and the two-tier approach for foundation models, to see what developers and users of such applications are likely to face.
What is "general purpose AI" under the AI Act?
The newly introduced definition is already causing a bit of a headache. This may be due to the fact that it is very difficult to legally define an application that is not particularly special but is characterized by a broad and general scope of application. Consequently, the attempt at a definition must also be general in order to cover as many forms as possible. This naturally leads to a certain vagueness.
According to the definition under the AI Act, general-purpose AI model means an AI model "…that displays significant generality and is capable to competently perform a wide range of distinct tasks regardless of the way the model is released on the market and that can be integrated into a variety of downstream systems or applications."
But what exactly is "significant generality"? What criteria may be used to determine that an individual AI model does not exhibit "significant" generality and thus falls outside the scope of the catalogue of requirements?
A newly inserted recital is intended to help with this, in which the number of parameters used during development to achieve significant generality is used as a criterion. The recital refers to "(…) at least a billion of parameters and trained with a large amount of data (…)".
This number seems quite low when you consider that even GPT-2 in 2019 used 1.5 billion parameters while GPT-3 from 2020 was trained already with 175 billion parameters. Since then, we have seen exponential growth with models like Gopher (280 billion parameters). Megatron-Turing NLG (530 billion) or Google’s GLaM with even 1.2 trillion parameters in its full version.
Consequently, the new recital considers this development and states that "it is appropriate for the Commission to be empowered to update the technical elements of the definition of general-purpose models in the light of market and technological development".
However, in view of these unprecedented dynamics, it must be questioned whether this regulatory approach can keep pace with these technical developments at all. This goes hand in hand with the question of legal certainty if the AI industry regularly has to deal with new criteria from the EU Commission that are highly likely to be outdated by the time the Commission has reached an agreement. For the time being it is fair to say that based on these criteria almost every known general AI model will likely have to be considered to display a "significant generality" and thus fall within the scope of the Act.
Transparency obligations for general-purpose AI models
Falling within the scope of the Act leads to a variety of obligations for general purpose AI providers. They will have to:
- Draw up and keep up to date the technical documentation, including their training and testing processes and the results of their evaluation for the purpose of providing it, upon request, to the competent European and national authorities.
- Draw up, keep up-to-date and make available information and documentation to providers of AI systems who intend to integrate the general-purpose AI model in their AI system.
- Put in place a policy to respect EU copyright law and in particular to identify and respect any reservations rights expressed in the exceptions for text and data mining, unless the rightsholders have expressly reserved by the rightsholders (Art. 4 No. 3 of the European Directive an copyright and related rights in the Digital Single Market).
- Draw up and make publicly available a sufficiently detailed summary of the content used for training of the general-purpose AI model, according to a template provided by the AI Office.
With the obligation to have a copyright policy in place, the AI Act requires non-European general purpose AI providers to comply with European copyright law. This is expressly emphasized in a further recital:
"Any provider placing a general-purpose AI model on the EU market should comply with this obligation, regardless of the jurisdiction in which the copyright-relevant acts underpinning the training of these foundation models take place. This is necessary to ensure a level playing field among providers of general-purpose AI models where no provider should be able to gain a competitive advantage in the EU market by applying lower copyright standards than those provided in the Union."
Use of Model Cards
Interestingly, a further recital is apparently to be added which permits these documentation obligations through the use of "Model Cards". This was already the subject of the non-paper in which Germany, France and Italy proposed a code-of-conduct approach in the final phase of the trilogue, but which has now resulted in a legally binding documentation obligation.
Such Model Cards are by no means a new invention. Rather, model cards are already widely used in the development of AI. This approach was developed and presented back in 2018 in a joint paper by computer scientists and lawyers, which also proposed a framework for the use of model cards. Such model cards are "short documents accompanying trained machine learning models that provide benchmarked evaluation in a variety of conditions, such as across different cultural, demographic, or phenotypic groups and intersectional groups that are relevant to the intended application domains. Model cards also disclose the context in which models are intended to be used, details of the performance evaluation procedures, and other relevant information."
A typical Model Card for an AI model could be structured like this:
1. General Description
- Policies and Instructions
2. Model Details
- Underlying infrastructure
- Design specifications
3. Model Architecture
- Number of parameters
4. Intended Use
- Use cases that were envisioned during development.
- Factors could include demographic or phenotypic groups, environmental conditions, technical attributes, or others.
- Metrics should be chosen to reflect potential real world impacts of the model.
7. Evaluation Data
- Details on the dataset(s) used for the quantitative analyses in the card
- Datasheets describing the training methodologies and techniques
8. Training Data and methodology
- Type and provinence of data, data points
9. Computational Ressources used to train
- Number of floating point operations (FLOPs)
- Known or estimated energy consumption of the model
10. Quantitative Analyses – Unitary results – Intersectional results
11. Responsible AI Considerations
The detailed content of what will be required under all these transparency obligations remains to be seen until the final and detailed version of the AI Act is available. It is planned that the individual elements will be regulated in a new annex to the Act. It will be interesting to see whether and to what extent existing industry standards will be taken into account there. We will keep you posted!
Sign up to our email digest