Whilst Austria is behind schedule with its national AI Act – in particular with regard to the appointment of the future AI authority or authorities and the implementation of sanctions – the obligations for providers of general-purpose AI models (“General Purpose AI” or “GPAI models”) are due to come into force on 2 August 2025 as planned. These obligations apply not only to manufacturers of GPAI models, but also to operators who integrate and modify these models into their systems. As a result, they may assume the role of a “downstream provider”. In such cases, they must also comply with the requirements set out in Article 51 et seq. of the AI Regulation. The European Commission provides supplementary guidance on this in its recently finalised guidelines on GPAI (see here) and in the Code of Practice.
Key aspects of the GPAI for businesses and relevant organisations (e.g. research, education and the public sector), as well as the implications of the lack of a national legal basis for imposing fines under the AI Regulation:
Identifying GPAI models
In order to assess the relevance of the new obligations to their operations, GPAI models must be identified as part of the AI inventory analysis. Not every model within an AI system is subject to regulation. According to Article 3(63) of the AI Regulation, GPAI models demonstrate significant general applicability and are capable of competently performing a wide range of different tasks, regardless of the manner in which they are placed on the market. Another characteristic is that they can be integrated into a wide range of downstream systems or applications. If models have been developed for specific purposes, they cannot be classified as GPAI. This must be critically assessed from a technical perspective in practice.
According to the Explanatory Notes, a model is presumed to meet the GPAI requirements if the AI model has at least one billion parameters. Building on this, the European Commission has drawn up indicative criteria based on the model’s training calculations. According to these criteria, a model is considered to be a GPAI model if
a training calculation value of 10²³ FLOP is exceeded and
language-based, text-to-image or text-to-video outputs are generated.
It is questionable that, according to the guidelines, the output is decisive for classification and that, for example, models for chess and computer games or for weather forecasts are to be excluded per se. After all, the statutory definition does not focus on the results of the work carried out. Regardless of this, however, the examples provide a good point of reference for better identifying and classifying GPAI.
To make it easier to carry out the inventory, those subject to the obligation can use the form of provision as a guide. For example, GPAI models are regularly made available via repositories (e.g. Hugging Face, GitHub), application programming interfaces (APIs), by direct download or as a physical copy.
Once GPAI models have been identified, the question arises as to whether they pose systemic risks and, consequently, whether the extended provider obligations apply. Classification is assumed under Article 52(2) of the AI Regulation if the cumulative volume of computations used for its training, measured in floating-point operations, exceeds 10²⁵ FLOPs. Furthermore, the Commission may also designate models as GPAIs posing systemic risks.
Provider Obligations
In accordance with Article 53 et seq. of the AI Regulation, providers of GPAI models must now, amongst other things, meet the following requirements:
- Documentation: technical documentation for each individual model, including its training and testing procedures and the results of its evaluation throughout its lifecycle.
- Provision of information for providers of AI systems: To facilitate integration for providers of AI systems, so that they can fully understand the capabilities and limitations of the model and fulfil their obligations.
- Copyright compliance: Implementing a strategy to ensure compliance with EU copyright law.
- Transparency: Publication of a summary of the training content used. A template is provided for this purpose.
Providers of GPAI models that pose systemic risks are subject to additional obligations: risk analyses and model evaluations, reporting of serious incidents to the AI Office, and ongoing cybersecurity measures.
However, under certain conditions, there are exceptions to the scope of application and the obligations for GPAI providers in the case of open-source software, provided that such software does not pose any systemic risks.
From downstream modifier to downstream provider
If GPAI models are substantially modified by downstream actors who integrate the model into their own systems, they become downstream providers. However, in line with the Blue Guide, the Guidelines make it clear that not every modification is relevant. It must be “substantial” or “significant”. This can, for example, also be achieved through fine-tuning of the model.
In addition, the new GPAI guidelines provide guidance on distinguishing between immaterial and material changes using indicative thresholds:
It will use more than a third of the training data for the original model;
Where the training effort required for the original model is neither known nor reliably estimable, the following indicators may be used: if the original model is a GPAI model involving systemic risk, one-third of the threshold for AI models involving systemic risk should be applied, which currently amounts to one-third of 10²⁵ FLOP. For GPAI models without systemic risk, one-third of the general indicative threshold should be applied, which is currently one-third of 10²³ FLOP.
Caution is therefore advised when making in-house adjustments and customisations. Hasty modifications to licensed software – including open-source software, unless specifically exempted – may also result in a new role under the AI Regulation.
No enforceable fines, reduced liability?
Furthermore, the fines under the AI Regulation came into force on 2 August 2025. An exception applies to GPAI models; the corresponding fines for breaches of the provisions that have only now come into force may only be imposed from 2 August 2026. As the powers of the national authorities and the accompanying criminal and procedural provisions have not yet been established, no penalties can currently be imposed in Austria for breaches of the AI Regulation, in particular the prohibitions under Article 5 of the AI Regulation that are already in force (up to EUR 35 million or 7 per cent of global annual turnover). However, this does not constitute a free pass. In the event of a breach of applicable laws, regulations and directives, the civil liability of the company and its authorised bodies – including under the Unfair Competition Act – remains unaffected.
An overview of GPAI compliance tasks:
Identify GPAI models and roles
Classify GPAI models
Question significant modifications to GPAI models
Fulfil obligations as a provider of the developed model or as a downstream provider for the modified part of the model
Documentation of models, roles and obligations
Ongoing re-evaluation of obligations and updating of documentation throughout the entire lifecycle