Understanding GPAI Models

Chapter V of the AI Act introduces a distinct regulatory framework for General-Purpose AI (GPAI) models. This represents a significant regulatory innovation—recognising that foundation models and large language models present unique governance challenges due to their versatility and potential for widespread downstream use.

Legal Definition of GPAI Models

Article 3(63) Definition

"General-purpose AI model" means an AI model, including where such an AI model is trained with a large amount of data using self-supervision at scale, that displays significant generality and is capable of competently performing a wide range of distinct tasks regardless of the way the model is placed on the market and that can be integrated into a variety of downstream systems or applications, except AI models that are used for research, development or prototyping activities before they are placed on the market.

Breaking Down the Definition

The "Foundation Model" Relationship

While the AI Act uses "GPAI model," this concept aligns closely with what the AI community calls "foundation models"—large-scale models trained on broad data that can be adapted to many downstream tasks. Key characteristics:

Examples: GPAI vs. Non-GPAI

Clear GPAI Model Examples

NOT GPAI Models

Edge Cases Requiring Analysis

⚠️ Classification Challenge: The boundary between "general-purpose" and "specific" is not always clear. When uncertain, assess whether the model displays "significant generality" for a "wide range of distinct tasks."

GPAI Model Provider Definition

GPAI Model Provider

The concept of a GPAI model provider is not defined in a single Article 3 definition. Rather, it derives from the obligations set out in Articles 53-55, which establish the responsibilities of natural or legal persons, public authorities, agencies, or other bodies that develop a general-purpose AI model or that have a general-purpose AI model developed and place it on the market. Note: Article 3(66) defines "general-purpose AI system," not the provider of a GPAI model.

When Are You a GPAI Provider?

Free and Open Source Considerations

Article 53(2) addresses open source GPAI:

💡 Open Source Distinction: Open source GPAI benefits from reduced requirements, but this exemption does not apply if the model presents systemic risk (e.g., exceeds 10^25 FLOPS threshold).

The GPAI-AI System Distinction

Different Regulatory Frameworks

How They Interact

GPAI Model (e.g., GPT-4)
         ↓
    [Integration]
         ↓
Downstream AI System (e.g., HR screening tool using GPT-4)
         ↓
    [If high-risk use case]
         ↓
High-Risk AI Obligations Apply (to downstream provider)

Key Principle: GPAI model obligations apply to the model provider. If the GPAI is integrated into a high-risk AI system, the downstream provider bears high-risk compliance obligations.

Cumulative Obligations Example

GPAI Model Classification Framework

Classification Decision Tree

Step 1: Does the AI model display "significant generality"?

• If No → Not GPAI → Standard AI system rules apply
• If Yes → Proceed to Step 2

Step 2: Is it "capable of competently performing a wide range of distinct tasks"?

• If No → May not be GPAI → Analyse further
• If Yes → Proceed to Step 3

Step 3: Can it be "integrated into a variety of downstream systems or applications"?

• If No → May not be GPAI → Analyse integration potential
• If Yes → GPAI Model → Proceed to Step 4

Step 4: Does it present systemic risk (>10^25 FLOPS or Commission designation)?

• If Yes → GPAI Model with Systemic Risk → Enhanced obligations
• If No → Standard GPAI Model → Baseline obligations

Classification Documentation

When classifying a model, document:

• Model architecture and training approach
• Range of tasks the model can perform
• Integration capabilities and APIs
• Computational resources used in training
• Reasoning for GPAI/non-GPAI determination
• Date of classification assessment
• Plan for re-assessment if model capabilities change

Compliance Timeline