GPAI (General-Purpose AI) — Definition, Obligations & Examples

Key Takeaway: GPAI (General-Purpose AI) is the EU AI Act's formal term for AI models such as GPT-4, Claude, and Gemini — capable of performing a broad range of tasks rather than a single defined function. From 2 August 2025, GPAI model providers face binding obligations under Title VIII of the Act, regardless of where they are headquartered.

What Is GPAI?

GPAI (General-Purpose AI) is an AI model trained on large amounts of data at scale that displays significant generality and can competently perform a wide range of distinct tasks. Under Article 3(63) of Regulation (EU) 2024/1689 — the EU AI Act — a GPAI model is specifically defined by its breadth: it is not a system designed and optimised for a single, narrow use, but one that can be integrated into a wide variety of downstream applications across sectors.

The term "general-purpose" is regulatory shorthand for what the industry more commonly calls foundation models or large language models (LLMs). OpenAI's GPT series, Anthropic's Claude, Google DeepMind's Gemini, Meta's Llama, and Mistral's open-weight models are all GPAI models within the Act's scope. The definition captures both text-based and multimodal systems, provided they exhibit the requisite generality of function.

Why It Matters

GPAI models sit at the top of the AI supply chain. When enterprises integrate a GPAI model API into their products or workflows, they are building on infrastructure that the EU regulator has specifically identified as posing systemic risk if not properly documented, tested, and governed. Enterprises that rely on GPAI models inherit compliance exposure: if the provider has not met its Article 53 obligations, the downstream deployer cannot rely on that model as a compliant component in any high-risk application.

The August 2025 applicability date for GPAI obligations is the most immediate compliance deadline in the EU AI Act timeline — it precedes the full general application date of 2 August 2026.

Core GPAI Obligations Under the AI Act

The Act establishes a two-tier framework for GPAI models based on systemic risk.

All GPAI model providers must (Article 53):

  • Prepare and maintain technical documentation covering training methodology, data sources, computational resources used, and evaluation results
  • Produce and make available a summary of the content used to train the model, in sufficient detail to enable copyright compliance assessment under the Text and Data Mining Directive (Directive 2019/790)
  • Implement a policy to comply with EU copyright law, including respecting opt-outs by rights holders
  • Make available to downstream providers all information necessary for those providers to comply with their own obligations under the Act

These obligations apply from 2 August 2025 for newly released models, and providers had until 2 August 2025 to bring models already on the market into compliance.

GPAI models designated as posing systemic risk face additional obligations (Articles 55–56):

The systemic risk designation applies to models trained with compute exceeding 10²⁵ FLOPs — a threshold intended to capture frontier models with broad societal impact potential. For these models:

  • Adversarial testing and red-teaming must be conducted before and after release
  • Serious incidents and near-misses must be reported to the European AI Office
  • Cybersecurity measures adequate to the model's systemic risk profile are mandatory
  • Providers must assess and mitigate potential systemic risks to democratic processes, critical infrastructure, and fundamental rights at the EU level

The European AI Office, established within the European Commission, supervises GPAI providers directly — distinct from the national market surveillance authorities that supervise other AI system categories.

Foundation Model vs GPAI vs Frontier Model

These three terms are frequently conflated but carry different connotations:

  • GPAI model is the EU legal term. It refers to any AI model meeting the definition in Article 3(63), regardless of size or capability.
  • Foundation model is the ML research and industry term for large pre-trained models that serve as a base for fine-tuning and downstream applications. Substantially overlaps with GPAI, though the legal definition is the controlling one.
  • Frontier model is a non-regulatory term referring to the most capable, most recently released models at the edge of what is technically achievable. Frontier models typically fall within the systemic risk tier under the Act, but frontier status is not itself a legal category.

For compliance purposes, the GPAI definition is what governs. An organisation's legal and compliance team should assess any AI model it uses or provides against the Article 3(63) criteria, not against informal industry labels.

Knowlee and GPAI Compliance

Knowlee's governance scaffold maps directly to the GPAI compliance layer. For every AI model used in production workflows, Knowlee records the model identity, version, provider, and documented compliance status in job-level metadata. When a workflow invokes a GPAI model, that invocation is captured in the audit trail with timestamp and context — enabling deployers to demonstrate, in any regulatory review, which model was used, when, and under what governance conditions. This operational record satisfies the deployer's downstream documentation obligations under Article 26 without requiring the enterprise to build a separate compliance system from scratch.

Related Terms