General-purpose AI refers to AI systems, particularly foundation models and large language models, capable of performing a wide range of cognitive tasks across diverse domains without being designed for a specific narrow application. The term carries weight in three contexts: a technical sense (models trained on broad datasets that develop emergent capabilities across tasks they were not explicitly trained for); a regulatory sense (the EU AI Act creates a specific "GPAI model" category with its own obligations, regulating AI models directly rather than only their applications); and a governance sense (the risks of GPAI models emerge from capabilities and downstream uses that cannot be fully anticipated at training time).
EU AI Act definition and obligations
Under Regulation 2024/1689, a GPAI model is defined as an AI model trained with a large amount of data using self-supervision at scale, displaying significant generality, capable of being integrated into a variety of downstream systems or applications.
Standard GPAI obligations under Articles 53–56 cover technical documentation disclosing training methodology, evaluation results, and known limitations; a copyright compliance policy for EU Directive 2019/790; and a publicly available summary of training data content.
Models trained with cumulative compute exceeding 10^25 FLOP are classified as GPAI with systemic risk and face enhanced obligations: adversarial testing and red-teaming before deployment, serious incident reporting to the EU AI Office, cybersecurity measures, energy efficiency disclosures, and cooperation with model evaluation.
As of 2024, only GPT-4 class models exceeded the 10^25 FLOP threshold. GPT-5 and Anthropic's Opus/Mythos class also exceed it. (Sources: EU AI Act (Regulation 2024/1689), A Taxonomy of Systemic Risks from General-Purpose AI)
Compliance instruments
The GPAI obligations are operationalized through two Commission instruments. The EU General-Purpose AI Code of Practice, published July 10, 2025 by the AI Office and endorsed via adequacy decisions on August 1, 2025, is the voluntary instrument through which providers can demonstrate compliance with Articles 53 and 55; the Commission's signatory list as updated April 23, 2026 names 23 full signatories, including Amazon, Anthropic, Google, IBM, Microsoft, Mistral AI, and OpenAI, while xAI signed only the Safety and Security chapter (Source: digital-strategy.ec.europa.eu). The Commission's Guidelines on the scope of obligations for providers of general-purpose AI models (last updated March 26, 2026), summarized at EU AI Office — GPAI Provider Guidelines and Enforcement Framework, are the interpretive layer: who counts as a provider, what counts as a significant modification, and how open-source exemptions apply. Commission enforcement powers over GPAI obligations, including fines, activate August 2, 2026 (Source: digital-strategy.ec.europa.eu).
Governance challenge
GPAI models present a governance gap that earlier AI regulation did not anticipate. Application-level governance addresses it only partly: a GPAI model might be used for medical diagnosis (high-risk under the EU AI Act) or for writing emails (minimal risk), so regulating only at the application level leaves systemic risks at the model level unaddressed, and a malicious actor can use a general-purpose model for harm regardless of its intended application.
A second factor is that capabilities are emergent. GPAI models develop capabilities their developers did not explicitly train for, including capabilities that create risk, which makes pre-deployment risk assessment inherently incomplete. The scheming, emergent misalignment, and biosecurity uplift risks all emerge from GPAI capabilities rather than from narrow application design. (Sources: Open Problems in Technical AI Governance, We Need a Science of Scheming)
A third factor is fast-follow capability transfer: once a GPAI model is trained, its capabilities can be distilled into smaller, more easily deployable models. Restricting frontier GPAI access while smaller models replicate key capabilities remains an unresolved problem. (Source: epochai.substack.com)
Open-source dimension
The open-weight model debate is partly a GPAI governance debate. When a GPAI model such as Llama-3 or DeepSeek-V3 is released as open-weight, its compute can fall far below the 10^25 FLOP systemic risk threshold while the model still has significant capabilities, so the EU AI Act's threshold approach may fail to capture the risk from open-weight GPAI models trained below the threshold. (Source: Open-Source AI / Open-Weight Models) Meta's Llama series and DeepSeek's open-weight releases represent GPAI deployment without the governance obligations that apply to frontier closed models. The capability distance at issue has been measured: UK AI Security Institute research reported in July 2026 found open-weight models, including GLM-5.2 and DeepSeek V4-Pro, trailing frontier closed models by four to seven months on cyber capabilities (Source: transformernews.ai). Frontier-scale open-weight releases continued through mid-2026 with Moonshot AI's 2.8-trillion-parameter Kimi K3 and Thinking Machines Lab's Inkling (see Open-Weight Frontier Models).
GPAI versus narrow AI as a spectrum
The Jeon (2025) standardization survey models AI as a four-level progression: (1) narrow AI and LLMs (language models for specific NLP tasks); (2) multimodal LLMs (text, image, and audio, still narrower in scope); (3) GPAI/AGI (able to understand and solve a variety of problems like humans); and (4) ASI (surpassing humans in all intellectual tasks). Most frontier models, including Claude Opus 4.6, GPT-5.x, and Gemini 3, sit in category 3 (GPAI/AGI) or approach it, while current standards and regulations are calibrated for category 2 (the multimodal-LLM level), leaving a governance gap. (Source: Standardization Trends on Safety and Trustworthiness Technology for Advanced AI)
Debates and tensions
The EU and US diverge in their regulatory approach. The EU Act is the only framework that regulates based on the model itself, through the GPAI category and compute threshold, whereas all US regulatory approaches regulate at the application level. This divergence produces (1) a different accountability architecture; (2) the potential for regulatory arbitrage (deploying in the US what cannot be deployed in the EU); and (3) different innovation incentives.
Threshold calibration is contested. The 10^25 FLOP threshold is a single bright line that GPAI developers are approaching from below and above, and it does not scale with capability improvements: a model trained with 10^24 FLOP today may match the systemic risk of a 10^25 FLOP model from 2023.
Certification also lags. Unlike hardware safety standards, where certification bodies test physical products, GPAI certification requires running evaluations on models whose capabilities are not fully characterized, and there is no equivalent of crash-testing for GPAI systemic risk.
See also
- EU AI Act — core regulatory definition and obligations
- GPAI Code of Practice — the voluntary compliance instrument for Articles 53 and 55
- AI Office GPAI Guidelines — the Commission's interpretive layer on provider scope
- Taxonomy of Systemic Risks from GPAI — academic mapping of 13 risk categories following EU Act definitions
- Open Problems in Technical AI Governance — governance challenges specific to GPAI
- Standardization Trends — where international standards bodies are on GPAI
- Open-Source AI — the open-weight dimension of GPAI governance