Work

Mirror Mirror on the wall, is this GPAI model in the AI Act's Scope at all?: Necessary clarifications for establishing regulatory certainty and mitigating evasion

Year: 2026
Type: preprint
Abstract: The regulation of General-Purpose AI (GPAI) Models under the AI Act was a late addition in the drafting process owing to its rapid development and initial focus on regulating `AI systems', as well as achieving a tricky and negotiated balance between precision and future-applicability. Unlike AI systems, GPAI models are not required to declare conformity or labelling indicating their status under the AI Act, and are instead regulated by meeting three criteria: being an AI model, showing generality, and being placed on the EU market. This has created a confusing and potentially unenforceable ecosystem where more AI models are being released and utilised everyday within the EU with little understanding of their status under the AI Act. At the same time, with the start of enforcement from 2nd August 2026, evidence is emerging of providers seeking to strategically place their models outside the scope of the AI Act. To address these challenges, we present a framework enabling step-by-step analysis of the AI Act's scope of GPAI models, including whether models meet criteria for exemption. We present a classification of 3,706 models from Epoch AI and identify models potentially subject to the AI Act, including an analysis of generality and capability as well as placement on the market. We also discuss our framework and the challenges of applying the AI Act for new model architectures such as decision models and sound-only models, hypothetical scenarios that can act as edge-cases, and address claims of non-applicability by certain providers regarding their released models. We believe our work provides necessary clarity for institutions such as the AI Office to enable effective enforcement of the AI Act, enable accurate assessment of obligations such as Article 53 transparency measures, and provides directions for future policy-making to deal with the scale and scope of AI models.
Language: English
Subfield: Law
Open Access status: green
OpenAlex ID: W7225636386
Cites:
Cited by:
Related to: 10