Regulating the Frontier: Model-Level Evidence from the EU AI Act

Engineering notes

Engineering notes will be added by the aipentium editorial team.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为大语言模型、生成式AI、ChatGPT相关技术、计算机视觉、深度学习等高价值论文补充中文说明。

Original abstract

The European Union's Artificial Intelligence Act creates a binding, economy-wide transparency regime directed specifically at providers of general-purpose AI models. Whether such rules alter model-development practice, rather than merely producing legal documentation outside public view, is an empirical question. This article offers an early model-level assessment using a daily updated public database of 3,571 machine-learning models. The main analysis compares 604 language-based models that satisfy a transparent general-purpose-AI proxy with 401 narrow-task models released during an approximately two-year window from 2 August 2024 to 21 July 2026, when the Act's general-purpose-model obligations began to apply to newly marketed models. A seven-component Structured Disclosure Index records whether public sources report parameter count, training compute, training-data scale, training time, training hardware, training-code status, and model-access status. Organization and calendar-month fixed-effects estimates associate the post-application period with an 8.7 percentage-point increase in structured disclosure for the general-purpose group. The change is concentrated in training-compute reporting. The estimate remains positive in several within-provider specifications, although it becomes small and statistically indistinguishable from zero when organization fixed effects are replaced by country and organization-category controls. No corresponding change is detected in open-weight release, model parameter scale, or disclosed training compute. Only two post-application observations with disclosed compute exceed the Act's $10^{25}$ floating-point-operation presumption, making a threshold design premature. These findings are best read as pre-enforcement evidence of selective documentation adaptation, not a definitive causal estimate. The article contributes a reproducible measurement framework, identifies the distinction between structured and substantive transparency, and provides code for extending the analysis to versioned Hugging Face model cards, safety reports, and copyright disclosures.

Links and sources

Comments