Study tests how prompts shape LLM-generated EU compliance documents
Researchers benchmarked large language models on producing Digital Product Passports and Data Protection Impact Assessments, finding prompt context and rule strictness affect output quality.
In plain words
EU rules like the ESPR and GDPR require companies to produce documents proving compliance, and some researchers propose using large language models to draft them. A new study tests how data extraction instructions and vague regulations affect the quality of those AI-drafted documents. Looser rules made outputs less consistent unless prompts included more context, while stricter formats were consistent but more prone to invented content.
What was announced
The arXiv paper examines two EU compliance artifacts: Digital Product Passports under the Ecodesign for Sustainable Products Regulation, which demand life cycle transparency, and Data Protection Impact Assessments required by the GDPR. The authors note that industrial data needed for these documents sits in heterogeneous formats scattered across company and supplier systems, and DPIA drafting requires interdisciplinary expertise without a standardized format. They investigate how data extraction instructions and regulatory vagueness affect LLM-produced artifacts, benchmarking different models against manually created ground-truth schemas.
The results split by rule strictness. Less strict guidelines, such as DPIA formatting, needed higher-context prompts to keep outputs consistent and complete. Stricter guidelines, such as Digital Battery Passport formatting, produced consistent results regardless of prompt context but may lead to more hallucinations in the output. The paper is a v2 replacement, and the available abstract does not report which models were tested or the size of any quality gaps.
What it does not show
- Evidence is limited to the paper's abstract, so detailed methods, model names, and quantitative results are not available for verification.
- The finding that stricter formats may increase hallucinations is stated as a possibility in the abstract, not a confirmed measurement.
Sources
Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 17 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.