> ## Content Index > Fetch the complete content index at: https://www.notatechguy.com/llms.txt > Use this file to discover other available public pages before exploring further. # LLMs hallucinate more under strict EU rules, study finds - URL: https://www.notatechguy.com/llms-hallucinate-more-under-strict-eu-rules-study-finds/ - Published: 2026-08-25T19:46:51.000Z - Updated: 2026-08-25T19:46:49.000Z - Description: Purdue researchers found stricter EU regulatory formatting makes LLMs hallucinate more, while vaguer rules need heavier prompts for consistent output. - Author: Marcello Babbili - Tags: Technology & AI, AI Models A preprint from Purdue University, published 21 August on arXiv, finds that stricter regulatory formatting guidelines cause large language models to hallucinate more, not less [S¹](https://arxiv.org/abs/2608.21317v1?ref=notatechguy.com)[P²](https://arxiv.org/html/2608.21317v1?ref=notatechguy.com). The finding upends the assumption that rigid templates keep AI-generated compliance documents on a short leash. It raises an uncomfortable question for any company hoping to automate EU Digital Product Passports or GDPR impact assessments: which kind of rule is safer to hand to a model, and what breaks when you do? **My read:** This is the first study I've seen that systematically pits regulatory strictness against LLM reliability, and the result is genuinely counterintuitive. I'd expect rigid templates to constrain a model's output. Instead, the paper suggests they constrain it in the wrong direction, toward confident invention. I don't buy the implication that vaguer rules are strictly better. The study itself says they need much heavier prompt context to stay consistent. The real signal is that neither extreme works well unsupervised, and the gap between benchmark schemas and live industrial data is where this will actually fall apart. ## Why strict templates backfire The EU has positioned itself as the world's most aggressive regulator of sustainability and privacy [S¹](https://arxiv.org/abs/2608.21317v1?ref=notatechguy.com). Two frameworks sit at the centre of this paper. The Ecodesign for Sustainable Products Regulation (ESPR) introduces Digital Product Passports, a kind of digital birth certificate tracking a product's full life cycle [S¹](https://arxiv.org/abs/2608.21317v1?ref=notatechguy.com). The General Data Protection Regulation (GDPR) requires Data Protection Impact Assessments (DPIAs), documents that identify and mitigate privacy risks before a system is deployed [S¹](https://arxiv.org/abs/2608.21317v1?ref=notatechguy.com). Both demand paperwork that is expensive to produce. Industrial data lives in scattered, incompatible formats across companies and suppliers, making it hard to extract into compliant Digital Product Passport formatting [S¹](https://arxiv.org/abs/2608.21317v1?ref=notatechguy.com). DPIA documents need interdisciplinary expertise and follow no standard template, so building one for a novel system is slow and inconsistent [S¹](https://arxiv.org/abs/2608.21317v1?ref=notatechguy.com). The promise of LLMs is obvious: feed the model the raw data and the regulation, and get a compliant artifact back. Researchers have proposed exactly that. But as the paper notes, nobody had rigorously tested what happens when the messiness of real industrial data and the vagueness of some regulations collide with model output [S¹](https://arxiv.org/abs/2608.21317v1?ref=notatechguy.com). ## The paradox at the centre The authors benchmarked multiple LLMs against manually created ground-truth schemas for two compliance artifacts: Digital Battery Passports (a stricter, more structured format) and DPIAs (a looser, less standardised format) [S¹](https://arxiv.org/abs/2608.21317v1?ref=notatechguy.com). The results cut in opposite directions. For DPIAs, the less strict guideline, models needed higher-context prompts to maintain consistency and completeness [S¹](https://arxiv.org/abs/2608.21317v1?ref=notatechguy.com). Give the model more background and it stays on track. Strip the context away and it wanders. For Digital Battery Passports, the stricter guideline, models produced consistent results regardless of how much context the prompt provided [S¹](https://arxiv.org/abs/2608.21317v1?ref=notatechguy.com). That sounds like a win. But the paper found that the stricter format led to more hallucinations in the output [S¹](https://arxiv.org/abs/2608.21317v1?ref=notatechguy.com). Rigid templates make models confident and consistent, and consistently wrong in a specific way. They fill gaps with invented details that fit the format so neatly they are hard to catch. This connects to a broader pattern. Confident invention is not unique to compliance work. ## What to do about it For a manufacturer preparing for ESPR's Digital Product Passport requirements, the takeaway is not to avoid LLMs. It is to never ship the raw output. Consider a mid-size battery maker in Stuttgart that needs to file Digital Battery Passports for every cell it sells in the EU. An LLM can draft those passports quickly and in the correct format. But the stricter the template, the more the model may invent values for fields it cannot find in the supplier data: a recycled-content percentage, a cobalt provenance claim. It presents them with the same confidence as a real figure. A human reviewer who sees a well-formatted, internally consistent document may not think to check whether the 12% recycled cobalt figure came from a supplier record or the model's imagination. The practical step: treat LLM-generated compliance artifacts as first drafts, and build a review checkpoint that flags every numeric or factual claim against the source data before submission. If the source data does not contain a field, the model's output for that field is fiction until proven otherwise. ## What we don't know yet The study benchmarks models against manually created ground-truth schemas, not against live regulatory submissions or real-world heterogeneous industrial data [S¹](https://arxiv.org/abs/2608.21317v1?ref=notatechguy.com). A schema is a clean abstraction. A supplier's spreadsheet with missing columns and conflicting units is not. The gap between the two is where compliance automation will live or die. The paper is a preprint that has not been peer-reviewed [S¹](https://arxiv.org/abs/2608.21317v1?ref=notatechguy.com), so the findings are preliminary. It also focuses exclusively on EU regulations, offering no analysis of how these dynamics would play out under Australian or US frameworks. The deeper question is whether any prompting strategy can close the hallucination gap for strict-format compliance, or whether the answer is architectural: smaller, specialised models constrained to extract rather than generate. The next signal: peer review of this preprint and the authors' planned testing against live industrial data, both flagged as future work in the paper. We'll check the hallucination claim against both when they land. Subscribe for more reporting on what AI actually does, not what the demo promises. --- *Sources: [S1 — From Regulation to Implementation: A Critical Evaluation of LLM-Assist](https://arxiv.org/abs/2608.21317v1?ref=notatechguy.com) · [P2 — From Regulation to Implementation: A Critical Evaluation of LLM-Assist](https://arxiv.org/html/2608.21317v1?ref=notatechguy.com) · [P3 — open-compass/CriticEval](https://github.com/open-compass/criticeval/?ref=notatechguy.com) · [P4 — De Jure: Iterative LLM Self-Refinement for Structured Extraction of Re](https://arxiv.org/html/2604.02276?ref=notatechguy.com) · [P5 — alektebel/regllm](https://github.com/alektebel/regllm?ref=notatechguy.com)* ## More from Not A Tech Guy - [SRPO trains Qwen3-8B to fix its own errors, hits 73.3% on AIME'24](https://www.notatechguy.com/srpo-trains-qwen3-8b-to-fix-its-own-errors-hits-73-3-on-aime-24/) - [Distilled AI safety guard runs on CPU in 24ms, matches teacher](https://www.notatechguy.com/distilled-ai-safety-guard-runs-on-cpu-in-24ms-matches-teacher/) - [MiroFish: 71,000-star AI prediction engine hits GitHub trending](https://www.notatechguy.com/mirofish-71-000-star-ai-prediction-engine-hits-github-trending/) --- *Generated from an audited evidence pack with primary-source research. Social-media items are discussion signals, not verified facts. Nothing here is financial, legal or medical advice.*