With LARA, the regulatory risk of LLMs becomes an audit item for DPOs

Create a monitoring on this topic Receive upcoming developments if any, with an AI monitoring configured from this article.

Create my free AI monitoring

None of the twelve major AI models evaluated meet the compliance threshold for GDPR and the AI Act, and the risk that previously weighed on integrators is now measurable, documentable, and disputable. This is what emerges from the LARA (Legal Assessment for Real-world Agents) benchmark, published on May 27, 2026 by the non-profit research foundation Aithos. The verdict is clear: none of the evaluated models fully meet the EU regulatory requirements, and some violate European rules in up to 93% of the studied cases. Among the twelve tested models, Claude Opus 4.1 from Anthropic achieves the best score, around 54% compliance, while Google's Gemini 3.1 Pro drops to 10%. The findings come ten months after the AI Act obligations for general-purpose model providers came into effect on August 2, 2025. Article 25 of the regulation, which came into force on August 1, 2024 and became fully applicable on August 2, 2026, had already established the integrator's joint liability in the texts: what LARA changes is the shift from theoretical risk to model-specific quantified risk.

A grid of 3,000 scenarios, twelve models, few survivors

LARA is based on more than 3,000 scenarios covering the main risks addressed by the GDPR and the AI Act. The protocol places each model in agentive situations where it must navigate legal or regulatory dilemmas: obtaining consent before collection, avoiding psychological profiling, refusing to manipulate a user, protecting vulnerable audiences, ensuring transparency in automated decisions. The most frequently documented violations concern the use of personal data without a clear legal basis, the implicit creation of psychological profiles and behaviors that may influence or manipulate users. The rating is severe: between the top (Claude Opus 4.1) and the bottom (Gemini 3.1 Pro), models from OpenAI, Meta, Mistral AI, xAI, and DeepSeek also show significant non-compliance rates in several categories. Aithos, a non-profit foundation of which Daan Henselmans is the research director, published the study on its own platform Substack on May 27, 2026; the work has not been peer-reviewed and has not received formal regulatory accreditation. The author's verbatim summary bluntly states the conclusion: "No leading model achieves acceptable levels of compliance with the AI Act and European privacy legislation" (free translation). The source article does not specify the date of the tests or the exact versions of the models evaluated beyond the two named — a methodological shadow zone that the tool's public documentation has not yet clarified.

GDPR / AI Act Compliance — Aithos LARA benchmark (May 2026, 3,000+ scenarios)

Model	Publisher	Compliance Rate
Claude Opus 4.1	Anthropic	~54%
Gemini 3.1 Pro	Google	~10%
OpenAI, Meta, Mistral AI, xAI, DeepSeek Models	Diverse	Not published individually

None of the 12 tested models fully meet European requirements. Some violate rules in up to 93% of cases. Source: Aithos LARA benchmark, May 2026.

Article 25 AI Act: the integrator becomes a co-supplier, and already knew that

The central legal mechanism that LARA makes operational is not new. Article 25 §1 of the AI Act, enacted in the regulation that came into force on August 1, 2024 and applicable from August 2, 2026, provides that a deployer who modifies the intended use of a high-risk AI system or applies its own brand to it becomes a co-supplier. Specifically, this classification triggers the direct application of Article 16, which defines the core obligations of the supplier: risk management, technical documentation, post-market surveillance. A DPO integrating a general-purpose model into a high-risk use inherits, through this mechanism, the same due diligence regime as the upstream publisher.

Above this integrator-supplier joint liability, the regime specific to general-purpose model suppliers applies in parallel. Article 53 sets since August 2, 2025, the specific obligations for providers of general-purpose AI models: documentation, copyright policy, summary of training data. In case of non-compliance, Article 101 of the regulation caps fines for these same providers at 3 % of global annual turnover or 15 million euros, the higher amount being applied - a sanctioning power that the Commission will however only exercise from August 2, 2026.

Aithos emphasizes the integrator link: "companies that build their own AI agents in addition to these models could also be held legally responsible" (free translation). The foundation adds, with an assumed methodological reserve, that "GenAI systems are not yet ready to be deployed in highly regulated environments without additional controls and safeguards" (free translation). Joint liability existed in the texts; LARA quantifies liability per model - a unit of analysis that legal departments had not yet had in this form, usable as an audit preparatory element. It is this shift, from theoretical risk to documented risk, that changes the situation for legal departments, more than the score itself.

Precedents, effective sanctions, and the compliance layer that is financed

The LARA assessment arrives in a tense regulatory landscape. In Italy, the Data Protection Authority imposed a fine of 15 million euros on OpenAI in December 2024 for processing personal data through ChatGPT, according to the information published at the time, as part of the procedure initiated in 2023 - ActuIA had already documented the initial restriction of the service in Italy. This penalty was annulled on appeal by the Rome Tribunal in the first quarter of 2026 on procedural grounds: the Irish Data Protection Commission having become the lead authority for OpenAI since February 2024, the Data Protection Authority no longer had jurisdiction under the GDPR one-stop shop. The probative basis remains, however, entirely open. On the providers' side, Meta had preferred to suspend the European launch of its multimodal Llama 3 models rather than face the trio DMA/RGPD/AI Act - a signal that geographic avoidance remains an option for actors who consider exposure too costly. On the other end, the market is preparing: the startup ZeroDrift, which develops compliance middleware positioned between LLMs and end-users, announced, according to its own communications, a raise of 10 million dollars early in June 2026, a few days after the publication of LARA. As of June 8, 2026, none of the named providers - Anthropic, Google, OpenAI, Meta, Mistral AI, xAI, DeepSeek - had publicly contested the Aithos methodology, claimed a score, or proposed an alternative protocol. The lock is in the silence of the text: the AI Act does not specify the distribution of the probative burden between the general model provider and the deployer when a damage is proven to an end-user. It is in this vacuum that the first cases brought before national regulatory authorities, starting from August 2, 2026, will have to decide.