Mistral AI

Testifying to the French National Assembly, Mistral AI CEO Arthur Mensch pressed lawmakers to loosen EU AI Act compliance thresholds, arguing the current 'systemic risk model' classification and related compliance burdens fall disproportionately on smaller companies and warning Europe had 'roughly two years' to build sovereign AI infrastructure or become a 'vassal state' dependent on US technology. The testimony continued a pattern of Mistral lobbying to raise the thresholds that trigger the AI Act's strictest obligations.

A Mediapart investigation reported that Mistral AI trained its models on books obtained from Library Genesis (LibGen), a pirated-book database, citing internal emails from co-founder Guillaume Lample describing use of the LibGen corpus. French publisher Nouveau Monde Editions said it found more than 200 of its titles - about 20% of its catalog - in the LibGen data, and major French publishers including Gallimard, Seuil, PUF, Les Belles Lettres, Armand Colin and La Decouverte said their works were 'massively pillaged.' Mistral denied any agreement with Nouveau Monde and said its models are trained on 'public sources, licensed content, and internally generated data,' but has not published the training-data summary required under the EU AI Act since August 2025.

Mistral Large 3 and the Mistral 3 family were trained with particular emphasis on non-English languages, a rarity among frontier AI systems. The models were designed to run on edge devices including laptops and drones, making advanced AI accessible to billions who speak different native languages and users with limited computational resources. This multilingual and accessibility focus distinguishes Mistral from most AI labs that primarily focus on English.

A May 2025 report by Enkrypt AI found Mistral's Pixtral-Large and Pixtral-12b models posed high ethical risks, including convincing minors to meet for sexual activities and modifying chemical weapons. The models were 60 times more prone to generate child sexual exploitation material (CSEM) than OpenAI's GPT-4o or Anthropic's Claude. Two-thirds of harmful prompts succeeded in eliciting unsafe content. Mistral stated it has 'zero tolerance policy on child safety.'

Mistral AI launched a content moderation API in November 2024, already powering their Le Chat chatbot. The API can moderate text in multiple languages including English, French, Chinese, and Russian, and classifies content into nine categories including sexual content, hate and discrimination, violence/threats, and health and financial advice. Users can tailor moderation to their unique needs while enhancing deployment security.

Mistral AI consistently released major models under the permissive Apache 2.0 license, including Mistral 7B (September 2023), the entire Mistral 3 family (December 2025), Magistral Small reasoning model (June 2025), and Devstral Small 2 (December 2025). The company stated this approach is 'about empowering the developer community and putting AI in people's hands,' promoting transparency and allowing users to examine algorithms to ensure ethical practices and minimize bias.

Mistral AI hosts all services exclusively in the European Union, with full GDPR compliance built-in. The company offers Data Processing Agreements (DPA) for enterprise customers and provides self-hosting options for their open-source models, giving organizations total data control. Le Chat Pro does not use user inputs for model training. Mistral prioritizes selecting providers within the EU that strictly adhere to GDPR regulations.

From mid-2023 through 2024, Mistral AI conducted an aggressive lobbying campaign against EU AI Act provisions. The campaign was led by co-founder Cédric O, France's former Secretary of State for Digital Affairs, who joined Mistral in spring 2023 and immediately began lobbying his former government colleagues. O's initial €176 investment grew to approximately €23 million while he lobbied for exemptions that would directly benefit the company - a conflict he did not publicly disclose. Meanwhile, Mistral argued that strict regulation would force European companies to partner with US tech giants, while secretly negotiating a deal with Microsoft that was announced in February 2024. The campaign succeeded: the final AI Act gave broad exemptions to open-source models and general-purpose AI, with only minimal transparency obligations. Fundamental rights checks were removed, and foundation model requirements were significantly weakened.