Mistral CEO warns: closed AI models give excessive power to providers

What happened?

Arthur Mensch, co-founder and CEO of Mistral AI, published a warning on LinkedIn aimed at business leaders: closed artificial intelligence models (such as those from OpenAI, Google, or Anthropic) give their providers "immense power" over the businesses that use them. According to Mensch, when companies connect these models to their internal data (for example, for search or automation), providers can observe that traffic, learn from it, and in some cases retain the information to improve their own systems. This creates a dangerous power asymmetry, where the customer is completely dependent on the provider and loses sovereignty over their data. In his post, Mensch stated that "closed providers are now forcing data retention" and that "they gain immense leverage over their customers' businesses." The statement comes in a context where generative AI adoption has skyrocketed: according to Gartner, 80% of companies will have implemented some form of AI by 2026, and many do so without evaluating the risks of dependency.

Why is it important?

Mensch's warning is no coincidence. Mistral is one of the main drivers of open-source AI in Europe, directly competing with US giants like OpenAI (ChatGPT), Google (Gemini), and Anthropic (Claude). Its flagship model, Mistral Large, is distributed under open licenses, allowing companies to download, audit, and run it on their own servers. The warning comes at a time when many companies are adopting generative AI without evaluating the risks of technological dependency. If closed model providers can access proprietary data, they could use it to train models that compete with their own customers, or simply retain it to improve their business. This is especially critical in regulated sectors such as finance, healthcare, or defense, where data leaks could have legal and reputational consequences. Moreover, the debate over data sovereignty intensifies with the recent approval of the EU AI Act, which requires transparency and control over data used in high-risk systems.

Consequences for companies and the market

Mensch's stance reinforces the debate over data sovereignty and open vs. closed models. Companies that opt for closed models will need to negotiate contracts that clearly specify what data is shared, how it is used, and when it is deleted. However, many current contracts from providers like OpenAI and Google Cloud Platform (Vertex AI) allow the use of customer data to improve models, unless explicit opt-out is requested. On the other hand, open models allow auditing, customization, and on-premise deployment, reducing the risk of leaks or misuse. But they require more investment in infrastructure and talent, which can be a barrier for small and medium-sized enterprises. The market could segment: companies with sensitive data will migrate to open solutions, while others will prioritize the ease of use of closed models. According to a McKinsey report, generative AI could add between $2.6 and $4.4 trillion annually to the global economy, but trust in providers will be key to unlocking that value. Recent examples, such as the Zoom controversy over user data for AI training, show how lack of transparency can erode customer trust.

What should readers know?

If your company uses generative AI, review the provider's terms of service. Can they retain your data? Do they use your queries to train models? For example, OpenAI allows enterprise users to opt out of training, but does not offer contractual guarantees on data deletion. Consider open-source alternatives like Mistral, Meta's Llama 3, or TII's Falcon, which allow code auditing and local deployment. Also evaluate platforms that offer private deployment, such as Azure AI or AWS Bedrock with open-source models. Transparency and control over data will be key competitive factors in the coming years. As a practical recommendation, companies should conduct a data risk audit before integrating any AI model, and establish clear data governance policies. Mensch's warning is a reminder that digital sovereignty is not just a technical issue but a strategic one for long-term competitiveness.