Mistral's free tier trains on user conversations unless they opt out

Mistral raised €3 billion this month at a valuation above €21 billion, led by Samsung, on a proposition built around data control. President Macron described the round as a step towards "a third way in AI." The company's commercial argument to European buyers is that they can determine where their data goes and who can compel access to it.

A default that routes free-tier conversations into training is not inconsistent with that — the paying customers whose sovereignty is the pitch are excluded — but it does mean the sovereignty is a feature of the contract rather than of the company. A European user of the free product is in the same position as a free user anywhere.

That is defensible and it is worth saying plainly, because the marketing does not.

The opt-out is in the wrong place

The more substantive criticism is mechanical. The setting sits in the admin panel.

An admin panel is where an administrator goes. It is not where a person using a free chat product goes, and most free users will never open it. A default that requires navigating to an administrative settings page to change is, in practice, not a choice most users make — which is presumably reflected in the volume of data collected.

European regulators have views about this. Consent under the GDPR must be freely given, specific, informed and unambiguous, and burying the control in an administrative surface is the kind of design that has drawn enforcement against other companies. Whether Mistral relies on consent or on legitimate interests as its lawful basis determines how much of that applies, and the help centre update does not say.

The context it arrives in

Training data provenance is under more legal pressure than at any point so far.

The US Justice Department filed a statement supporting OpenAI and Microsoft's fair use position on 2 September. The Seattle Times and Newsday sued both companies three days later. Authors are contesting how Anthropic's copyright settlement gets divided. 404 Media documented a warehouse where books are cut apart and scanned for AI training.

Against that, a provider being explicit about what it does with free-tier conversations is better behaviour than most of the sector has managed — the update makes the position clearer than it was.

The question Mistral has not answered is what happens to conversations already collected under a less explicit policy, whether users can have that data removed from future training, and how a model that has already trained on it would be affected. Deletion of training data is straightforward. Removing its influence from trained weights is not, and no provider has a good answer.