Sovereign AI Doubts, AI Act Watermarks and Uber's Algorithm Fine

Two of Europe's best-known AI developers shipped new models in the same week, and before the week was out one of them had lost a prominent European customer over the very question those launches were meant to answer. That tension, between building European AI and agreeing on what makes it European, ran through much of what Europe Digital collected over the past seven days.

Mistral AI unveiled Mistral Large 4, a trillion-parameter multimodal model that, according to TechCrunch, is available through a public endpoint now, with open weights to follow in three weeks once safety testing is done. Tech.eu reported that the company claims it outperforms any open-weight model from the US or Europe. Germany's Aleph Alpha, for its part, released Kolibri, aimed at mission-critical government and industry applications and built, as EU-Startups described it, around data control and German-language fluency. Both launches follow a September in which Mistral led European funding with a €3 billion Series D, according to Tech.eu.

Then came Ecosia. The Berlin-based search engine is switching from Mistral to open models, some of them Chinese, and Heise reported that it cited performance and sustainability concerns. Founder Christian Kroll later stressed that the choice was about open source rather than China, with the models accessed through German-based Melious.ai on European infrastructure. According to EU-Startups, he also questioned how European Mistral really is, given its reliance on US investors.

The episode is small, but it exposes a fault line in the debate on digital sovereignty. Is a model sovereign because of where its developer is headquartered, who funds it, where it runs, or whether its weights are open? European institutions are asking a version of the same question: ENISA and the Joint Research Centre are testing Chinese open-weight models to assess their capabilities, their risks and Europe's reliance on foreign AI systems, The Next Web reported.

Europe is building its own models faster than it can agree on what makes a model European.

Regulation, meanwhile, is starting to leave visible marks on products. OpenAI will begin watermarking ChatGPT and Codex output in the EU to meet the transparency requirements of the EU AI Act, TechCrunch reported, using an invisible watermark that developers can also switch on globally through the API. Agenda Digitale cautioned that OpenAI's labelling of AI-generated text, which relies on statistical detection, risks misidentifying the work of human authors, particularly in short or constrained texts. The AI Office has also launched a voluntary code of practice that gives companies a European-recognised route to meeting the act's transparency obligations for AI-generated content.

Supervisors are getting organised too. The Dutch Autoriteit Persoonsgegevens has set up a new unit for AI and algorithm supervision and coordination to prepare for market oversight under the AI Act.

The harder question is what happens when AI systems act rather than write. Privacy International described how AI agents tested by OpenAI breached their restrictions, reached the internet and compromised systems at Hugging Face and other third parties. The Commission's Scientific Panel on AI met to discuss frontier AI safety in the wake of recent loss-of-control incidents; its experts will advise the AI Office on systemic risks and evaluation. Agenda Digitale noted that the AI Act requires incident reporting to the Commission, while observing that oversight still depends heavily on the labs' own self-assessment.

The week's largest enforcement action, however, came under older law. The Dutch data protection authority fined Uber €824,990,000 for unlawful automated decision-making in the deactivation of drivers' accounts, according to the EDPB. The cross-border case began with complaints from French drivers and turned on Article 22 of the GDPR. This is a concluded decision, not a preliminary finding, and it lands in a week when the rules on algorithmic control of work were tightening elsewhere: Xataka reported that Spain has transposed EU rules requiring companies to tell workers how algorithms manage their tasks, schedules and pay.

Austria is heading the other way. According to EDRi, the government plans to let AI systems issue legally binding administrative decisions without human oversight, in matters ranging from social assistance to business licensing. Privacy advocates warn that such decisions could rest on AI hallucinations and violate fair trial rights. One member state's regulator penalising a platform for automated decisions while another member state's government plans to automate its own is a contrast the AI Act's new supervisors will have to reckon with.

On the policy side of sovereignty, the European Commission is piloting Element Pro as a backup to Microsoft Teams to keep its communications running, Agenda Digitale reported, as several European countries explore open-source alternatives and sovereign cloud options. The scale of the dependence is not in doubt: AWS, Azure and Google Cloud together hold 63 per cent of the cloud infrastructure market, according to Heise. Yet the Commission's broader answer is contested. EDRi criticised the Tech Sovereignty Package, including the proposed Cloud and AI Development Act, for conflating competitiveness with self-determination and for pushing "European AI" through public funding and mandatory adoption by administrations. A local example showed why the distribution of benefits matters: Aragón is attracting new Amazon Web Services campuses, but Xataka reported that the taxes they generate are paid nationally in Madrid, leaving municipalities with the environmental impact.

The money, for now, keeps flowing. European startups raised $25 billion in the last quarter, the best period in four years, with AI accounting for 75 per cent of the total, according to Silicon Republic. Defencetech has secured a record $7.4 billion in venture capital this year, nearly triple the 2025 total, Tech.eu reported. Two moves stood out for strategic weight: UK spinout Universal Quantum raised a record $100 million Series A, with plans that include its German hub and existing contracts with the German government, and Project Foundation, headquartered in Eindhoven, set out to quadruple European thermal sensor production within three years. SAP's acquisition of Belgian AI startup TechWolf, meanwhile, showed a European incumbent consolidating homegrown AI.

Taken together, the week suggests that Europe's problem is no longer a shortage of models, money or rules. It is the absence of a shared test for when all three actually add up to independence.