In the recent proceedings involving DeepSeek, Mistral AI and NOVA AI, the Italian Competition Authority (AGCM) makes one point clear: what matters is not only that AI systems can make mistakes, but also how that risk is communicated to users.
With decision No. 31864, the Italian Competition Authority (AGCM) addresses one of the most delicate aspects of generative AI services: managing the risk of inaccurate outputs (“hallucinations”) from a consumer protection standpoint.
Proceeding PS12968 concerns the “Le Chat” service by Mistral AI. The Authority alleged a potential infringement of Articles 20, 21 and 22 of the Consumer Code, taking the view that users were not informed in a sufficiently clear, immediate and intelligible manner about the possibility that the system could generate incorrect or misleading responses. The issue was not so much the absence of information per se — which was available in the terms of service and help centre — but rather its actual accessibility and visibility within the user experience.
The decision fits within an increasingly consolidated approach: in digital practices addressed to consumers, relevant information must be effective, not merely formal. In other words, it is not enough for a risk to be described in contractual documents if those documents are not readily accessible or do not reach the user at the moment decisions are made or the service is used.
In this context, Mistral AI offered commitments under Article 27(7) of the Consumer Code, thereby avoiding a formal finding of infringement. The commitments — deemed suitable to address the concerns — develop along four main lines: (i) inclusion of disclaimers directly within chat interfaces (“Le Chat may make mistakes. Please check responses”); (ii) strengthening and Italian localisation of the terms of service, with explicit reference to the potential unreliability of outputs and the need for verification; (iii) improved accessibility of the terms throughout the user journey (homepage, login, registration, app store, interface); and (iv) full translation of the website and help centre into Italian.
The distinctive feature of the decision is the affirmation of “contextual” transparency: users must be warned of risks not only “somewhere”, but at the time and place where those risks materialise — i.e. during interaction with the system. This marks a shift towards a compliance-by-design logic, already familiar in other regulatory domains.
The decision also has a broader systemic implication: “hallucinations” are no longer merely a technological limitation, but become a legally relevant factor in assessing the fairness of a commercial practice. A generative AI system that — even implicitly — leads users to consider its outputs as reliable may be regarded as misleading, unless accompanied by clear and appropriate warnings.
Practical takeaways
For operators developing or integrating generative AI solutions:
