Meta begins Ai training on public posts and comments. Here's how to object

Meta will use public content shared by adult EU users on its platforms, such as public posts and comments on Facebook and Instagram. It will also exploit interactions with Meta AI - such as questions and requests made to the assistant.

But 'we do not use private messages exchanged with friends and family to train our generative AI models. Furthermore, public data from accounts belonging to people in the EU under the age of 18 will not be used for training purposes'.

Objective, to improve the quality and accuracy of the models.

"Last month we launched Meta AI in the EU, the first step in making Meta AI's chat function available for free across Europe, via the messaging apps people already know and love: WhatsApp, Instagram and Faceboo's Messengerk" .

"We believe we have a responsibility to build artificial intelligence that is not only accessible to Europeans, but truly designed for them. That is why it is crucial that our generative AI models are trained on a variety of data that allows them to understand the incredible and varied nuances and complexities that characterise European communities'. This includes everything from dialects and idioms to local knowledge and the distinctive ways in which different countries use humour and sarcasm in our products. This is particularly important at a time when artificial intelligence models are becoming more and more advanced, with multi-modal functionalities including text, voice, video and images'.

Meta points out that this training is already being done by its competitors, 'including Google and OpenAI, which have already used European user data to train their own AI models. We are proud of the fact that our approach is more transparent than that of many others in the industry'. And that the delay in the launch was precisely dictated by the desire to fully comply with the European regulatory framework. "We are sorry that it has taken almost a year to get to this point, we welcome the clarity provided by both the Irish Data Protection Commission (IDPC) and the European Data Protection Board (EDPB), which has allowed us to take this next step," the company writes.

Meta thus tries to balance two requirements. Pushing on AI and meeting the EU's stringent requirements, which are often also supported by citizens (as the Meta AI controversy shows). Meta aims to strengthen its presence in the generative artificial intelligence market, where it lags behind already established competitors. Meta's Llama model is unique due to its open source nature (limited to the weights used in neural networks), but it is judged by independent benchmarks (such as those of Artificial Analysis) to be less performing than OpenAi Gpt, Anthropic Claude, Google Gemini, on tasks of complex reasoning, multimodality, contextual comprehension in long texts.

A few days ago, Meta criticised Chatbot Arena, the open-source platform that compares language models through anonymous user votes, because its new LLaMA 3-70B model was ranked below GPT-4, Claude 3 Opus and Gemini 1.5, and only at the level of GPT-3.5. According to Meta, technical benchmarks show superior performance, and Arena's ranking does not reflect the real value of the model. The academic community defended Arena, arguing that it measures real human preferences, which are more representative of everyday user experience.

It should also be considered that Meta's privacy game is not to be considered definitively won.

It was only in December 2024 that a form of green light came for this AI training: the Edpb confirmed the conformity of Meta's initial plan with legal obligations, thus allowing it to proceed.

The method chosen by Meta - i.e. a mechanism of opposition (opt-out) instead of explicit consent (opt-in) to the use of data for training - remains controversial, however. Organisations such as Noyb, founded by activist Max Schrems, criticise this approach, arguing that it does not comply with the spirit of the privacy regulation GDPR, which favours explicit, free and informed consent on the part of the user. According to these critics, an email or notification is not sufficient to guarantee truly informed consent.

This tension, between big tech that has to push on AI and Europe, will only escalate if the two blocs US and EU do not agree on trade tariffs.

In short, a very open game and big tech in Europe is now walking a fine line between innovative drives, regulations and rights.