Meta: Multimodal AI model not coming to the EU
Because the regulation is not clear enough, Meta is not bringing its latest AI model to Europe. At least that's Meta's reasoning.
(Image: Koshiro K/Shutterstock.com)
"We will be launching a multimodal Llama model in the coming months, but not in the EU as the European regulatory environment is unpredictable," a Meta spokesperson told heise online. Reports had previously suggested that Meta was taking this step. Meta AI, which includes most of Meta's AI applications, is also not yet available in the EU. It is not entirely clear whether this is due to uncertainties arising from the AI Act, the Digital Markets Act (DMA) or the GDPR.
In the EU, the AI Act was only recently published in the Official Journal. This means that it now applies with different deadlines. From February 2025, for example, there will be a ban on certain particularly risky AI applications, such as real-time biometric monitoring, remote identification systems and social scoring. Six months later, the regulations for GPAIM - general purpose AI models, including ChatGPT and the like - and high-risk applications will come into force. Some critics say that the AI Act is too vague. However, it by no means prohibits chatbots or multimodal models.
Apple is also citing regulatory uncertainties in the EU - and is therefore unlikely to launch Apple Intelligence on the market for the time being. However, they cite the DMA as the reason. Here, too, it remains unclear why exactly the AI services should violate the regulation.
The background to Meta's decision could also be the GDPR. This concerns the training data for the AI models. Meta recently changed its terms of use worldwide. This means that all public posts by people who use Facebook, Threads and Instagram can be freely used for Meta's AI training. This also means, for example, that publicly posted photos are included in the AI models. In the EU, it was initially possible to object. This form of opt-out called the data protection experts into action, who said it did not comply with local data protection laws. Meta subsequently withdrew the changes - initially, no user data from the EU will flow into the training.
Meta's multimodal model is to be integrated into a variety of products - including smart glasses from Ray Ban. These would then be at most half-smart in the EU. So far, Meta has published its Llama models as open source. If there is a restriction for the EU, the model can still not be used.
It remains unclear whether the multimodal model is the one that has already been speculated about. Meta is expected to release a version of Llama 3 this week - with 400 billion parameters.
(emw)