Update June 14, 2024 at 7:30AM PST:
“We are disappointed that our lead regulator, the Irish Data Protection Commission (DPC), acting on behalf of the European DPAs, has asked us to delay training our large language models (LLMs) with public content shared by adults on Facebook and Instagram, especially since we have taken regulators’ feedback into account and notified the European DPAs since March. This is a setback for European innovation, the race to develop AI and will further delay bringing the benefits of AI to Europeans.
We are highly confident that our approach complies with European laws and regulations. AI training is not specific to our services, making us more transparent than many others in the industry.
We are committed to bringing Meta AI and the model that powers it to more people around the world, including in Europe. But, simply put, without including local information, we’re only offering people a second-rate experience, which means we can’t release Meta AI in Europe at this time.
We will continue to work with the DPC to ensure that Europeans have access to the same level of AI innovation and receive the right services as the rest of the world.
The delay will also enable us to respond to certain requests we have received from the UK regulator, the Information Commissioner’s Office (ICO), before we can begin training.
Originally published on June 10, 2024 at 6am PST:
We’ve been hard at work for years building next-generation AI capabilities across our family of apps and devices, and this year we’ll be expanding our collection of generative AI capabilities and experiences, and the model that powers them, AI at Meta, to people in Europe.
Meta’s AI is already available in other parts of the world, including: Llama – Cutting-edge open source large-scale language models – and Meta AI Assistant – The most intelligent AI assistant that is free to use. To properly serve our European community, the models that power Meta’s AI need to be trained with relevant information that reflects the diverse linguistic, geographical, and cultural references of the European people who use it. To achieve this, we want to train the large-scale language models that power our AI capabilities with the content that EU people choose to publish in Meta’s products and services.
“Without training our models on public content, such as public posts and comments that Europeans share on our and other services, our models and the AI features they power will not be able to accurately understand important regional languages, cultures, or trending topics on social media. We believe that Europeans will be disadvantaged by AI models that do not take into account Europe’s rich cultural, social, and historical contributions.”
Meta is not the first company to do this. Open AIBoth companies already train their AI with data from European users, and our approach offers more transparency and control than many of our industry peers who already train their models using similar public information.
We are committed to responsible and transparent AI development
Building this technology comes with a responsibility to develop best practices and policies that comply with local laws and regulations. In line with this commitment, we’ve consulted with the Irish Data Protection Commission, the EU’s lead privacy regulator, and incorporated their feedback to date to ensure that the AI training methods in Meta comply with EU privacy law. We also continue to work with experts, including academics and consumer advocacy groups, to ensure what we build follows best practices.
We want to be transparent so people are aware of their rights and the controls available to them, which is why since May 22, we have sent over 2 billion in-app notifications and emails to people in Europe explaining our efforts. These notifications include links to objection forms that people can use to object to the use of their data in AI modeling efforts.
In crafting our notice, we reviewed the approaches of our industry peers as well as previous policy update notices. As a result, our form is easier to find, read, and use than those offered by other companies providing generative AI in the EU. It’s accessible in just three clicks and has fewer fields to complete. To further aid comprehension, although we do not train our Llama model on content from the accounts of Europeans under the age of 18, we designed it to be accessible to those with a younger reading age.
We respect all European objections, and if an objection form is submitted before Llama training begins, that person’s data will not be used to train those models, either in the current training round or in future ones.
To be clear, our goal is to build useful features based on the information that European users aged 18 or over choose to make public on Meta products and services (such as public posts, public comments, and public photos and their captions). Our models may be trained based on the posts that users publicly share; however, this is not a database of each user’s information, nor is it designed to identify individuals. Rather, these models are built by looking at users’ information and identifying patterns, such as understanding colloquial phrases and local references, rather than to identify specific users or their information.
As We said, We do not use private messages with friends and family to train our AI systems, although in the future we do plan to use other content, such as your interactions with AI features and chats with companies using AI in Meta AI.
Like many industry peers who have preceded us in using European data to train language models at scale, we rely on the “legitimate interest” legal basis to do this work in the EU and to comply with the EU’s General Data Protection Regulation (GDPR), which we believe strikes the most appropriate balance for processing public data at the scale necessary to train AI models while respecting people’s rights.
We feel a responsibility to build AI that is actually made for Europeans, not forced upon them. To do that while respecting the choices of European users, we believe the right thing to do is to inform them of our plans, and give them the choice to tell us if they don’t want to participate. And we believe the best way to strike this balance is for companies to be transparent about the information their AI uses, while still providing users with prominent controls to opt out of those uses if they wish. And that’s exactly what we’ve done.
Europe is at a crossroads
As Europe stands on the cusp of society’s next big technological evolution, some activists are arguing for an extreme approach to data and AI. Let us be clear: this position does not reflect European law and amounts to an argument that Europeans should not have access to, or be properly served by, AI used in the rest of the world. We strongly disagree with this conclusion.
As one of the most influential regions in the world, Europe has the potential to lead the competitive AI innovation race. But questions remain: Will Europeans have equal access to groundbreaking AI? Will our AI experiences reflect our culture, humor, and history? Or will Europe want to stand by and watch while the rest of the world benefits from truly innovative technologies that build community and foster growth?
AI is the next frontier of innovation. We are living in the most exciting technological era of a generation, with breakthroughs happening right before our eyes and endless possibilities. At Meta, we want Europeans to be part of it.
