Takeaways:
We are imitating other companies like Google and OpenAI, which have previously trained AI using data from Europeans. Compared to many of our industry peers who are currently training their models on identical publicly available data, our approach is more transparent and provides easier controls.
Models are not meant to identify a single person or their information; rather, they are constructed by analyzing people's data to find trends, such as comprehending slang terms or local allusions.
We don't use information from European accounts belonging to people under the age of eighteen, nor do we use people's private chats with friends and family to train our AI systems.
We are building our open-source, fundamental AI model with content that users have chosen to make public.
Update on June 16, 2024 at 10:35am PT:
We are still very sure that our strategy conforms with all applicable laws and regulations in Europe. We offer greater transparency than many of our competitors in the market, and AI training is not exclusive to our offerings.
Our goal is to make Meta AI and the models that drive it available to as many people as possible, especially those in Europe. To put it another way, though, we could only provide a mediocre experience to users if we didn't provide local information. This indicates that Meta AI cannot currently be launched in Europe.
We will keep collaborating with the DPC to make sure that Europeans can access and benefit from the same degree of AI innovation as people in other parts of the world.
In addition, by delaying the start of the training, we will have time to respond to certain demands that our UK regulator, the Information Commissioner's Office (ICO), has made.
Originally published on June 10, 2024 at 6am PT:
We've been putting a lot of effort into developing the upcoming AI features for our line of apps and gadgets for years. Additionally, we want to make AI at Meta—a collection of generative AI experiences and features, as well as the models that underpin them—available to people in Europe this year.
Meta's AI is already accessible in various regions of the globe. This includes the most intelligent AI assistant you can use for free, the Meta AI assistant, and Llama, our cutting-edge large language model that is available as an open source project. The models that drive AI at Meta must be trained on pertinent data that represents the various languages, geographic locations, and cultural contexts of the people in Europe who will utilize them in order for them to effectively serve our European communities. In order to achieve this, we wish to use the content that EU citizens have decided to post openly on Meta's products and services to train our massive language models that drive AI capabilities.
The public content that Europeans share on our services and elsewhere, such posts or comments, is what trains our models on. Without this, models and the AI features they drive won't be able to comprehend crucial regional languages, cultures, or popular social media subjects. AI models that are not influenced by Europe's rich cultural, social, and historical contributions, in our opinion, will be detrimental to Europeans.
We are not the first firm to do this; Google and OpenAI, for example, have already trained AI using user data from Europe, and we are only following their lead. Compared to many of our industry peers who are currently training their models on identical publicly available data, our approach is more transparent and provides easier controls.
Our goal is to develop AI in a transparent and responsible manner.
Developing best practices and policies that go by regional laws and regulations is a duty that comes with building this technology. As part of this commitment, we are consulting with the Irish Data Protection Commission, the primary EU privacy regulator, and have taken their suggestions into consideration thus far to make sure that Meta's AI training adheres to EU privacy regulations. In order to make sure that what we develop adheres to best practices, we also keep collaborating with specialists like academics and consumer advocates.In order for individuals to understand their rights and the controls at their disposal, we want to be open and honest with them. We have therefore explained what we're doing to over two billion individuals in Europe through emails and in-app notifications since May 22. People can object to the usage of their data in our AI modeling efforts by clicking on the link in these alerts to fill out an objection form.
We studied the strategies used by our industry peers and our earlier policy update notifications while developing our own. We therefore made our form simpler to locate, read, and utilize than those provided by other EU-based companies providing generative AI: it only needs three clicks to access and requires the completion of fewer fields. Even though we are not training our Llama models on content from accounts of Europeans under the age of 18, we nevertheless built it to be more readable for those who are less literate in order to further facilitate understanding.
All complaints from Europe are respected. Data from that individual will not be utilized to train those models, either in the current training round or in subsequent ones, if an objection form is provided before to the start of Llama training.
To be explicit, our objective is to develop valuable features based on data that Europeans over the age of 18 have voluntarily chosen to share publicly on Meta's products and services. Examples of this type of data include public posts, public comments, and public images with captions. While publicly published posts may be used to train models, the material is not stored in a database or intended for individual identification. Rather of identifying a single individual or their information, these models are constructed by analyzing people's information to find trends, such as comprehending slang terms or local allusions.
We have said that we do not train our AI systems using private messages that users exchange with friends and relatives. We want to use additional content in the future, such talks with businesses employing AI at Meta AI or interactions with AI features.
Similar to several industry peers who have preceded us in training extensive language models with European data, we will utilize a legal foundation of "Legitimate Interests" to ensure compliance with the General Data Protection Regulation (GDPR) in order to carry out this work within the EU. We think that this legal foundation strikes the best possible compromise between upholding people's rights and processing public data at the volume required to train AI models.
We believe it is our duty to create AI that is truly intended for Europeans rather than imposed upon them. We believe that informing European users of our goals and giving them the option to opt out would be the appropriate course of action to accomplish that while honoring their choices. We also think that giving people clear tools to opt out of these uses, should they so choose, and being open and honest about the data that AIs are utilizing are the best ways for businesses to strike this balance. That's exactly what we've done.
dsEurope is at a Crssroa
Europe, one of the most powerful continents on earth, has the capacity to lead the world in AI innovation competitively. But there are still unanswered concerns: will innovative AI be equally accessible to Europeans? Will our history, humor, and culture be reflected in AI experiences? Or is Europe content to stand by and watch as the rest of the world gains from genuinely ground-breaking technology that fosters togetherness and promotes development?
Artificial Intelligence is the next big thing. In a generation, we are living through one of the most exciting times in technology history, with incredible advancements occurring right before our eyes and countless opportunities. We at Meta want Europeans to be involved as well.