We use AI to help the billions of people who use our services every day discover content that they will hopefully find useful and interesting–whether it’s a new creator to follow on Instagram, or a post they may enjoy on Facebook.

Our goal in having these systems is to make sure people see content that they will hopefully find relevant and valuable. There isn’t a singular AI system that determines everything people see on Facebook and Instagram. Instead, many AI systems work separately–and in some cases, together–to seamlessly power these experiences behind the scenes – in just a fraction of a second. Diving even deeper, each AI system has multiple models that identify content and predict how likely a person is to be interested in it or interact with it.

It is important that the people who use Facebook and Instagram have access to information about the technology that powers their experiences. This information must also be accessible and explained in a way that non-experts and experts alike can understand.

• An overview of the AI system
• A section explaining how the AI system works, which includes a summary of the steps involved in creating experiences on Facebook and Instagram
• A section describing how to customize the content that is shown. This includes descriptions of system controls and instructions for how each person can control and customize their experience.
• A section explaining how the AI delivers content, which includes an explanation of how some of the significant prediction models inform the overall AI system and produce product experiences

The prediction models of an AI system may use information such as the features of posts and a person’s history of interacting with similar posts when making predictions about levels of interest. There are thousands of these signals used across Facebook and Instagram.

For example, when predicting how likely it is that a person will engage with a post, AI systems consider signals that include:

• How many likes, comments, and views a post or video has received
• How often or how long a person watches videos or views posts
• How a person interacts with a given author, such as how many times they have seen and liked similar posts from that person

Importantly, our system cards also describe the controls that exist for each AI system that people can use to customize their experiences. For example, if someone wanted to see less of a certain type of post, they could unfollow the author, temporarily hide content, or click “Show less” on Facebook and “Not interested” on Instagram to temporarily decrease the ranking score for similar types of content.

One of our biggest challenges in creating these system cards was figuring out the best way to explain highly technical information in a way that everyone can understand. There isn’t an industry standard way to do this, so we created a uniform approach at Meta to explain these systems. By listening to people who use our services and speaking with a diverse group of experts throughout the design and development process, we gained insights that helped us determine how to present this information in a meaningful way that will reach our intended audiences. We heard that people want more transparency and control over what they see, so we added a customization section to each system card. We also learned that giving people too much technical detail can sometimes obfuscate transparency, which is why we present the top ten most important prediction models, rather than everything in the system.

In order to keep our approach consistent, we chose a glossary of terms to use when we talk about AI. When explaining terms that might be unfamiliar, we include a tooltips prompt to remind people what we mean when we say things like “connected content” and “AI system.” By taking a consistent approach to language, we enable people to compare and contrast multiple system cards.

To maintain consistency across sections in the system cards that explain how our AI systems work and deliver content, we developed internal tools to analyze the impact of the models that make up an AI system. With the help of our engineers, we translated this information–from signals to words–to help explain how each AI system makes predictions. Notably, these models and signals are dynamic as the system learns, and they change frequently over time.

As the industry evolves and discussions about system documentation and transparency continue, we will further identify opportunities to undertake and iterate on our approach over time, so we can reflect product changes, evolving industry standards, and expectations around AI transparency.

We will continue to incorporate feedback from diverse audiences to improve our products and empower people who use them. In our research we have learned that people want to explore system cards when the information provided is relevant to them, with a mix of text and visuals, so we are working on continuously improving system cards.

People can find our system cards by visiting our Transparency Center. We hope this effort encourages people to learn more about how AI powers their experiences. We believe that system cards will empower people to educate themselves about AI and control and customize how they experience our products.