Meta recently released 22 different information cards that explain how AI affects the content users see on Facebook and Instagram.
Written by: Wang Mei
Meta has recently released 22 different information cards that provide specific explanations to users about how the company uses AI to control what they see on Instagram and Facebook.
Meta says this move is intended to provide more transparency to its AI system, but the specific explanations in these cards still remind users that they are all living in algorithms that precisely target them.
- Ubisoft Releases Blockchain Game “Champions Tactics” – Is the Big Player Entering the Game?
- Is GameFi about to make a comeback? Analyzing 3 different types of blockchain games
- Sovereign SDK: Significantly reduce bridging delays for optimistic rollups
Some of the 22 information cards from Meta are shown below:
Facebook information card
Instagram information card, source: Meta Transparency Center
Each card provides detailed and easy-to-understand information on how the AI system behind these features ranks and recommends content.
Here are a few examples:
How Facebook Reels displays content to users
When users view content and engage on Facebook, an underlying AI system will provide Reels (short videos), including content posted by creators that users may be interested in but have not yet followed, and cross-app recommended content from Instagram.
The AI system that Facebook Reels relies on automatically determines the Reels to be shown to users and their display order by predicting the content most likely to interest or interact with users. These predictions are based on a variety of factors, including users’ recent attention, likes, or interactions with users and content. The operation is as follows:
1. Collect inventory information
First, the AI system collects all the Reels that users may be interested in, including Reels posted by users they follow or accounts they follow, or Reels similar to Reels that users have recently interacted with. The AI system may also recommend Reels published by sources similar to users that they follow or interact with.
2. Use directional signals
Next, the AI system will take into account all the directional signals related to each Reel, which may include the duration of the Reel, its similarity to other Reels, and its match with the content that users are willing to interact with, and run a simple model to select about 10-100 of the most relevant Reels (problem content will be filtered out).
3. Make predictions
At this stage, the AI system uses the model to help predict what content users will find most relevant and valuable.
4. Sort Reels by score
Finally, the system calculates relevance scores for approximately 200 posts and sorts them by score. Reels that the system predicts will provide higher value to users are displayed higher in the feed.
Meanwhile, users can signal to the system to show fewer or more similar Reels by hiding or saving and sharing Reels.
How Facebook Stories shows content to users
According to the explainer card, here’s how Facebook Stories (a feature that allows users to post photos and videos that disappear after 24 hours) works:
1. Collect Stories
First, the system collects all relevant stories shared by users or public Pages in the past 24 hours (filtering out any that violate rules).
2. Predict and analyze
Next, the system collects all the Stories it can show to a given user and predicts which ones they’ll find most relevant and valuable. It keeps these Stories and removes the rest.
Factors the system considers include how many times the user clicked to view the Story full-screen, the total time spent viewing stories from the author, the number of different authors and collections of Stories the user has viewed, how many times the user responded to an author’s Story by liking or chatting and how long users typically spend viewing Stories.
3. Sort Stories
The system then sorts this small set of Stories by the likelihood that the user will interact with each.
4. Apply additional rules
Finally, the system applies additional rules to ensure a balanced mix of Stories from users and public Pages is shown.
Users can share Stories by sending them to others via Messenger or adding them to their own Stories or a public Page’s Stories. They can also opt not to view Stories. If they do, they won’t see any other Stories posted by the user or public Page that created the Story, unless they choose to view them again.
How Instagram Explore shows content to users
Instagram’s Explore feature shows users recommended content, such as photos and Reels from accounts they don’t follow that might be related to their interests or similar to content they’ve interacted with before.
Its operation is similar to Facebook Reels:
First, collect inventory information. The AI system will collect some public content displayed on Instagram (excluding problematic content), such as photos and Reels;
Then, use instruction signals: the AI system will comprehensively consider the user’s participation in similar content or interested content. This includes the length of time the post has been published, the total number of times the user has viewed or clicked the post thumbnail, the number of times the post is prioritized in a series of posts, the number of times the user clicks “not interested” in the post, and the number of times the user browses the posts published by authors who share similar interests with the author of this post;
Finally, rank the content. The system predicts that it will provide users with higher value content and push it to the front of the “Discovery” tab.
Similarly, users can affect this process by saving or marking the content as “not interested” to encourage the system to continue displaying or filtering similar content in the future.
Users can also view Reels and photos that the algorithm did not specifically select for them by selecting “non-personalized” in the Explore filter.
How Instagram Search displays content to users
When users view content and interact on Instagram, a underlying AI system provides results when users search for content.
The AI system that Instagram Search relies on automatically sorts search results by predicting the most relevant and valuable content for users.
Its operation is as follows:
1. Collect inventory information.
First, the system collects all search results that meet the criteria to sort them for users. Such content may include topic tags, locations, Reels, posts, homepages, audios, or other results related to the words searched by users.
2. Score the results.
Then, the system scores each search result based on various factors, such as the type of content and the degree of matching with the content with which users usually interact. Factors considered include:
The similarity between the words used in the search and the words used in the account name or homepage name of the account, the similarity between the words used in the search and the words used in the suggested keywords, the similarity between the words used in the search and the words used in the topic tags, and the number of times users in the same country/region who perform the same search click on the topic tags.
3. Apply more filtering conditions.
The system will apply “additional filters” and “integrity processes” to narrow the scope of qualified content to the search results most relevant to users.
4. Sort the results by score.
Finally, the system will prioritize displaying the results that the system predicts to be the most valuable or relevant to the user based on the score.
Meanwhile, the system will customize the Instagram search experience dynamically for users, who can choose to control or customize the displayed content, or view non-personalized search results.
Can users resist?
Anyway, in the smart age where everyone lives in algorithms, Meta’s approach of increasing algorithm transparency is commendable.
Research shows that algorithm transparency can play a role in two dimensions: accountability and the right to know.
First, algorithm transparency can make algorithm operators more accountable, and once there is a bias in accuracy and fairness, the responsibility of algorithm operators can be asserted based on the disclosed algorithm.
Second, algorithm transparency also gives the regulated object of the algorithm a certain degree of the right to know, and this right to know is beneficial to third parties (especially professionals) to supervise, and also beneficial to the regulated object of the algorithm to question the fairness and rationality of algorithm decisions based on the disclosed algorithm after the fact.
Of course, Meta’s release of this information is also in line with the regulatory trend. Currently, European legislators are rapidly advancing legislation to provide new interpretations and transparency requirements for companies using AI technology, and US legislators have also expressed a desire to begin drafting similar legislation later this year.
In addition to the 22 cards already published, Meta also said that it will expand the scope of explanations to include “Why am I seeing this post” and other functions in the coming weeks.
In addition, Meta also offers a feature that allows users to centrally control the content they want to see on Facebook and Instagram.
Instagram already supports selecting “Not Interested” for some posts to increase the recommendation of content that is less similar to it. In the future, users will soon be able to select “Interested” to view certain types of content, and Meta will provide richer options in the future.
Over time, users can also resist AI to some extent.