Updated on December 16, 2020, to share some of our learnings from prompting Quote Tweets and sharing an update that we are re-enabling the standard Retweet behavior.
It’s been just over a week since Election Day in the US and the culmination of an unprecedented voting period that resulted in the highest voter turnout rate in 50 years. While we’re still seeing record levels of election-related conversation on Twitter and our teams continue to actively enforce our rules to protect the integrity of this public conversation, we want to provide an update about this ongoing work.
We also want to be very clear that we do not see our job as done — our work here continues and our teams are learning and improving how we address these challenges. We’ll be sharing a comprehensive report on the election early next year.
In the months leading up to Election Day, we announced a set of policy, enforcement and product changes to add context, encourage thoughtful consideration, and reduce the potential for misleading information to spread on Twitter.
Here’s what we’ve learned about the impact of those changes so far, and how we plan to adapt based on those findings:
Labels, warnings, and prebunks
More than a year ago, the public told us they wanted Twitter to offer context on misleading information. This is the exact approach we’ve implemented on Tweets about COVID-19, synthetic and manipulated media, and the 2020 US election that could contribute to offline harm. We continue to apply labels to add context and limit the risk of harmful election misinformation spreading without important context.
Below are some key statistics about the labels, warnings, and additional restrictions we applied to Tweets that included potentially misleading information about the US Election from October 27 to November 11:
We also got ahead of potentially misleading information by showing everyone on Twitter in the US a series of pre-bunk prompts. These prompts, which were seen 389 million times, appeared in people’s home timelines and in Search, and reminded people that election results were likely to be delayed, and that voting by mail is safe and legitimate.
These enforcement actions remain part of our continued strategy to add context and limit the spread of misleading information about election processes around the world on Twitter.
In the weeks leading up to and during election week, we implemented significant product changes intended to increase context and encourage more thoughtful consideration before Tweets are amplified. Starting today, we are reverting some of these changes and providing more detail on the impact they had.
Encouraging Quote Tweets
We encouraged people to add their own commentary when amplifying content by prompting Quote Tweets instead of Retweets. This change introduced some friction, and gave people an extra moment to consider why and what they were adding to the conversation. Since making this change, we observed a 23% decrease in Retweets and a 26% increase in Quote Tweets, but on a net basis the overall number of Retweets and Quote Tweets combined decreased by 20%. In short, this change slowed the spread of misleading information by virtue of an overall reduction in the amount of sharing on the service. We are taking more time to study and fully understand the impact of this change and are leaving it in-place for now.
Updated on December 16, 2020: We’ll no longer be prompting Quote Tweets, and are re-enabling the standard Retweet behavior. We hoped this change would encourage thoughtful amplification and also increase the likelihood that people would add their own thoughts, reactions and perspectives to the conversation. However, we observed that prompting Quote Tweets didn’t appear to increase context: 45% of additional Quote Tweets included just a single word and 70% contained less than 25 characters.
Removing Tweet recommendations in Home Timeline & Notifications
We stopped providing “liked by” and “followed by” Tweet recommendations from accounts you don’t follow in the Home Timeline and through notifications. While we had initially hoped that this would help reduce the potential for misleading information to spread on our service, we did not observe a statistically significant difference in misinformation prevalence as a result of this change (nor any meaningful reduction in abuse reports). Instead, we found that pausing these recommendations prevented many people from discovering new conversations and accounts to follow. As of today, we are reverting this change.
Our goal is to eventually replace these “liked by” and “followed by” recommendations with ones that are based on the Topics you follow or Topics we think you might like. We believe this will provide you greater control to tell us what you are and aren’t interested in, which will make our recommendations more relevant to you. Also, we don’t believe the “Like” button provides sufficient, thoughtful consideration prior to amplifying Tweets. We’ve grown our coverage of Topics recommendations significantly in the last year, with now more than 5,000 Topics available, and hope to be able to complete this transition soon.
Required context before showing Trends in “For You”
We only showed Trends in “For You” that had added context, meaning our team added a description, representative Tweet, or article to help people more quickly gain an informed understanding of high volume public conversation.
While we saw a significant reduction in reports (on Trends, and Tweets within Trends result pages) as a result of this change, we also recognized that it placed a significant limitation on the number and breadth of Trends that we could show people, making “For You” less relevant for many people’s interests. Moving forward, we’ll continue to prioritize reviewing and adding context to as many Trends as possible, but won’t make this a requirement before a Trend can appear in “For You.” Our goal is to help people see what’s happening, while ensuring that potentially misleading trends are presented with context.
We remain vigilant and will continue working to protect the integrity of the election conversation on Twitter. As we’ve done for many elections around the world, we will produce a longer-form retrospective of all of our work around the 2020 US Election in early 2021. We will continue to research, question, and change features that may incentivize or encourage behaviors on Twitter that negatively affect the health of the public conversation. For more, follow @TwitterGov and @Policy.
Did someone say … cookies?