In 2020, we tested prompts that encouraged people to pause and reconsider a potentially harmful or offensive reply before they hit send. In that initial analysis, we wanted to know if prompting individuals to take another look at their words could foster healthier conversations. 

While it was clear that prompts cause people to reconsider their replies, we wanted to know more about what else happens after an individual sees a prompt. To understand this, we conducted a follow-up analysis to look at how prompts influence positive outcomes on Twitter over time. Today, we are publishing a peer-reviewed study of over 200,000 prompts conducted in late 2021. We found that prompts influence positive short and long-term effects on Twitter. We also found that people who are exposed to a prompt are less likely to compose future offensive replies. 

How we got here

Offensive content online is a familiar challenge for social platforms. A common approach used to address this problem involves identifying and removing harmful content reactively, after it’s been posted. However, this content ideally wouldn’t exist to begin with. As such, there’s been an industry-wide shift toward more proactive strategies, like these prompts that check-in with individuals posting on Twitter before they even send the content. 

What we learned

Prompts change actions in the moment

We found that out of every 100 Tweets where users were prompted to reconsider, the following actions were taken:

• 69 were sent without revision
• 9 were canceled
• 22 were revised

The quality of the 22 revisions can be broken down further:

• 8 Tweets were revised to be less offensive than the original Tweet
• 13 Tweets were rewritten but were neither more nor less offensive than the original Tweet
• 1 Tweet was revised to be more offensive than the original Tweet

Prompts change future action

We also found the effects of being presented with a prompt extended beyond just the moment of posting. We saw that:

• After just one exposure to a prompt, users were 4% less likely to compose a second offensive reply.
• Prompted users were 20% less likely to compose five or more prompt-eligible Tweets.
• Prompted users received fewer offensive replies themselves. The proportion of replies to prompt-eligible tweets that were offensive decreased by 6% for prompted users.

This represents a broader and sustained change in user behavior and implies that receiving prompts may help users be more cognizant of avoiding potentially offensive content as they post future Tweets.

What’s next

Overall, we’re optimistic to see how prompts can improve the health of conversations. When offered a moment to reconsider their potentially harmful comments, many people will take it. Since our 2020 and 2021 studies have wrapped, we’ve:

• Rolled this feature out to English-language enabled settings (May 2021)
• Launched an experiment to test changes to prompt wording to determine if we can increase the efficacy of the prompts (August 2021)
• Rolled this feature out in Portuguese to all people in Brazil (March 2022)

In addition to the changes we’ve made to Twitter based on these learnings, we’re still continuing to understand how we can improve the efficacy of prompts and other forms of intervention to encourage healthier conversations on Twitter. 

Acknowledgements 

We would like to thank our co-author Matthew Katsaros, as well as the following people for their contributions and feedback: Alberto Parrella, Stefan Wojcik, Shaili Jain, Charis Redmond, Cody Elam, Allen Lee, and the rest of the Incentives team at Twitter.