In March, we introduced our new approach to improve the health of the public conversation on Twitter. One important issue we’ve been working to address is what some might refer to as “trolls.” Some troll-like behaviour is fun, good, and humorous. What we’re talking about today are troll-like behaviours that distort and detract from the public conversation on Twitter, particularly in communal areas like conversations and search. Some of these accounts and Tweets violate our policies, and, in those cases, we take action on them. Others don’t but are behaving in ways that distort the conversation.
To put this in context, fewer than 1% of accounts make up the majority of accounts reported for abuse, but a lot of what’s reported does not violate our rules. While still a small overall number, these accounts have a disproportionately large – and negative – impact on people’s experience on Twitter. The challenge for us has been: how can we proactively address these disruptive behaviours that do not violate our policies but negatively impact the health of the conversation?
A New Approach
Today, we use policies, human review processes, and machine learning to help us determine how Tweets are organised and presented in communal places like conversations and search. Now, we’re tackling issues of behaviours that distort and detract from the public conversation in those areas by integrating new behavioural signals into how Tweets are presented. By using new tools to address this conduct from a behavioural perspective, we’re able to improve the health of the conversation, and everyone’s experience on Twitter, without waiting for people who use Twitter to report potential issues to us.
There are many new signals we’re taking in, most of which are not visible externally. Just a few examples include if an account has not confirmed their email address, if the same person signs up for multiple accounts simultaneously, accounts that repeatedly Tweet and mention accounts that don’t follow them, or behaviour that might indicate a coordinated attack. We’re also looking at how accounts are connected to those that violate our rules and how they interact with each other.
These signals will now be considered in how we organise and present content in communal areas like conversation and search. Because this content doesn’t violate our policies, it will remain on Twitter, and will be available if you click on “Show more replies” or choose to see everything in your search setting. The result is that people contributing to the healthy conversation will be more visible in conversations and search.
In our early testing in markets around the world, we’ve already seen this new approach have a positive impact, resulting in a 4% drop in abuse reports from search and 8% fewer abuse reports from conversations. That means fewer people are seeing Tweets that disrupt their experience on Twitter.
Our work is far from done. This is only one part of our work to improve the health of the conversation and to make everyone’s Twitter experience better. This technology and our team will learn over time and will make mistakes. There will be false positives and things that we miss; our goal is to learn fast and make our processes and tools smarter. We’ll continue to be open and honest about the mistakes we make and the progress we are making. We’re encouraged by the results we’ve seen so far, but also recognise that this is just one step on a much longer journey to improve the overall health of our service and your experience on it.