On Tuesday, Twitter announced a massive change to the way its conversations will work, evaluating not just the content of individual tweets, but the way users behave more broadly on the service. Twitter will now use thousands of behavioral signals when filtering search, replies, and algorithmic recommendations. If it believes you are trying to game its system, or simply acting like a jerk, it will push your tweets lower down. It’s the biggest update so far in the company’s push to create healthier conversations, an initiative announced by its CEO Jack Dorsey in March.
Among the signals Twitter will use: whether you tweet at large numbers of accounts you don’t follow, how often you’re blocked by people you interact with, whether you created many accounts from a single IP address, and whether your account is closely related to others that have violated its terms of service.
“A lot of our past action has been content-based, and we have been shifting more and more toward conduct and behaviors on the system,” Dorsey said in a briefing at the company’s San Francisco headquarters on Monday.
The push is meant to get out ahead of problems that might normally result in an abuse report under the existing system. In testing, Twitter said the changes led to an 8% drop in abuse reports on conversations (the discussions that happen in the replies to a tweet) and a 4% drop in abuse reports in search. These drops, the company believes, indicate that something is working. Fewer than 1% of total accounts make up the majority of those reported for abuse, Twitter trust and safety VP Del Harvey said in the briefing. The company believes identifying a chunk of them and proactively decreasing their reach could generate big results.
“Directionally, it does point to probably our biggest impact change,” Dorsey said. “This is a step, but we can see this going quite far.”
Twitter, of course, has promised many fixes to its problems in the past, and delivered mixed results. It’s still far from troll-free, regularly used to spread misinformation, and home to legions of identity-stealing crypto-scammers. It spent a good portion of 2017 apologizing for its missteps.
The company believes these changes can be effective, particularly because they can be applied globally, without regard for language, since they are entirely action based. “It’s not dependent on hiring more people,” Dorsey said. “It’s a model built into the network.”
Twitter also said more changes along these lines are on the way, but declined to further describe its roadmap.
The company will make the new behavioral filters optional, and they will be on by default. People will be able to turn them on or off with a “show everything” toggle in search. The changes will begin to roll out this week.
Twitter’s product tweaks are meant to roll back the reach of everything from abusive tweeters to scammers using fake accounts to boost their presence in search results. Dorsey was also clear that while he’s optimistic, he’s under no illusion that they’ll fix everything.
“This is not an endpoint,” he said. “We have to be constantly ten steps ahead. Because even a system like this, a new model, people will figure out how to game it, take advantage of it.”