Don’t believe everything you read on Twitter. New research shows that roughly a quarter of tweets are not credible.
False tweets, posts or videos often go viral, boosted by news websites thirsty for clicks, but it’s not always easy to know what’s true and what isn’t. Now tools are arriving to help.
CREDBANK, a database compiled by computer scientists at the Georgia Institute of Technology in Atlanta, is one. It couples crowdsourcing with machine learning to filter and study our social networks.
Advertisement
Researchers Tanushree Mitra and Eric Gilbert started by scraping up just 1 per cent of the tweets in Twitter’s entire feed. Their software filtered and trimmed the tweets for spam before automatically sorting them into topics.
The tweets were then sent to human workers on crowdsourcing site Mechanical Turk to confirm the topics and rate the messages on scales of certainty and accuracy.
The sorting takes place right after an event unfolds on Twitter, taking just a few hours. Over 96 days in 2014, the system assessed 60 million tweets about 1000 news events.
Their base finding is that roughly a quarter of tweets could not be trusted. The biggest hoax they found in the data was last year’s “Ebola zombie virus” rumour, a story about Ebola victims rising from their graves. It seems obviously untrue but was still shared on social networks millions of times.
Mitra and Gilbert have made CREDBANK publicly available so that others can hone its judgement and use it to build apps. The study was presented at the AAAI Conference on Web and Social Media in Oxford, UK, last week.
“Nowadays propagation and verification operate at a different timescale,” says Iyad Rahwan of the MIT Media Lab. “Understanding the dynamics of rumours and falsehood will help us reduce this gap by building tools that speed up verification.”
Google is pioneering a similar service for search results. Its Knowledge-Based Trust is a research attempt to rank pages according to the number of “true” facts they contain, rather than how many links point to them.
A system could be developed from CREDBANK that would automatically notice your uncle’s unbelievable reports of a celebrity death, says Christian Sandvig of the University of Michigan.
“It would then alert you that his tweets match similar reports that were not credible,” he says.”I imagine a future where bizarre stories, political hatchet-jobs, misinformation, scams, and rumours are marked with different warning shades of red.”
