Twitter expands efforts in AI-assisted war on COVID fake news

Twitter has acknowledged it has struggled to contain the volume of COVID-19 misinformation on its platform despite removing 40,000 posts and suspending about 1500 accounts globally since the pandemic began.

New figures in the social media company’s biannual transparency report show that after instituting rules specific to COVID-19 content in April last year, Twitter removed only 3846 pieces of COVID misinformation in the six months from July to December last year. Added to the 4658 pieces of content reported in the previous report, that makes fewer than 10,000 pieces in 2020.

Twitter has published a transparency report for the second half of 2020.Credit:AP

But the company said the total number as of this week is 43,010, indicating a significant acceleration in takedowns over the past six months. Twitter puts this down to difficult conditions at the start of its efforts and new AI-powered detection methods that have been recently developed.

“Varying country-specific COVID-19 restrictions and adjustments within our teams affected the efficiency of our content moderation work and the speed with which we enforced our policies,” the company wrote in its report.

“We increased our use of machine learning and automation to take a wide range of actions on potentially misleading and manipulative content.”

According to Twitter’s rules on COVID-19 misinformation, the company only removes content if it: asserts a fact intended to influence others’ behaviour, is demonstrably false or misleading or would lead to harm if believed.

The spread of misinformation online can be correlated directly to community attitudes that could run counter to health and recovery measures like vaccinations.

Research from Reset Australia found that the number of Australians following anti-vax Facebook groups grew by almost 300 per cent in the first three months of this year, while a University of Melbourne survey showed that between October and February the proportion of Australians willing to receive a COVID-19 vaccination fell from 74 to 66 per cent.

Twitter said its goal was to make sure as few people as possible saw dangerous tweets before they were detected and taken down.

From July 1 to December 6 Twitter removed a total of 3.8 million tweets for violating its rules (not just those related to COVID-19), and has determined that 77 per cent of those tweets were viewed less than a hundred times before removal. However 6 per cent (or 228,000 tweets) were seen more than a thousand times before they were removed.

“Our goal is to improve these numbers over time, taking enforcement action on violative content before it’s even viewed,” Twitter said.

Elsewhere in Twitter’s transparency report, the company said it took action against close to a million accounts for abuse and harassment in July-December 2020 (up 142 per cent from the previous half). The category of violable tweets that saw the largest increase was “non-consensual nudity”, with Twitter reporting action against 27,087 accounts for a 194 per cent increase.

There was a 35 per cent decrease in the number of accounts permanently suspended for terrorist or violent extremist messages, and almost all of them were proactively identified with the help of AI.

Get news and reviews on technology, gadgets and gaming in our Technology newsletter every Friday. Sign up here.

Tim is the editor of The Age and Sydney Morning Herald technology sections.

Share this post