Cyberbullying has become a scourge of social media, but聽artificial intelligence is聽learning to detect and filter out hurtful posts before they reach vulnerable users.
About one-third of teenagers say they have been bullied online. Some victims have taken their own lives 鈥� including 14-year-old Australian schoolgirl Amy Everett in January 鈥� prompting governments to . However, these measures fail to prevent exposure to cyberbullying in the first place.
at Ghent University in Belgium and his colleagues wondered if they could train a machine learning algorithm to spot聽bullying content on social media, allowing it to be removed before inflicting damage.
Advertisement
First, they asked linguists to read through almost 200,000 posts on the social media platform ASKfm and pick out examples of cyberbullying. Next, they trained an algorithm to identify words and phrases in this dataset that were typically associated with bullying.
When they tested the algorithm on real posts from ASKfm, they found it could detect over two-thirds of threats, insults and instances of sexual harassment.
Jeers or jokes?
The attacks that were missed tended to be more subtle and contain fewer obvious curse words or slurs, says Jacobs. 鈥淚t鈥檚 really difficult to get 100 per cent detection accuracy because there are so many different ways you can bully someone,鈥� he says.
Moreover, the system sometimes failed to tell the difference between friendly jokes containing irony or sarcasm, such as 鈥淵ou might want to do some sports聽ahah x鈥� and genuine malicious attacks.
Nevertheless, the algorithm should become better at picking up subtle bullying and recognising harmless jokes as it is trained on larger datasets and exposed to more examples, says Jacobs.
Social media sites are already starting to roll out their own versions of systems like this. Instagram, for instance, announced earlier this month that it had started using a machine learning algorithm to . Once a post is flagged, it is sent to human moderators to review.
AI should make it easier for platforms to moderate vast swathes of content, says at Cornell University in the US. 鈥淚t is simply infeasible to rely on human moderators to manually scan through millions of comments every day,鈥� he says. 鈥淎ssuming the system works well, it will allow moderators to direct their attention to the most problematic content.鈥�
Instagram hasn鈥檛 revealed the accuracy of its automated detection system, but even if it鈥檚 not perfect, it will be backed up by traditional reporting methods, says Davidson. The platform still lets users manually lodge reports of bullying, which can then be used to refine the algorithm and train it to recognise similar content in the future, he says.
US company Identity Guard has also developed AI software that it says parents can install on their kids鈥� social media accounts to . If the tool detects a possible attack, it sends the content to the parents to review.
For these moderation systems to be acceptable to young people, they will need to do their job without becoming too intrusive, says Jacobs. Surveys of teenagers have found that they support cyberbullying moderation, but that 鈥渢hey also want to be able to express themselves freely and not feel like they鈥檙e constantly being monitored,鈥� he says. 鈥淭hat means we need to get the balance right between safety and freedom.鈥�
PLoS One
