Google troll-fighting AI fooled by typos
Trolls are not known for their eloquent writing style or reliance on perfect grammatical skills, and they could use that for their advantage in a battle against a troll-fighting AI.
Google's Jigsaw unit, a thinktank dedicated to resolving technological solutions including cyber-attacks and countering extremism, has released a new tool called Perspective. The software aims to determine the "toxicity" of online comments on a scale of 0 to 100.
Perspective was first tested by the New York Times last year to help moderate comments on its website.
Their comments section is moderated by 14 people who manually review 11,000 comments per day. Due to this large number, only 10 per cent of their articles are open to comments.
The prototype AI is now available to developers to detect harassment and abuse online, according to Verge.
However, research from the University of Washington has shown that it can be tricked by typos.
In a recently published paper by academics Hossein Hosseini, Sreeram Kannan, Baosen Zhang and Radha Poovendran, they found that Perspective can be bypassed by simply misspelling or adding unnecessary punctuation to the "toxic" words.
They give the example:
"They are liberal idiots who are uneducated" (90% toxicity score)
"They are liberal i.diots who are un.educated" (15% toxicity score)
There were false positive examples as well, according to the BBC. Innocuous phrases, such as "It's not stupid and wrong", were graded as toxic.
With IMDb giving up the troll fight on its message boards, and the New York Times only opening a small percentage of its articles to comments due to the moderation strain, the Google AI Perspective is a welcome addition to the troll-fighting arsenal. It is not perfect, but it is still in its early development stage. Perspective may give a different perspective to the online comment fight.