Google builds an AI to fight online hatred

Rooting out the trolls turns out the be harder than expected
Perspective's web page allows you to test sentences for "toxicity." (Google/Alphabet)

Trolls. They used to guard bridges.

Now they burn them, at least online.

Online hatred is one of the a painful reminders of the politically and socially divided world we live in.

To see how nasty some people can be, you only have to visit the comments section of a newspaper, or scroll through a Twitter feed.

And right now, weeding out those trollish remarks is a very difficult job. So is it something we can delegate to AI? Can a machine learn to tell the difference between constructive criticism and outright hatred?

This is what Google has set out to do with its new platform, called Perspective.

Google has partnered with Wikipedia, The New York Times, and some other media companies. The goal is to help them weed out offensive entries and comments on their websites.

It's also released the platform publicly, so anyone can try it out, by typing in a sentence and and having it scored in terms of its "toxicity."

David Auerbach is a former Google software engineer who now writes for the MIT Technology Review.

David Auerbach.

He recently put Perspective through its paces giving it a bunch of random phrases to see how well it could pick the malevolent from the benign.

The results? Not so good.

For example, the phrase, "I f**king love you, man," returned a rate of very high toxicity, even though it's a compliment, if a little rough around the edges.

The words "garbage truck" also returned a high toxicity score.

Conversely, "Some races are inferior to others," scored very low in toxicity.

Why does Perspective have such a hard time discerning the context of a sentence - and whether the arrangement of words in it constitutes something offensive?

Mainly because it's an incredibly hard thing to do. Unlike recognizing the difference between a cat and a dog, or even different breeds of cats, human language doesn't just jive with the kind of pattern recognition AI, or machine learning, is good at.

Language is nuanced and constantly evolving, and, most important, entirely contextual, David says. And it takes a very broad and flexible knowledge of language to be able to tell the whether the word "garbage" is being used to denigrate something, or simply refer to the trash in your kitchen bin.

Theoretically, the more Perspective is used, the better it will become. But David believes its accuracy will only increase incrementally, because of those challenges involved in deciphering context.

Nonetheless, he says the New York Times has notices an off-label benefit: Perspective is able to identify what it calls "quality" comments, that is to say, comments that are intelligent, well-put together and thoughtful. That's because those comments are more likely to have good sentence structure and grammar, which are things that do closely follow a pattern that AIs can recognize.

But removing the trolls? For now, that's going to have to remain a job for humans well-versed in understanding context. And that's a friggin' tough job to do.


To encourage thoughtful and respectful conversations, first and last names will appear with each submission to CBC/Radio-Canada's online communities (except in children and youth-oriented communities). Pseudonyms will no longer be permitted.

By submitting a comment, you accept that CBC has the right to reproduce and publish that comment in whole or in part, in any manner CBC chooses. Please note that CBC does not endorse the opinions expressed in comments. Comments on this story are moderated according to our Submission Guidelines. Comments are welcome while open. We reserve the right to close comments at any time.

Become a CBC Member

Join the conversation  Create account

Already have an account?