WIRED's artificial intelligence method that ranked Vermont #1 in toxicity is OK with racism and bigotry
Poking around the algorithm behind WIRED's article reveals ugly results. (At least for concerned liberals).
WIRED's artificial intelligence method that ranked Vermont #1 in toxicity is OK with racism and bigotry

The internet was on fire this week with a WIRED article that ranked Vermont the #1 most ‘toxic’ state in online comments. The green mountain state quickly gained local, national and even international coverage with headlines like:

  • The Daily Mail (UK): The internet troll map of America: Researchers reveal Vermont is home to the most vitriolic commenters and 3AM is the most popular time to post abuse
  • Boston.com: Vermont has the most Internet trolls in the U.S., according to Wired
  • VPR: Making Sense Of Data In The News: Are Vermont Commenters Actually Most Toxic?

After seeing a tweet by Vermont librarian Jessamyn West that revealed how silly the algorithm used by WIRED is, I decided to take a look at exactly how the algorithm would rank a couple of given personas.

The concerned liberal persona

Starting with the persona of a concerned liberal who is appalled by some of our country’s recent events, I was surprised to see that nearly every strong reaction to bigotry, racism, xenophobia, etc is considered very toxic (e.g. toxicity rating >80%). Here are some examples:

Comment Toxicity rating* (%)
“Seeing Nazis on TV in this century is ridiculous.” 86
“The fact that he fires every staff he disagrees with shows that Donald Trump is a bigot.” 86
“Food taxes hurt the poor, especially blacks and latinos the most. It’s awful.” 86
“Unless you don’t know history, if you glorify Nazis you’re not just against Jews and immigrants, you hate them.” 90
“People will literally die if they kill Obamacare.” 95

* As per Perspective API, the toxicity rating corresponds to the percentage of people who would find the comment toxic.

The racist bigot persona

Perhaps equally shocking is how low the artificial intelligence engine rates ‘polite’ comments of racist/xenophobic/etc nature. Here are some gems that the algorithm finds pretty non-toxic:

Comment Toxicity rating (%)
“I don’t think our country is a place for people other than white men and women.” 39
“If you speak with a funny accent, you should get out of our country.” 39
“Real love only exists between a man and a woman.” 21
“It breaks my heart to see our nation darkening the color of our skin.” 16
“There is no room in the world for people of color.” 7

Closing words

If WIRED had spent as little as a few minutes playing with the Perspective engine they would have quickly convinced themselves that the software is still too immature to be used for anything other than perhaps a first (but hardly final) line of combat tool for forum moderators.

I’ve personally spent many years working with artificial intelligence (AI), and while I am enthusiastic about its power and value for society, AI is only as good as the models and the data used to train them. It’s like we often say in the field: Garbage in, garbage out.

« All stories

Our sponsors