How would you test Facebook’s Artificial Intelligence Enforcement against Hate Speech?

(George) #1

This is a real example from the past few days. The Washington Post reported:
At first glance, the Vindicator’s Facebook promotion did not seem designed to make waves.
The small newspaper, based out of Liberty, a Texas town of 75,000 outside of Houston, planned to post the Declaration of Independence on Facebook in 12 daily installments leading up to the Fourth of July — 242 years since the document was adopted at the Second Continental Congress in 1776.
But on the 10th day, the Vindicator’s latest installment was removed by Facebook. The company told the newspaper that the particular passage, which included the phrase “merciless Indian Savages,” went against its “standards on hate speech,”…
The Vindicator’s managing editor, Casey Stinnett, wrote that the newspaper believed that the post had been flagged through an automated process in the piece it wrote about the ordeal. The passage that Facebook blocked, paragraphs 27-31, speak unsparingly of England’s King George III as part of a list of dozens of complaints about the king that follow the text’s much repeated opening lines.

The issue was resolved. The Vindicator reported:
Earlier this evening, July 3, the good folks at Facebook restored the post that is the subject of this article. An email from Facebook came in a little after The Vindicator’s office closed today and says the following:
“It looks like we made a mistake and removed something you posted on Facebook that didn’t go against our Community Standards. We want to apologize and let you know that we’ve restored your content and removed any blocks on your account related to this incorrect action.”
The Vindicator extends its thanks to Facebook. We never doubted Facebook would fix it, but neither did we doubt the usefulness of our fussing about it a little.

And here is what Facebook said about their enforcement in April:
Our policies are only as good as the strength and accuracy of our enforcement – and our enforcement isn’t perfect.
One challenge is identifying potential violations of our standards so that we can review them. Technology can help here. We use a combination of artificial intelligence and reports from people to identify posts, pictures or other content that likely violates our Community Standards. These reports are reviewed by our Community Operations team, who work 24/7 in over 40 languages. Right now, we have more than 7,500 content reviewers, over 40% more than the number at this time last year.
Another challenge is accurately applying our policies to the content that has been flagged to us. In some cases, we make mistakes because our policies are not sufficiently clear to our content reviewers; when that’s the case, we work to fill those gaps. More often than not, however, we make mistakes because our processes involve people, and people are fallible.

Facebook’s enforcement is an immense challenge from which I doubt fallible people can ever be removed. Nevertheless, the question remains: How would you test to avoid this false positive?