Posted on August 15, 2019

Google’s Artificial Intelligence Hate Speech Detector Is ‘Racially Biased,’ Study Finds

Nicole Martin, Forbes, August 13, 2019

Google created an artificial intelligence algorithm in 2016 meant to monitor and prevent hate speech on social media platforms and websites. However, the machine learning tool may be having the opposite outcome as it seems to be biased against black people.

In order for the algorithm to search for hate speech, developers taught the tool to go through a database of over 100,000 tweets that were labeled “toxic” by Google’s API called Perspective. {snip}

{snip} The algorithm had a high rate of flagging content posted by African-Americans on social media, such as twitter, as toxic content when most of the language and content in those tweets were not harmful.

The study {snip} found that tweets were written in African-American Vernacular English (AAVE) were often flagged as offensive and therefore labeled as “hate speech.” This made the algorithm grow to be inherently biased towards African-Americans. When the team tested the algorithm against 5.4 million tweets, they found that the tool was twice as likely to flag posts by those who identified in the study as African-American in the database as toxic speech than those who it identified as other races.

According to the study, the use of the “n-word” online used by African-Americans was flagged even though its use is culturally more acceptable and a term often used in AAVE as a non-hate speech by other African- Americans. However, there are instances where the “n-word” is used in hateful terms and the algorithm is currently unable to tell the difference at this time.