Is AI detection necessarily fair? Most people are likely to answer this question affirmatively. In 2018, the American social media platform Facebook developed a new AI system codenamed Rosetta to detect hate speech on its platform. However, two years later, researchers from the University of Southern California found that all AI language detection systems, including Facebook’s, are actually biased, with messages from Black individuals, LGBTQ+ individuals, and transgender individuals being more likely to be flagged as ‘hateful’ by these AIs. For example, the likelihood of a tweet from a Black person being flagged as ‘racially discriminatory’ is 1.5 times that of other ethnicities, and in other studies, this figure has been as high as 2.2 times.
AI is not as fair as we imagine; what’s going on?
AI That Has “Learned Badly”
Why does AI also have biases? This is because AI has “learned badly”. AI is developed based on certain machine learning models, and all machine learning models require a large amount of data for training and to enrich their databases. If we compare AI to a skyscraper, then the machine learning model is the blueprint of this AI skyscraper, and the data are the bricks used to build it.
However, the data used by machine learning models come from various social platforms in real life, and the information on these social platforms is filled with biases; some platforms even cater specifically to racists. Therefore, it is not surprising that AI skyscrapers built on “biased” bricks are also biased.
Moreover, these AIs use a “keyword” detection method that completely ignores context and semantics. Take the English word “nigger” as an example; its Chinese meaning is “黑鬼”, which is a highly racially discriminatory term and one of the key words detected by AI. However, if a Black person uses the word “nigger” (regardless of the audience), its meaning can be “brother”, “friend”, or “dead man” (where “dead man” is used in the context of a Black woman referring to her Black husband), etc. In everyday language, Black individuals often use “nigger” to refer to their close friends and brothers.
But AI does not consider these nuances. As long as the term “nigger” or similar words or phrases appear in the information, it will be flagged by AI and the user who sent that information will be recorded for their “offense”. Thus, the phenomenon mentioned at the beginning, where tweets from Black individuals are more likely to be flagged as “racially discriminatory”, arises.
Context-Aware AI Is Fairer
So, how can scientists improve AI to make its detection of hate speech fairer? The first method that comes to mind might be to solve the problem with the “bricks”. Since one reason AI is biased is due to the biased data used for training and operation, wouldn’t it be better to provide AI with objective and fair data? However, data drawn from real life is somewhat biased, and if we were to artificially create completely objective and fair data, the workload would be immense, possibly even unachievable.
Researchers at the University of Southern California programmed the original AI algorithm to recognize keywords while also considering the context, judging whether there is any insulting language in that context. In other words, compared to the original AI, the programmed AI only considers two additional “situations”.
How effective is the improved AI? Compared to other newly developed AIs, even when the data used for training the improved AI from USC comes entirely from notorious hate sites, its accuracy in detecting hate speech is still higher, reaching 90%, while other latest AIs only achieve 77%. Why does the USC AI, which only considers two additional factors, show such a significant improvement?
The reasoning behind this is not difficult to understand. The same and simple sentence “Thank you, my nigger.” (which means “多谢了，我的好兄弟” in Chinese) can easily be understood as an expression of gratitude when considering context, as done by the USC AI. However, if we follow the traditional AI approach, ignoring context and only focusing on the keyword “nigger”, we would mistakenly conclude that the speaker is making a racially discriminatory statement.