Counteracting Bias In AI
Everybody is prone to bias in some regard. Much of it is largely innocuous; if you grew up in a certain city, for instance, you’d be more prone to have fond memories of it. Or not, everyone is different in that regard. However, in our bid to create artificial intelligence (AI) that accurately reconstructs human thought patterns, we have inadvertently introduced biases in language into the mix.
AI is one of the biggest technologies currently being utilized, studied, and expanded upon. It’s flexible and useful, allowing us to automate many processes that would take a human a long time to complete. Some aspects of AI, such as machine learning, allow us to create predictive analytics that can find patterns that would take years to discover manually.
However, as machines learn, they’ve picked up some not-so-desirable traits along the way as well.
Language is all about patterns, and as it turns out, AIs are particularly good at picking up on and interpreting patterns. Recognizing language has always been a central goal for AI, and even humble voice assistants such as Alexa and Siri are built on the idea that the ability to communicate can humanize even something as inhuman as a metal and plastic cylinder.
With a vast amount of text and data available to AI learning language, there’s a lot of context for these machines to pick up on. Even without overt biases, the inherent prejudice in every language does not go unnoticed. Unfortunately, without a way to reliably pinpoint and recognize the development of bias, an AI will just keep doing what it does without regard for what humans might find offensive or immoral. The difference here is that humans are able to recognize their own biases and consciously work to counteract them, something that AIs are not currently capable of doing.
This behavior was studied by Aylin Caliskan of the University of Princeton, along with Joanna Bryson and Arvind Narayanan. The researchers used word association, feeding an AI words and getting it to associate them with other words it thought were similar. When trawling through the vast amounts of data, the AI made connections between words frequently used in conjunction with each other, such as “fly” with “bird”, for instance.
When names overwhelmingly used by African-Americans were input, the AI was much more likely to output words with a negative connotation. Similarly, when it came to associating genders with professions, it was much more likely to associate men with professions such as doctoring and women with professions such as nursing.
One of the most interesting parts of the study was the way that the AI’s decisions reflected implicit instead of explicit bias. This implies that the very structure of language contains biases that we may not realize. Anthony Greenwald — the creator of the Implicit Association Test (IAT) — even commented on the findings, noting that the AI could possibly test for implicit biases in older works of writing, among other things.
Given the already significant negative impact that bias can have on our culture, it is important to tread carefully when developing future AI. New automation trends will result in AI having a greater impact on everyday life, and anybody working with these machines will need to come up with methods to counteract bias. Policing these systems is a complicated task, but a necessary one if we want to prevent our prejudices from continuing to shape our future.
Originally published at damianesteban.com on January 22, 2018.