AI EDUCATION: What Is AI Bias?

This week on AI Education we’re going to focus on bias in AI. The “how we got here” this week should be a little obvious, we cover some recent research in our AI & Finance column showing both a shortage of trust in AI and a hope that artificial intelligence is going to be less biased than human actors—or not biased at all. Then there was my introduction to AI & Finance a couple of weeks ago, where I lamented the prevalence of extreme bias among journalists, both in the financial trade press and in journalism at large, and my own personal hope that an AI-driven future would somehow produce less-biased journalism than the industry now offers consumers. 

Well, let’s go ahead and dispel some of those illusions right now: AI is already irrevocably biased, it will be biased in the future, and it is difficult to imagine a future in which AI will not be biased. Artificial intelligence today is filled with biases and preferences and is in reality no better than the deeply flawed human beings which design, build and use it. 

To be clear, the technologies underlying generative AI are, in and of themselves, free from bias. Absent content to digest, an artificial neural network is a close to the proverbial blank slate (though I’m not sure which proverb I’m referring to) as exists. The technologies themselves have no preferences or prejudices, they just consume and generate data. 

How, then, can AI be so biased? 

What Is AI Bias? 

Artificial intelligence bias “refers to the occurrence of biased results due to human biases that skew the original training data or AI algorithm—leading to distorted outputs and potentially harmful outcomes,” according to IBM. In other words, the bias doesn’t coy from the technology, it comes from us. Before human intervention, AI is no more biased than a blank page in a word processing program—it takes content to create the bias. 

And we’ve created a lot of content to train AI on over the years, and a lot of that content contains legacies of bias based on identifying characteristics like race, sex, sexual orientation, ability, nationality and geography. The data collectors of the past were often less concerned about accurately representing certain groups of people, leading to skewed data. In some cases, data were falsified to reinforce cultural views on some groups and individuals. 

But AI bias can also be introduced wherever a human interacts with the technology. In some cases, AI algorithms may be designed in a biased manner or trained to create flawed or unfair output. In other cases, AI is prompted or queried in a manner that creates a biased outcome—this was the downfall of many early attempts at AI social media chatbots when malicious users eventually trained the AI to produce offensive results. What’s more, a person can introduce just about any cognitive bias and fallacy that exists in their own minds to an artificial intelligence. 

Trust us human beings to screw up a good thing, right? 

So What? 

I mean, humans are biased to varying degrees—many of us are very biased in our everyday dealings despite claims to the contrary—and we’ve managed to build thousands of years of civilization. Why should we care if AI, too is biased? 

Well, for one thing, biases reduce the accuracy of an artificial intelligence’s output, according to IBM, which may limit its potential and reduce its benefits for businesses: Bias is one of the causes of AI hallucinations. For another, biased output from AI may foster distrust among certain groups and communities, particularly at a time when the technology is pushing to win over more people. In worst-case scenarios, biased AI may touch off scandals that negatively affect businesses. 

Also, importantly, we’re rapidly moving towards a world that is going to be managed, in large part, by artificial intelligences. We need now, at these early stages, to make sure that world is as fair and free as possible, and to instill trust between the people being governed and the people and technologies doing the governing. 

Finally, we’re already presented with risks and dilemmas from various forms of AI bias. 

What AI Bias Looks Like 

• Diagnostic software in healthcare that favors an average or median based on a majority population and is less accurate when diagnosing ailments in minority populations. 
• AI creating insurance premiums based on historically biased data, creating higher costs for certain population groups. 
• Facial recognition systems that excel with faces with lighter skin tones but struggle to identify people with darker skin. 
• Voice recognition systems that struggle with particular dialects, jargon, slang and accents. 
• Generative AI that produces inappropriate or offensive graphic or video content. 
• Content recommendations that are culturally or politically biased, adding to the creation of “echo chambers.” 

HR software that favors certain demographics when making hiring recommendations—particularly concerning in a moment where attempts to mitigate historical human bias are being accused of overreach. 

So, clearly, AI bias is a problem, particularly for highly regulated industries. 

The Good News 

Well, I’ve already delivered most of the bad news, which is that we’re not going to get rid of AI bias altogether. We can, according to IBM, mitigate to a great extent the biases that enter into our technology by building diverse teams and using diverse sets of data to develop and train AI, by adopting and adhering to ethical frameworks to guide how AI is developed and used, and by retaining a skeptical perspective on the fruits of AI’s labor. That sounds an awful lot like what diversity and inclusion initiatives were originally intended to be.