Originally published on Forbes
As AI continues to permeate all aspects of our lives, questions of ethics and fairness are being vehemently debated. In a recent survey of IT professionals, 85% stated that their biggest concern was ethics. And rightly so. From gender-biased recruitment tools to racist computers, the implications of AI bias have begun to make themselves seen, heard and felt in a big way.
As technologists, we have a responsibility to identify and eliminate bias in AI. But why is it such a huge challenge and how realistic are our hopes of banishing it altogether?
Both the beauty and fundamental flaw behind AI is that it is designed to mimic human thought and behavior. And as humans, we are by nature subjective and filled with personal biases, whether conscious or not. Though not all these biases have the potential to be malicious – as Google’s shoe analogy illustrates – the fact that biases can and do make their way into data becomes a problem when they carry racial, gender-based and other harmful prejudices.
Ultimately, AI is a mirror. If we feed our systems with data embedded with our biases, those systems will reflect them right back at us – and worse – and out into the world.
The bubble effect
The rise of social media has seen us increasingly influenced by algorithms designed to serve up content that matches our interests. While this does deliver a more personalized experience, in effect it also encases us in a bubble of “similarity” or “confirmation” bias: validating our existing belief sets, while eliminating contrasting viewpoints that might challenge our perspective.
As humans we’re highly susceptible to this bubble effect. And since AI relies on human input and intervention, there’s always a risk that personal bias “bubbles” filter through to data, despite genuine attempts at objectivity.
In her Ted Talk, “The Danger of a Single Story”, Chimamanda Ngozi Adichie reminded us how easily biased views can be formed when not enough versions of reality are represented. This is true not only of how we view others, but also how we view ourselves.
“The single story creates stereotypes, and the problem with stereotypes is not that they are untrue, but that they are incomplete. They make one story become the only story.”
– Chimamanda Ngozi Adichie
Stereotypes and self-bias
Bias is most often something we project outwards, but self-bias is instrumental in propagating stereotypes. For example, it’s been found that women presented with jobs labeled as “Nursing” rather than “Medical Technician” will often tend toward the first, because they unconsciously align themselves more with the female stereotype of a nurse.
If the algorithms behind AI-powered recruitment systems are designed to generate the highest number of clicks, and people tend to click on jobs that confirm their self-view, then they will continue to present jobs that reinforce stereotypes. And so the cycle continues.
History repeats itself
Self-selection aside, bias plays an even greater role in systems that determine the treatment of others. Sometimes, this can be quite deliberate and widely accepted as fair – for example, people applying for credit being assessed on their personal payback history. But when similar systems deny loans to women simply because of their gender, that alleged fairness comes into question.
The US sentencing tool COMPAS, designed to measure the likelihood of recidivism, was found to label black people twice as likely to reoffend than whites, despite criminal records suggesting otherwise. The tool’s developers resisted disclosing details of its proprietary algorithm, which clearly displayed historical racial prejudices with serious consequences.
Unless data collection and algorithmic development are carefully monitored, AI can be a dangerous catalyst for perpetuating bias and allowing history to repeat itself in undesirable and antisocial ways.
Quality vs quantity
A common misconception in AI is that by using multiple samples in datasets, we can ensure a fair result that outweighs human bias. But if the data itself carries bias, a higher quantity will only reinforce it. Though unintentional, gaps or disproportional representation in data can lead to dire results – as was the case with HP’s skin tone issues in their facial recognition software, among countless other examples.
In order to minimize and eliminate bias, data needs to be carefully vetted, right from the point of collection to structuring, testing and tuning. Scale is important, but quality is the key to ensuring training data and the AI systems it feeds are not only accurate, but fair and free from prejudice.
From bias to BHAG
As we’ve begun to understand the impact of AI in our daily lives, so too have its moral and social implications come to light.
In an ideal world, AI systems would be completely objective. But as they’re built by humans, they subsequently end up reflecting and projecting our biases into the world. By understanding bias, the sources from which it originates and the processes by which they infiltrate data, we can actively design systems to avoid, minimize, and – as much as is humanly and mechanically possible – eliminate them.
Is the notion of utterly banishing bias from AI a utopian ideal? Perhaps. Is it a social imperative and the moral obligation of AI technologists to make it our common Big Hairy Audacious Goal? Absolutely.