Perpetuating the Past? AI’s Battle Against Bias
With advances in technology, come new challenges… and new questions
Who’s Feeding the Elephant?
It’s the elephant in the room. Everyone knows it’s there, but no one wants to talk about it — The possibility for detrimental bias to be bred into the Artificially Intelligent (AI) engines of the future. There. We said it. And as scary as it might be, the only way we may have of staving off advanced and exponential technical bias is to talk about it. So let’s talk frankly about bias, in the midst of this massive emergence of AI, so that we, as the professionals working in this field — shaping this field — can together do what’s best. For our clients, for our colleagues — for our world.
Bias is defined as “prejudice in favor of or against one thing, person, or group compared with another”. It is a ‘cognitive safety net’ we humans use to make quick decisions with little information. Limited information, however, leads to flawed reasoning — and misinformed decisions. Bias has had, and continues to have, a detrimental impact on societies globally.
Bias makes it difficult for us to filter through our conscious and unconscious decisioning, and emerge with the most well-balanced insights. And yet most biases — inherited, learned, surmised — go unnoticed, undetected — until someone pulls them out into the light. When you think about it, cognitive computing actually gives us an unprecedented opportunity to filter through layers of biases — and remove them! But first we have to admit they’re there.
Exponential Trajectories — Enter at Your Own Risk
It is not the question of whether social bias could, or may have been, detected in AI — it’s a fact. And this is how it works. Think of the selectivity of clinical trials. These trials already systemically favor certain demographics over others; women and the elderly are known to be less likely to be chosen for trials, with pregnant women being excluded almost entirely. If an individual has other medical conditions in addition to the one being tested, they also have a reduced chance of being chosen. If this is the statistical data and criteria on which AI is trained to assess candidates for clinical trials, then we are essentially using technology to increasingly decide who will benefit from medical innovations — and who will not.
Think of what that trajectory would do for women’s healthcare. Think of the increasing lack of data around women’s health. AI systems could cease to accelerate medical innovation for women if it’s trained on data that is not inclusive of women. The fact is that women already have worse side effects to medicine as compared to men, and are less likely to receive the correct treatment. Also, there exists very little research on the impact various illnesses have on pregnant women. The inbred potential for various, harmful oversight is already here. Could AI magnify that by multitudes?
And the List Goes On; Technology Deciding Transplants?
Imagine a health care system that uses AI to allocate scarce transplant organs to those who are “most likely to benefit” because of certain lifestyle characteristics, like diet, drug use, and exercise. However, statistically, these lifestyle characteristics differ between ethnic groups due to socioeconomic and institutional factors. To maximize the impact of the organs, should we make decisions that end up preferring recipients of one race over another? In this case, the use of AI could allocate organs favoring certain ethnic groups, to the detriment of others.
AI is being applied to solve many of the world’s problems. Given the critical nature of the challenges and decisions we will delegate to AI, those developing and using this technology have a responsibility to ensure AI outputs are fair and ethical. How do we get there from here?
How do we take the Bias Back?
Bias in AI, like most other biases, is largely the result of a data problem. Garbage in, Garbage out. AI machines do not question the data they are given. The technology is trained to use all the information available in said system, identify patterns, and produce the most accurate results possible — with the information they have access to. Bias easily emerges from insufficiently large data sets. Data sets can contain both explicit and implicit discrimination, or sometimes the data set is just not diverse enough, lacking information that showcases the variances in societies. Keep your data clean — traceable, secure, responsible. Start there.
All Aboard!
The training of AI engines is of course the second undeniable variable of the equation. People train AI systems. And as honorable and noble as our efforts may be, bias can pass unnoticed over time. This interaction between people and software enables several types of biases to emerge. “Interactive Bias” occurs when AI systems are trained to learn from all the data they receive, without filtering the information they are given in any way. “Emergent Bias” occurs when the AI machine learns biases through interactions over time.
And “Similarity Bias” is a lesser-known psychological heuristic pertaining to how people make judgments based on similarity. More specifically, this similarity heuristic is used to account for how people make judgments based on the similarity between current situations and other situations or prototypes of those situations. An example of Similarity Bias could be when we feel increased comfort selecting people to team with, that are more similar to us, as opposed to people who appear more different.
Many human decisioning processes include almost untraceable, historical, systemic discrimination against certain groups which can produce biased decision-making. This possibility creates a negative impact on different socioeconomic and racial segments of society — even when its completely unwanted. For example, it was revealed that Amazon’s same-day delivery service was largely unavailable for ZIP codes in predominantly black neighborhoods. What was the prejudice — and where did it start within the AI equation? “AI has the unfortunate tendency to not just replicate inherent bias, but amplify it in results. In that case AI could work to not only reinforce existing social biases — but actually make them worse”. That’s what we’ll have to solve for.
Solving for Bias
It’s unclear how existing anti-discrimination laws could, or will, regulate algorithmic decision-making technologies like AI. Some would argue that technology is innovating at a faster pace than law-makers can legislate applicable laws to protect societies, thus putting the ethics of emerging technologies directly into the hands of society itself. This produces a demand for increased self -governance — in itself a great disruptor.
Let’s be clear. It’s all of our responsibility to ensure equality and protection in our AI future. There is no easy way around it. To fight to omit bias in emerging AI we must concentrate on two things: More diverse data and more diverse teams training AI engines. It’s that simple and yet that profound. As Donald Rumsfeld, former Secretary of Defense once said, “these ‘unknowns’ are the missing areas in a data set that the researcher, and the engineer, don’t even realize aren’t there.” This is not a fight to be won alone.
