The Next Level of Discrimination: Algorithmic Bias

Recently, my family finally decided to transition from the Wii console (yes people still use these) and invested in a PS4 with the Camera. When we were setting up the camera, it would not detect us. We went online and searched up what we could do. We ended up buying 2 huge lights to add to the living room and the camera still couldn’t recognize. However, our mom (who is a few shades lighter) happened to walk behind us and the camera immediately caught her.

This was the basis for me exploring and writing this article on algorithmic bias. I’m sure many of you have heard about it but just assumed that this would be a problem that someone would eventually fix the problem. But through this article, you’ll understand that algorithmic bias is more complicated than that and we also have a part to play.

How does AI fit into my PS4 situation?

A facial recognition system uses AI biometrics — which uses algorithms to read the geometry of your face and record key factors like the distance between your eyes or distance from forehead to chin — to map facial features from a photograph or video. The software identifies these facial landmarks and in result creates a facial signature that is then compared to a database of known faces to find a match or validate the features it distinguished.

The way AI is designed is to learn the same way as a child. With a dataset, AI can find patterns and build assumptions based on those findings and like a child, this process is repeated over and over again until perfection.

However, because AI is applying it’s algorithms to vast amounts of data, the biases within the design, the algorithms, and the data can become magnified at a fast rate. And suddenly we’re looking at a giant projection of a traits we didn’t even know we had.

How can we solve AI Bias

The only way to solve AI Bias (like solving a question in physics class) is to first understand the problem? What is AI bias? That’s actually a hard question to answer because “bias” can mean different things depending on the context. For this article, we’ll stick to the general one that most people think of.

Algorithmic bias occurs when a computer system reflects the implicit values of the humans who created it.

I do recommend reading this article to know about other AI Biases. It is important to know AI Bias doesn’t really come form the AI models themselves but from the people who create the model. And for those who say that AI Bias comes from the data and not people, remember that people create the data.

The reason AI bias is a problem is because whether intentional or unintentional these biases are mimicked or exaggerated. Though AI Bias can be seen as an insight to understanding real-world biases; it is something that affects the results of whether YOU get a job, credit card, jail and so much more.

4 Causes of AI Bias

1. The Data reflects our Current Bias

AI Biases are not just a reflection of who we are today, but of everything that’s led up to this cultural state since the beginning of history — determining who lives in safe neighborhoods, who can afford education, who gets good jobs and who goes to jail. The two main reasons for the lack of data that leads to our current bias.

Reason 1: Minorities make up the largest percentage of low income households and many don’t have access to the internet, devices and apps necessary for contributing to the big data that’s fueling our AI systems.

Reason 2: Minorities that do have access, don’t trust the system enough as they have been used against them so they don’t provide they try their best to not contribute to the big data that’s fueling out AI system’s.

The lack of diversity in our data is how Google ended up with their infamous photo search fail in 2015, when the machine learning models that power their search engine labelled a photo of African Americans as gorillas.

Unlike the Google example, there are examples of bias explicitly being embedded into AI that is evident but we don’t ever seem to talk about. The gender of our most popular AI assistants and the type of tasks assigned to them.

Have you noticed that AI assistants, like Amazon’s Alexa and Apple’s Siri have female voice assistants to do tasks that suggests that women are “subservient and tolerant of poor treatment.” compared to male voiced Watson that is used to solve more complicated problems. I’d recommend reading this article about the situation.

2. Not enough representation (The Minority Effect)

AI lives on data, like humans live on food or water. If you think of Maslow’s Hierarchy of Needs, a theory of motivation that is depicted as a pyramid — many might have seen this in science class when learning about the flow of energy in the food chain — the most important needs at the bottom contribute to the main thing at the top. That is why, when many people talk about AI bias, they focus on the data first because the AI model will only be as good as the quality of data collected.

Data must be gotten from somewhere, if you don’t have enough current ones, you dig into the past and I think we can all agree that so many things have changed since last year. And whether it’s race, gender, age or religion — minorities would have been underrepresented wherever they went. There are more women in engineering now than there were 10 years ago. As a predictive system; using past items to predict the future or define the current status would not only be inaccurate but moving us backwards.

One of the most misunderstood things when collecting data in hopes of not having bias is believing that if you don’t collect mainstream data like gender or race then there would be less or no bias. However data that might not be mainstream can still have correlations with mainstream features. Let’s take this example where the data collected was based on looking for hot spots locations.

The popular game “Pokémon Go” used geographic data based on another game also created by Niantic, called Ingress. According to Urban Institute, “Ingress used to allow players to suggest relevant portal locations in their areas, but because Ingress players tended to be younger, English-speaking men, and because Ingress’s portal criteria biased business districts and tourist areas, it is unsurprising that portals ended up in white-majority neighborhoods.” In white-majority areas, there were on average 55 portals while in black- majority areas there were 14 portals on average. This indeed resulted into racialized’ increase social division.

3. Data is amplified through a positive feedback loop

A positive feedback loop means amplifying what happened in the past. This is one of the biggest what contributes to the bias found in crime prediction algorithms like PredPol. PredPol is a crime prediction

PredPol uses data that was already biased by past housing segregation and cases of police bias. The AI model would be frequently sent to minority neighborhoods. The arrest in the neighborhood increase, in the arrest gets fed back into the algorithm. Then the AI model would predict more future arrest in the same neighborhood and send the police there again. Even if their were crimes in other neighborhoods, the AI model wouldn’t send police officers there because their haven’t been any arrest made in those areas. The AI continues to repeat this process feeling that it has solved the problem but instead amplifies the error even more.

4. Malicious or Intentional or Manipulative

Last but not least, we’ll always have the people who try to use AI for malicious intent, either by embedding the bias intentionally or through manipulation. Take for instance, Twitter bot Tay. An AI chatbot that was made to learn (mimic) the language of teens. I’d recommend watching from 2:31.

A trustworthy AI model will still contain biases because bias (in its broadest sense) is the foundation of machine learning. The key question to ask is not is my model biased? but how can I mitigate negative bias? Because as Robin Hauser said “AI is meant to interact with humans, but if you add the human factor and don’t account for human behavior then you run into loosing control.”

What’s being or can be done to Mitigate AI Bias?

Maintaining diverse teams is important to avoiding and mitigating unwanted bias. As much as we gain continuous lip service on diversity by tech executives, minorities are still under-represented. Through this article and many others, AI/ML models continue to perform worse on minorities.

With more diversity in AI teams, these issues around unwanted bias can be noticed and mitigated before the model is released into production. After all, the first people to notice these issues are users who are female and/or people of color.

We need to be intentional about the data that we collect. More data doesn’t necessarily mean less bias. Instead, it could account for more bias as more unique features might be missed because the AI system is looking for patterns. If one model is different from the patterns in the database the AI system would easily ignore it.

One big IDEA to mitigate and monitor bias in AI models before releasing it to the public is testing AI models for bias like clinical drug trials. In this article by Kalev Leetaru, he suggest that AI developers would declare the demographics they designed their algorithms to work best for. They register with a testing agency that would test its accuracy on each of those demographics and generate report detailing on how it performed.

The report would come out looking like a nutrition label and be provided to the public. You might be wondering how this method helps? This is a way of monitoring the bias. For companies to keep loyalty of their costumers they will be held accountable to mitigate their AI systems for bias before releasing their product to society.

There have already been a variety of approaches taken to mitigate AI. Algorithms that help detect and mitigate hidden biases within training data or learned by the model. Processes that hold companies accountable and discuss the different definitions of fairness.

However, as much as we have this tools, not much progress has been in the past years to mitigate AI. AI developers have been aware of these biases and yet, I find myself not being able to use AI features to the full capabilities as others.

If your looking for a tool to use, check out IBM’s AI Fairness 360. An open source toolkit that helps developers test for bias in their datasets.

If you liked this article, please follow. I also have a monthly newsletter that you can subscribe to here!