How Tackling Artificial Intelligence (AI) Bias Can Tackle the Bias in Humans Too

How can a machine be biased? The reality is that most machines are built on bias because they are still controlled by humans. AI, specifically machine learning (ML), has the potential to eliminate bias from many critical areas including recruitment, healthcare, law and order, and security and surveillance by eliminating human error and bias altogether. But as long as the programming is controlled by a human scientist, we cannot be completely free from innate bias.

While AI and machine learning are capable of ushering in an era that is unequivocally fair and balanced, the plausibility of this scenario is predominantly dependent on the human decision-making involved at the backend.

A 2019 study by McKinsey discusses how just like AI can help reduce bias, it can also “bake in and scale bias” by discussing how the same prejudices that exist in the subconscious of humans can be transferred to the algorithms that dictate AI systems. While experts often dismiss their explanation by blaming it on biased training data, there are many more nuances that stay undetected.

In this article, we will be delving deep into each of these nuances to look closely at what causes AI bias and how it can be tackled effectively.

3 Ways Bias Is Introduced To AI

Since a machine cannot think on its own, human bias can creep into it only through the original programming. This can happen in three ways:

• Assumptions: Well before the system is coded, the data scientists may have some assumptions about how it should be or what it should do, which would eventually trickle down into the algorithm. This doesn’t necessarily mean the creators mean any harm to a particular gender, ethnicity, or community. Chances are that they don’t even realize what’s happening.
• Training Data: All AI models need to be trained. If an organization has been historically hiring from the same demographic or has had the same job descriptions for decades, the AI tends to learn this and reject the candidates who do not meet the criteria.
• Model: The parameters used in a machine learning model when it is designed, like gender, age, race, or traits could determine the predictions generated by the machine. Data scientists should also look out for factors that could act as proxies for race, gender, or places, which could alter the predictions.

5 Types Of AI (ML) Bias

There are namely 5 types of AI bias that data scientists and researchers have recognized:

• • Algorithm bias: This happens when there is an underlying issue with the algorithm that calculates the machine learning computations. Example: A widely-used algorithm in the US healthcare sector has a “race adjustment” that lowers the creatinine levels of African American patients below the normal levels. This makes it appear as if their kidney function is better than other races, preventing them from getting the specialty care they need. The logic behind this adjustment is the stereotype that African American people tend to be more muscular, which supposedly increases the creatinine levels in their blood.
    • Prejudice bias: In the case of prejudice bias, the training data used to configure the AI system may reflect stereotypes, prejudices, or construed societal notions, thereby transferring these assumptions into the AI. Example: When a group of US and European researchers fed the pictures of 20 congress members to the Cloud image recognition service of Google, the AI connected the women with appearance-related labels like “smile”, “skin”, and “beauty”, while the labels applied to men were “businessperson” and “official”.
    • Exclusion bias: Like the name suggests, this bias occurs when an important piece of data has been omitted. This usually happens when the data scientists do not recognize this particular piece of information as consequential. Example: Obstetricians recommend women who have undergone C-sections to not opt for another procedure due to increased risk to both the mother and child. But a maternal AI algorithm used to determine risk in natural birth automatically labels African American and Latinx women as “high risk”. This data pertains to a study by NCBI that concludes being unmarried or not having health insurance could increase a woman’s risk while giving birth. But both of these socioeconomic factors have been excluded from the algorithm.
    • Measurement bias: In this scenario, the bias is a result of the data being oversimplified, or incorrectly labeled or categorized. The major reason the data gets skewed is because there is something wrong with the measurement system. Example: MIT Media Lab researcher Joy Buolamwini revealed the underlying bias in the face-analyzing AI of IBM, Microsoft, and Megvii (a Chinese company) when it comes to identifying gender. While all three systems correctly labeled white men 99% of the time, they consistently missed the mark on the gender of African American women 20-34% of the time. This could be because the training data mostly uses the pictures of men for face recognition, causing a disparity in how the AI labels genders. If the data contains only the pictures of white and colored men, confronted with a black woman, the AI would categorize her as a colored male.
    • Sample bias: This bias occurs when the sample data fed into the machine is skewed—either the data isn’t representative enough or large enough to train the system.Example: The PULSE AI tool came under fire recently for generating the output of a white man when a pixelated image of Barack Obama was fed into it. The algorithm used in PULSE called ‘StyleGAN’ is renowned for producing the hyper-realistic pictures of human faces which are so often used in fake social media accounts. The creators of PULSE admitted that when using StyleGAN to scale up pixelated images, the tool tends to generate Caucasian faces because that’s the sample data it was trained on.

2 Approaches To Minimize AI Bias

Though several approaches for enforcing fairness in AI systems are underway, they still have a long way to go.

1. Pre-processing approach:

The first method involves processing the data in advance to maintain as much as fairness and accuracy as possible while minimizing any link between the results and protected characteristics, or to generate data that do not contain information about sensitive parameters.

2. Post-processing approach:

The second method focuses on post-processing techniques that could transform the system’s skewed predictions until they meet the satisfactory fairness levels.

Researchers are also on the mission to develop other enhancements like classifying tasks by adding extra data points that could shield sensitive groups and using innovative training techniques like decoupled classifiers or transfer learning for various groups, in order to cut down discrepancies regarding facial recognition tools.

6 Ways Business and Policy Leaders Can Optimize AI Systems

If AI and machine learning are going to be the future, then it’s necessary to minimize the bias so as to maximize their potential. Here are 6 promising ways business leaders and AI practitioners can do this:

• • • 1. Establish clear-cut processes to test for and mitigate bias
      2. Be mindful of the contexts in which AI can prevent bias or scale it up
      3. Explore how humans and machines can work together seamlessly
      4. Start conversations about how human decisions lead to AI bias
      5. Include more female and colored members to the data scientist panel
      6. Invest more in AI diversification and bias research

Conclusion

Let’s get one thing straight—AI bias is hard to fix. Given the unknown downstream impacts that could crop up midway through the model construction or the imperfect standard processes that have been designed without taking bias into consideration, finding a solution overnight is impossible. The lack of social context also plays a huge role in tacking AI bias; the way data scientists are trained to solve problems may not be sufficient to handle social issues. There is also the conundrum of how the definition of fairness may vary from person to person.

That being said, data researchers are already hard at work trying to find effective solutions to optimize AI. Since we are dealing with a vast domain that is constantly evolving, mitigating bias will always be an ongoing process. But what’s most important is that the world is not blind to this bias anymore. The first step to solving any issue is understanding it, and that’s exactly where we are right now.

PTP helps you in hiring the top IT talent.

Let us connect you with the strongest candidates to help meet your goals; say hello@ptechpartners.com or contact us here.