Bias In AI

As AI becomes increasingly important to many fields, the racist and prejudiced views of our American society infiltrate its systems in ways that lead to algorithmic bias. The term artificial intelligence was coined in 1956 at a conference at Dartmouth College where it was defined as the “science and engineering of making intelligent machines” by John McCarthy. However, the field has grown rapidly since then and various definitions have followed suit. AI has a place in many of our daily lives because it increases the convenience and efficiency of many everyday activities. For example, the recommendations given on Netflix or any shopping site based on what the user has previously watched or shopped are based on machine learning algorithms. Since the future of technology is so dependent upon artificial intelligence, it is crucial that we ensure that the systems are fair so as to not create the same inequalities we have in the past, today.
In 1951, one of the first working AI programs was created for playing chess by Alan Turing. This program was improved upon by IBM who created Deep Blue. In the 1980s, Gary Kasparov claimed that AI chess systems could never reach a point where they could defeat him and other grandmasters. This remained true until 1997, when Deep Blue successfully defeated the world champion. The modern era of AI can be traced to the 1990s when Stuart Russell and Peter Norvig published their book, Artificial Intelligence: A Modern Approach, which became one of the foundational textbooks for the field. Today, AI is making large impacts in fields such as healthcare where it diagnoses diseases like Acute Kidney Injury more easily and faster than clinicians have historically been able to. There are various definitions of AI we have today, but for this discussion we’ll be defining it as IBM did in 2020, as “[a technology] that leverages computers and machines to mimic the problem-solving and decision-making capabilities of the human mind.” In 2021, AI algorithms are trained with data and make decisions based on what they learn from that information. The data being used are documents from the past which often include bias and discrimination againist race, gender, and more. The artificial intelligence technology learns these biases which creates inaccurate results and an overall lack of fairness for certain groups of people.

Facial Recognition is a system that relies heavily on AI algorithms, but the testing data available lacks diversity. The current standardized data used for this software is called “Faces in the Wild.” It is comprised of 80% white people and 70% males. The accuracy of the software, when identifying people of color or women, is very poor because of the lack of the data available. In Saving Face: Investigating the Ethical Concerns of Facial Recognition Auditing, Joy Buolamwini and colleagues looked at three companies: Microsoft, Face, and IBM, and the technologies that classified 1, 270 images from African and European countries. The subjects were grouped by gender, skin color, or both. The accuracy at identifying the images overall is pretty good, but there are notable differences because of race and gender. Looking at gender, Microsoft’s percent accuracy for a lighter male is 100% while only 98.3% for a lighter female and 94% for a darker male while only 29.2% for a darker female. This difference in accuracy between identifying faces of specific gender is consistent throughout all the companies. A large difference occurs when taking a look at skin-color: the worst recognition for all three companies was for darker females where they failed at identifying one in three women of color. This lack of accuracy stays consistent throughout many softwares where facial recognition is used. In 2015, AI in a Google online photo service grouped 80 photos of a black man and his friend into a folder labeled “gorillas.”

This illustrates a shortcoming of machine-learning technology because without enough diverse training data, software cannot be properly trained to categorize and recognize images. Another example that demonstrates this issue is when a black researcher in Boston realized that the AI system that she was working on could not recognize her as a person until she put a white mask on.

The historical institutions that have been perpetuating this sense of inequality that social movements have been attempting to remedy since the 1960s are the same institutions that are being utilized to influence AI data sets today. In the fall of 2018, BERT was developed as a search engine tool by Google to understand search queries better. The system uses information from news articles, wikipedia, and books, which are all filled with biases that may have been formed decades ago. When computer scientist Robert Munro put 100 simple words like “action,” “jewelry,” and “book” into BERT, he found that in 99 out of the 100 cases, BERT associated the words with men. “This is the same historical inequity we have always seen,” Robert Munro states in We Teach AI Systems Everything, Including Our Biases. The inequality represented in technology is not so different from the social fights for equality happening right now. It is crucial that we place an equal emphasis on the way people are treated online as we do in real life.

On a daily basis, users can experience the consequences of the shortcomings of training AI with biased data. For example, when Microsoft released a chatbot system on Twitter in 2016, its purpose was to interact with people, answer their questions, and provide better service. However, it learned with public data and had an internal learning feature. Because of this, it used anti-semitic, racist, and misogynistic language towards some of the users when replying to their messages. On the macro scale, we see criminal justice systems potentially demonstrating biases against specific races or socioeconomic groups because of their historical overrepresentation. We also see systems incorrectly judging less wealthy neighborhoods at a high risk of contracting diseases when determining health insurance benefits. As the usage of AI algorithms by the government continues to increase, it is crucial that we understand the implications of its use. There is no straightforward solution to the issues governments will face when implementing these algorithms as simply transparency is not enough. The more complex and better the performance of the system, the harder it becomes to understand and so, “machine learning will often be more useful the more of a black box it is,” as stated by Diane Coyle in her article The Tensions Between Explainable AI and Good Public Policy. This makes it difficult to design ways to evaluate and monitor a machine learning system. Additionally, policy making does not have one clear goal since many compromises are made throughout the process. In order for these algorithms to work effectively, there needs to be a well-defined goal. Lastly, there is an overall lack of trust towards policy-makers. The first step would be to address this issue rather than algorithmic biases.

According to Morgan Livingston’s article, Preventing Racial Bias in Federal AI, in 2020, the National Institute of Standards and Technology issued a report recommending that federal agencies regulate AI systems, but there is currently no federal policy doing this. From the 64 federal agencies using or considering using AI, none have addressed establishing protocols that assess the potential ramifications of bias. An example of the issues stemming from this lies in healthcare. Since black communities have historically lower health costs because of unequal access to care and this was mirrored in the prediction. With federal policy, this bias could have been eliminated. One of Livingston’s main recommendations is to perpetually have the agency workers be informed about the biases in AI and have the right to record and correct the errors of the algorithms as they are noted.

The bias in AI is often more rooted in society than the actual system itself. The 2019 AI Now Institute report also recommends AI bias research to move beyond technical fixes. “It’s not just that we need to change the algorithms or the systems; we need to change institutions and social structures,” explains Rankin. From her perspective, to have any chance of removing or minimizing and regulating bias and discrimination, there would need to be “massive, collective action.” Joy Buolamwini, a postgraduate researcher at the Massachusetts Institute of Technology, realizes the repercussions of algorithmic bias in our society, and to address it, she founded the Algorithmic Justice League. The organization’s primary goal is to highlight the social and cultural implications of AI bias using art and scientific research. The work of such organizations will be essential to addressing biases in artificial intelligence. Nevertheless, there are various technical fixes to the biases as well. McKinsey Company along with many others have developed ways for business practitioners and policy leaders to minimize bias. Another method is tools, some that have been created are AI Fairness 360, IBM Watson Open Scale, and Google’s What If Tool. These tools are all made for testing AI algorithms with more accurate data. In the future, we aim to use the power of AI in a way that delivers outcomes that are fair and beneficial to everyone. To accomplish this, there are two essential steps that need to be taken. First, before an organization deploys an AI system, it must employ a third-party audit that is regulated by the government to mitigate bias. Additionally, organizations should take a holistic approach in researching bias because many of the issues require perspectives from experts not in the computer science field such as social scientists and ethicists.

Bibliography

• Coyle, Diane. “The Tensions between Explainable Ai and Good Public Policy.” Brookings, Brookings, 15 Sept. 2020, www.brookings.edu/techstream/the-tensions-between-explainable-ai-and-good-public-policy/.
• IBM Cloud Education. “What Is Artificial Intelligence (AI)?” IBM, 3 June 2020, www.ibm.com/cloud/learn/what-is-artificial-intelligencexa.
• Kantarci, Atakan. “Bias in AI: What It Is, Types & Examples of Bias & Tools to Fix It.” AIMultiple, 17 Apr. 2021, research.aimultiple.com/ai-bias/#:~:text=AI%20bias%20is%20an%20anomaly,prejudices%20in%20the%20training%20data.
• Livingston, Morgan. “Preventing Racial Bias in Federal AI.” Impacts of Emerging Technologies on Inequality and Sustainability, vol. 16, no. 02, 2020, doi:10.38126/jspg160205.
• Metz, Cade. “We Teach A.I. Systems Everything, Including Our Biases.” New York Times, 11 Nov. 2019.
• Nouri, Steve. “Council Post: The Role Of Bias In Artificial Intelligence.” Forbes, Forbes Magazine, 4 Feb. 2021,
• Photopoulos, Julia. “Fighting Algorithmic Bias in Artificial Intelligence.” Physics World, 28 May 2021, physicsworld.com/a/fighting-algorithmic-bias-in-artificial-intelligence/.