AI now permeates our lives through job applications, credit scoring systems, public services, and even healthcare access. With the integration of AI systems deeply rooted into society, questions of fairness have turned from optional to critical.
While advocating for innovation and efficiency, AI technology also presents concealed hidden risks of human bias replication or amplification. The United Kingdom’s future is more data-driven than ever, and with the growing implementation of AI in sectors like education, policing, and the welfare system, understanding AI bias is an issue that cannot solely rest in the hands of computer scientists- it requires much broader public discourse.
What Are Bias and Fairness in AI?
Defining AI bias is when systematic and repeatable errors are made within an AI system, which then has a negative result on specific groups of people. These biases are not random and oftentimes stem from deep-rooted historical imbalances, poor training data, or assumptions deeply rooted in the foundational framework of algorithms guiding the designs.
Business ethics fairness focuses on ensuring that AI systems do not discriminate against people based on their gender, race, socioeconomic status, or other prejudiced characteristics. To put it simply, fairness in AI is one of the most difficult metrics to define and achieve simply because what is ‘fair’ is subjective. It can have multiple meanings, such as achieving equal outputs or providing equal treatment.
What Is the Main Reason for Bias in AI Systems?
Discrimination that occurs within AI systems can occur for a variety of reasons, but the root cause of the problem almost always stems from discriminatory data. AI systems learn from the datasets that they are given. For example, AI may reflect biased data in the absence of women in workforce sectors like technology or vacancies of certain ethnic groups in healthcare services, the AI will, most probably, follow such trends.
But there are other equally important problems contributing to bias in AI systems:
Overrepresentation of Specific Groups
If a facial recognition program is developed with little diversity among the participants in terms of culture, religion, and race, then it will turn out to be skewed. The same has been claimed in practice.
Labelling Errors
Misclassification of data, whether carried out by human annotators purposefully or mistakenly, will distort the learning outcomes of AI.
Proxy Variables
Harmless variables such as postcodes may serve as substitutes for income level, which may heighten discrimination.
Modelling Choice
When developing AI, there are different models to choose from, including ones that might focus on accuracy over fairness. This can have a significant effect on the bias or fairness of the output given by the AI system.
Why Does AI Equity Matter More Now in The UK?
AI is being incorporated into more aspects of day-to-day life in the UK, often operating unnoticed. From police predictive algorithms that identify crime “hotspots” to automated eligibility checks for government assistance. As these systems are used across different sectors, if bias enters into these AI systems, the impacts can cause discrimination and bias, causing large amounts of devastation.
Example of AI Bias in the UK
We can look at the A-Level grading algorithm incident in 2020, where students were assigned grades by an AI system which severely penalised pupils from low-performing schools. Outrage made the government withdraw the results, but not before a significant number of students had already suffered tremendous stress and anxiety from the situation.
Is It Possible to Ever Achieve Perfectly Fair AI?
True fairness in AI in its absolute form may never be possible. The fairness construct, to give a simple example, is diverse and has different meanings based on context and culture. While one person may consider being treated ‘equally’ in terms of ‘equal outcome’ to not be discriminatory, others may believe in equity over equality.
How Can We Improve Fairness in AI?
There are multiple ways that fairness can be improved in AI systems, including regular bias audits of the systems, ensuring inclusive data sets and more:
Bias audits: Organisations can conduct impact assessments of their models before deployment.
Inclusive datasets: Capturing relevant populations in data collection helps mitigate data gaps.
Fairness-aware algorithms: Models are being built that account for varying definitions of fairness.
Transparency and analyses: Understanding how an algorithm arrives at a conclusion makes bias easier to identify.
Regulation and oversight: Enforcement provisions are absent from the UK’s AI regulatory policy framework, but there’s mounting demand for legislation.
Human Values in a Machine World
A driving notion around fairness in AI is that we do not want to allow machines to inherit or aggravate existing inequities humans have created. AI, despite its immense capability, is void of the ability to distinguish right and wrong. So, if you’re a policymaker, a business leader, or just an observer of how technology influences your daily life, this question is for you: Is AI making these decisions fairly?
