Reducing Bias: Ensuring Just and Reliable AI Outcomes

In today’s rapidly evolving technology landscape, AI stands out as one of the megatrends shaping the future of industries. Recent developments have given us a glimpse of the economic and political impact that AI can achieve. For instance, China’s release of DeepSeek has underscored the transformative potential of AI across all facets of society. Society is driven by the decision-making capabilities of human beings, both in individual and collective capacities. The decision-making process of a human being is affected by multiple factors such as knowledge, experiences, the environment, and conscious/unconscious biases and stereotypes. Traditionally, AI has been regarded as a tool to assist in decision-making processes and, in certain contexts, to make autonomous decisions. However, this raises critical questions about AI's capacity to make unbiased and fair decisions. A significant concern is that AI systems can inadvertently inherit and even amplify human biases present in their training data. A biased decision made with the assistance of AI can have profound financial and reputational consequences for organizations. Nowadays, with more business functions adopting automation for their processes, the chances of such biased outcomes are higher. In a recent example, one of the fastest-growing quick commerce companies in India sent a ‘miss you’ notification about a contraceptive pill to a female customer, as part of its marketing techniques. The customer was offended and posted about it on social media, causing a dent in the company’s brand image.

The Nature of Human Bias in AI

Human bias in AI refers to the inclination or prejudice for or against a person or group, especially in an unfair manner. Biases can be explicit, involving conscious thoughts, or implicit, comprising unconscious attitudes or stereotypes. These biases influence decision-making processes, often without the individual even realizing it.

Types of Human Biases

• Confirmation Bias: The tendency to seek out, interpret, and recall information in a manner that confirms one's existing beliefs and preconceptions.
• Anchoring Bias: The tendency to place extreme importance on the initial piece of information encountered (the ‘anchor’) when making decisions.
• Selection Bias: The bias introduced by the non-random selection of data, which can lead to misleading or skewed outcomes.
• Groupthink: The practice of thinking or making decisions as a group, which discourages creativity or individual responsibility.

The Transference of Bias to AI

There are two major foundations for any AI solution: Mathematical models/algorithms and data. A complex AI model trains itself on the data which is fed into it. This process is iterative, and the AI model keeps learning on the vast amount of data it gathers or is fed. The output of an AI model depends on the quality of the data it learns from. Human biases can infiltrate AI systems in several ways:

Data Collection

The data used to train AI models may contain biases. Imagine a hiring process at a manufacturing organization where most common bias or stereotype is that female candidates are not good engineers or suitable for shopfloor jobs. An AI model trained on such data will have this stereotype in its output too.

Algorithm Design

The process of developing an AI algorithm is mostly human. The selections made by an AI engineer can embed bias in the AI. For example, the decision to include particular features in a model can result in biased outcomes if they are the result of individual biases/preferences.

Feedback Loops

AI models often learn over time from interactions with users. If users' interactions are biased, the AI will learn to replicate those biases. For example, a recommendation model that provides suggestions based on user behavior will replicate the prejudices (if any) in the loop.

Impact of Biased AI

The consequences of biased AI can be long-lasting and harmful. Biased AI systems can:

• Discriminate: Against certain ethnic groups/genders, leading to unfair treatment in areas such as hiring and law enforcement.
• Perpetuate Inequality: Bias in AI can reinforce existing social inequalities, making it harder for marginalized groups to achieve parity. For instance, in the US housing market, an AI model was found to reject up to 80% of mortgage applications from black families, maintaining historical discrimination.^[1]
• Undermine Trust: Public trust in AI will weaken if AI systems are perceived as unfair or biased.

Preventing Bias in AI

Preventing bias in AI requires a composite approach involving careful consideration at different stages of development and deployment. Diverse Data Collection Ensuring that the training data is representative of the diverse populations that AI will serve is important. Actively looking out and including data from diverse groups helps avoid skewed outcomes.

Inclusive Algorithm Design

Diverse teams of developers (gender/ethnicity/geography) should be involved in the design and development of AI models to reduce the risk of embedding biases/stereotypes. Regular assessments of algorithms should be done to identify and mitigate bias.

Transparency and Accountability

Organizations should be transparent on the data on which their AI models are trained and their decision-making processes. Implementing accountability mechanisms ensures oversight and responsibility for AI outcomes.

Continuous Monitoring and Feedback

AI systems should undergo continuous monitoring to detect biased outcomes. Incorporating user feedback and adjusting models accordingly can help mitigate bias over time.

Conclusion

In a recent article, Forrester identified five principles of AI ethics, namely fairness and bias, trust and transparency, accountability, social benefit and privacy and security.^[2] Recognizing that AI systems can inadvertently inherit and amplify human biases is crucial in our pursuit of ethical and equitable technology. As AI becomes increasingly integrated into decision-making processes, it's imperative to prioritize AI ethics to ensure fairness and trustworthiness. By remaining vigilant about the ways biases can transfer from humans to AI, and by implementing strategies to reduce and prevent such biases, we can develop AI systems that are more just and reliable.

References

• How AI is hardwiring inequality — and how it can fix itself, Brunel University of London, By Hayley Jarvis, 18 Dec 2024: https://www.brunel.ac.uk/news-and-events/news/articles/How-AI-is-hardwiring-inequality-%E2%80%94-and-how-it-can-fix-itself
• Five AI Principles To Put In Practice, Forrester, by Brandon Purcell (VP, Principal Analyst, Forrester), 13 Apr 2020: https://www.forrester.com/blogs/five-ai-principles-to-put-in-practice/