Fact or fiction: The bias in AI
The promise and problem
The existential promise of artificial intelligence (AI) has fascinated us from the early days of computing, when the internet emerged as a unifying force, setting the stage for the imminent AI revolution. As a convergence of cutting-edge technologies in machine learning, data analytics, computing power and high-speed connectivity, AI has the potential to revolutionize every facet of our personal and professional lives, from healthcare and education to finance and beyond. However, alongside this promise and power, comes challenge and responsibility, particularly the issue of bias, which can influence the outcomes of AI systems and perpetuate existing societal flaws. In this article, we will explore the intricate interplay between reality and fiction in AI, uncovering how bias can manifest in AI systems from their very inception and how the blurring of lines between truth and falsehood can uncover potential opportunities for future work.
The emergence of OpenAI's ChatGPT and its introduction of Artificial Narrow Intelligence (ANI) [1] to the mainstream in November 2022, along with its remarkable progress and impact in recent months, has further cemented the role of ANI as a driving force for business and societal advancement. Numerous reports are regularly published, showcasing diverse use cases that highlight the utility of ANI, prompting industry leaders to integrate ANI into their business models to gain a competitive edge. Nevertheless, it is crucial to objectively acknowledge that ANI possesses unique characteristics that make it susceptible to bias.
Mental exercise
Let's engage in a brief exercise. Take a moment to close your eyes and immerse yourself in the experience of driving your dream car. Hold that image in your mind for a moment, and now open your eyes. Let's quickly recap: What make and model was the car? How fast were you going? What color was it? Did you imagine a sunny day? Were you cruising along the picturesque Pacific Coast Highway or navigating a serene rural road in Kansas? It's fascinating to note that the answers to these questions may vary greatly from person to person, influenced by our internal biases. Similarly, it's crucial to acknowledge that ANI also possesses inherent biases. However, unlike human biases, these biases have the potential to become entrenched societal norms, spreading as supposed facts rather than editorial data points.
Nurture versus nature
Bias refers to systematic deviations or prejudices that can impact information processing, decision-making and perceptions, leading to unfair or unbalanced representations of reality. When it comes to ANI, it's crucial to understand how both nature (inherent characteristics) and nurture (external influences) can inadvertently introduce biases.
In the context of nature, biases can be unintentionally or intentionally introduced into ANI algorithms based on assumptions, objectives and design choices during development or due to the technology’s inherent limitations. For example, if an ANI system is designed to optimize for certain outcomes without considering ethical consequences, it can result in biased decisions or actions. In such cases, understanding client goals, along with the design mechanisms of ANI, can help mitigate inherent biases in algorithms. Additionally, auditing the internal controls of algorithms and design can ensure that proper risk management measures are in place as ANI becomes a more significant component of client decision-making.
On the other hand, nurture plays a role in ANI bias as well. ANI models are primarily trained using large amounts of data, often generated by humans from sources such as blogs, scientific papers or smartphones. However, this data can contain biases, leading to biased outcomes or predictions. For instance, if an ANI system is trained on data that focuses only on the positive attributes of a goldendoodle and the negative attributes of a golden retriever, it may result in biased outcomes when asked which breed provides the most value. As such, accountants and auditors have opportunities to contribute to the design of internal controls over applications by leveraging contextual understanding of the data sets used to train ANI to ensure client goals are properly contemplated.
Future impacts
ANI’s continued advancement means there will be a need for evolving regulations to govern ANI ethics and bias in the business world. For instance, in New York [2], recent regulations mandate employers to conduct audits of automated decision-making tools used for evaluating job candidates and employees, aimed at addressing potential biases in these processes. The American Institute of CPAs (AICPA) has also anticipated [3] outcomes for ANI and auditors, including assisting clients in implementing ANI tools into their processes, which can offer new assurance engagement opportunities. The future of ANI and work remains uncertain. Dell Technologies predicts that 85% of the jobs that will exist in 2030 have not yet been invented. Nevertheless, one thing is clear: Mitigating bias will pose a significant challenge that will require dedicated resources.
Final thoughts
The question of whether ANI is biased is complex, as it can be a combination of both or either, depending on various factors. ANI is inherently subjective and biased, but it has the potential to evolve with unbiased assistance. However, it is crucial to recognize and understand the limitations of ANI while exploring effective governance measures to mitigate bias, especially with the introduction of more regulations.
As ANI continues to improve at an exponential rate, many futurists predict the transition to Artificial General Intelligence (AGI) [4] (a topic for another discussion) may be closer than previously anticipated.
