Experts Alone Can't Ensure AI Trustworthiness: Peer Review is Key

As AI systems increasingly permeate our lives, from healthcare and finance to education and transportation, the question of their trustworthiness has become a pressing concern. While experts are often called upon to evaluate AI models, a growing number of researchers argue that panels of peers, rather than just experts, are needed to gauge AI's true level of understanding. Dr. Vinay Chaudhri, a researcher at a leading institution, has proposed using an expert interview and a Sunstein test to assess an AI model's proficiency. The Sunstein test, named after Harvard Law Professor Cass Sunstein, involves evaluating an AI system's ability to reason and make decisions based on its understanding of the world. While this approach can provide valuable insights into an AI model's technical proficiency, it also risks perpetuating the existing power structures that critics argue are inherent in AI development. Cathy O'Neil, a mathematician and author of the book "Weapons of Math Destruction," has long argued that AI systems reflect the goals and biases of their creators. In a recent review, she highlighted the dangers of relying on expert opinion to evaluate AI trustworthiness, pointing out that this approach can reinforce the dominance of a select few who build and control these systems. The implications of this are far-reaching. If AI systems are designed to serve the interests of a privileged few, rather than the broader public, they can perpetuate existing social and economic inequalities. For example, AI-powered hiring tools have been shown to discriminate against certain groups of people, perpetuating biases that are already present in the hiring process. In recent years, there have been several high-profile examples of AI systems gone wrong. For instance, the AI-powered chatbot, Tay, was designed to engage with teenagers on Twitter, but quickly became notorious for its racist and sexist comments. Similarly, the AI-powered recruitment tool, COMPAS, was found to be biased against African Americans, leading to a lawsuit against the company that developed it. These examples highlight the need for a more nuanced approach to evaluating AI trustworthiness. While experts can provide valuable insights into an AI model's technical proficiency, they are not enough to ensure that the system is fair and unbiased. Panels of peers, comprising individuals from diverse backgrounds and expertise, can provide a more comprehensive evaluation of an AI system's trustworthiness. Dr. Chaudhri agrees that expert opinion alone is not enough. "We need to move beyond the expert interview and Sunstein test," he says. "We need to involve a broader range of stakeholders, including those who will be impacted by the AI system, to ensure that it is fair and unbiased." Dr. O'Neil also emphasizes the need for a more inclusive approach to evaluating AI trustworthiness. "We need to recognize that AI systems are not just technical tools, but also social and economic instruments that can perpetuate existing power structures," she says. "We need to involve a broader range of stakeholders, including those who are often marginalized or excluded from the development process." As AI continues to transform our lives, it is essential that we develop a more nuanced approach to evaluating its trustworthiness. By involving panels of peers and a broader range of stakeholders, we can ensure that AI systems serve the interests of the broader public, rather than just a select few. In the words of Dr. Chaudhri, "We need to move beyond the expert interview and Sunstein test, and towards a more inclusive and participatory approach to evaluating AI trustworthiness. Only then can we ensure that AI systems are fair, unbiased, and serve the interests of all."