In recent months, generative AI models such as ChatGPT, Google’s Bard, and Midjourney have gained increasing popularity among individuals for both personal and professional purposes. However, research is uncovering a troubling trend: these models are embedding biases and negative stereotypes in their users while also generating and spreading nonsensical but seemingly accurate information.
The consequences of this phenomenon are particularly severe for marginalized groups, who bear a disproportionate burden of the dissemination of fabricated information. As these models become more prevalent on the World Wide Web, there is a growing concern that they could influence human beliefs. A new Perspective article published in the journal Science highlights the urgent need for collaboration between psychologists and machine learning experts to assess the scale of this issue and develop solutions.
Abeba Birhane, an adjunct assistant professor in Trinity’s School of Computer Science and Statistics and Senior Fellow in Trustworthy AI at the Mozilla Foundation, explains that generative models lack the ability to communicate uncertainty, resulting in confident and fluent responses without appropriate qualifiers or representations of doubt. Consequently, individuals may accept these answers as factually accurate, leading to a distortion of beliefs. Furthermore, financial and liability interests incentivize companies to anthropomorphize generative models as intelligent, empathetic, or even childlike, exacerbating the problem.
Disproportionate impact on marginalized groups
The repercussions of biases and fabricated information are particularly pronounced among marginalized populations. The study emphasizes the importance of conducting detailed analyses to measure the impact of generative models on human beliefs and biases. By focusing on marginalized communities most affected by fabrications and negative stereotypes in model outputs, future studies and interventions can effectively address these issues.
One concrete example discussed in the Perspective relates to the use of generative AI models in the legal system. Statistical regularities within these models have resulted in higher risk scores being assigned to Black defendants. Consequently, court judges, having internalized these patterns, may alter their sentencing practices to align with the predictions of the algorithms. Even if regulations are implemented to curtail the use of such systems, this mechanism of statistical learning can perpetuate the belief that Black individuals are more likely to re-offend.
Challenges in overcoming biases and fabrications
Once biases and fabricated information have been accepted by individuals, they become difficult to dispel. Children are particularly vulnerable as they are more likely to anthropomorphize technology and are easily influenced. Recognizing this, efforts should be made to educate the public, policymakers, and interdisciplinary scientists about the workings of generative AI models, dispelling existing misinformation, and countering exaggerated claims surrounding these technologies.
The Perspective article stresses the importance of prompt and comprehensive analysis to evaluate the impact of generative models on human beliefs and biases. It also calls for targeted interventions and resources aimed at mitigating the effects on marginalized populations. Furthermore, education and awareness campaigns should be undertaken to provide realistic insights into how these AI models function and to rectify misconceptions and hype surrounding their capabilities.
As generative AI models gain widespread adoption, the encoding of biases and the fabrication of nonsensical information pose significant concerns. Marginalized communities are disproportionately affected, exacerbating existing inequalities. Collaboration between psychologists and machine learning experts is crucial to assessing the scale of this issue and devising effective solutions. By understanding the impact of generative models on human beliefs and biases, it becomes possible to address and rectify the biases and fabrications embedded within these AI systems.
