A team of researchers has introduced a novel training tool designed to enhance artificial intelligence (AI) ‘s capabilities in recognising when humans provide deceptive information, particularly in scenarios involving economic incentives. The tool addresses a critical issue where individuals may falsify personal data, such as when applying for mortgages or seeking to lower insurance premiums.
As Mehmet Caner, co-author of the study and Thurman-Raytheon Distinguished Professor of Economics at North Carolina State University’s Poole College of Management, points out, AI systems are extensively used in business applications, such as assessing mortgage affordability and determining insurance premiums. These systems, which traditionally rely on statistical algorithms for predictive modelling, inadvertently create a space for individuals to manipulate information to their advantage, leading to the need for the development of more sophisticated AI tools.
The research aimed to adjust AI algorithms to better accommodate these economic incentives for deception. By developing a new framework of training parameters, the researchers enabled AI to adapt its learning process to identify situations where users may have motives to lie. This enhancement focuses on improving AI’s ability to anticipate and account for human behaviour influenced by economic incentives.
In simulated trials, the modified AI demonstrated improved accuracy in detecting inaccuracies in user-provided data. “This effectively reduces the incentive for users to provide misleading information,” Caner explains. Nevertheless, the study acknowledges the challenge of distinguishing between minor falsehoods and more significant deceptions, prompting further investigation into establishing clear thresholds.
The team is now making these cutting-edge training parameters available to the public, with a strong call to AI developers worldwide to embrace and refine their applications. Caner underscores that this advancement is a significant stride towards curbing the economic motivations for dishonesty in AI-interpreted contexts. The ultimate aim is to elevate AI systems to a level where they can potentially eliminate such incentives, thereby fostering greater trust and reliability in automated decision-making processes.
More information: Mehmet Caner et al, Should Humans Lie to Machines? The Incentive Compatibility of Lasso and GLM Structured Sparsity Estimators, Journal of Business and Economic Statistics. DOI: 10.1080/07350015.2024.2316102
Journal information: Journal of Business and Economic Statistics Provided by North Carolina State University