Reproduce “Utilizing Human Behavior Modeling to Manipulate Explanations in AI-Assisted Decision Making: The Good, the Bad, and the Scary” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zhuoyan Li, Ming Yin
Purdue University researchers developed a computational model of human decision-making that integrates both AI recommendations and explanations. This model enables the algorithmic manipulation of AI explanations, which can either improve human-AI team performance by promoting appropriate reliance or introduce bias, often without human detection.
Key insights
- Human decision-making, specifically how humans integrate AI recommendations and explanations, can be accurately modeled quantitatively, with trained models achieving 0.65 to 0.79 accuracy across tasks.
- AI explanations can be algorithmically manipulated to significantly influence human decisions towards either biased or more accurate outcomes, demonstrating a dual-use capability of such models.
- A critical vulnerability exists: humans generally fail to detect manipulated explanations, even when these manipulations profoundly alter their decision-making processes, highlighting a significant security and ethical concern.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.