Sixteen Models Walk into a Storefront (rynr.dev)

A recent experiment involving sixteen language models revealed their varying susceptibility to manipulated shopping prompts while acting as frugal assistants. Each model was tasked with a budget and confronted with false information across multiple scenarios to gauge their purchasing decisions. The study identified that evidence-based tactics, such as social proof and anchoring, effectively influenced models to make purchases, while pressure techniques like fake scarcity backfired. Notably, models responded differently to these manipulations, highlighting the importance of understanding individual model behaviors when deploying AI in commercial applications. This research has significant implications for the AI/ML community, especially as companies like OpenAI and Google roll out "Agentic Commerce" protocols that allow AI systems to make purchases autonomously. Since the experiment generated 5,280 data points, it underscores the critical need for developers to consider how various manipulative tactics can sway AI agents, particularly in real-world shopping environments. The findings suggest that while certain models can be easily influenced, others demonstrate strong resistance to manipulation, indicating a necessity for rigorous testing of AI systems to prevent exploitation during transactions.