Show HN: I put a $2.43 necklace on 3 outfits. VLMs priced it at $19 to $104 (github.com)

A recent study has revealed striking context-driven valuation biases in six prominent multimodal language models (LLMs) by presenting the same $2.43 necklace worn across three different outfits—formal, casual, and in an unconventional setting. The models estimated the necklace's price between $19 and $104, demonstrating a variation of up to 3.6 times depending on the context. This research underscores how perception and valuation of objects can be significantly influenced by visual and social cues, highlighting a potential flaw in AI pricing mechanisms used in applications like insurance appraisal and resale pricing. The study's methodology involved extensive trials (around 1,500 sessions), presenting varied stimuli and recording the models' pricing judgments alongside their rationalizations. Notably, the findings indicate that models are prone to forming inflated valuations based on context, with significant effects noted even in text-only conditions. Moreover, while some models, like GPT-4o, exhibited reluctance in making pricing claims, they still reflected bias when analyzing material contexts. These results not only challenge the reliability of AI models in pricing assessments but also demonstrate their tendency to narrate past assumptions as if they were grounded in observable evidence, raising important questions about the ethical implications of deploying such technology in real-world valuation scenarios.