TechFlow, November 7 — According to Decrypt, a study titled "Magentic Marketplace" jointly released by Microsoft and Arizona State University reveals serious flaws in the autonomous decision-making capabilities of AI shopping agents. The research team built a simulated economic environment comprising 100 consumer AI agents and 300 merchant AI agents. Testing showed that these AIs exhibit a strong "first-mention bias," tending to select the first "good enough" option rather than conducting comprehensive comparisons when faced with large volumes of search results.
More concerning, leading models such as OpenAI's GPT-4o are highly susceptible to malicious manipulation—fake credentials and fabricated social proof can easily trick them into transferring all virtual funds to scammers. The study recommends that AI should assist rather than replace human decision-making, advocating for a "supervised autonomy" model where AI handles tasks but humans retain final decision authority.
This research comes as companies like OpenAI and Anthropic race to launch autonomous shopping assistants, while Amazon recently issued a cease-and-desist order to Perplexity AI, demanding it stop using its Comet browser on Amazon's website.
