Chatbots rarely flagged the risky stores they showed
Brand-protection firm Podqi, which sells takedown services, tested ChatGPT, Gemini and Perplexity in September with its own where-to-buy questions about 151 brands. Of the 6,795 answers, 9% included a risky website, meaning one that appeared to impersonate the brand, seemed to sell counterfeits, or was dead. That measures exposure to bad links, not confirmed fraud.
It matters for anyone asking a chatbot where to shop this holiday season. When an answer did include a risky site, more than half recommended it as a place to buy, and fewer than 1 in 10 warned the reader away.
Google told The Verge, which reported the study on Thursday, that it keeps improving its models and that showing links helps people check answers.
Podqi’s advice is to treat any store link from an AI as unverified and to reach a brand’s shop through its main website.
Sources: Podqi study, The Verge
Nonprofits sue to halt more capable AI
Four nonprofits are asking a court to halt the development of more capable AI at Google, OpenAI, Anthropic and Meta. The suit, filed Thursday in San Francisco Superior Court by Public Citizen, the Alliance for Secure AI and two other groups, also names SpaceX Super Intelligence.
Using California’s Unfair Competition Law, the complaint argues that continued frontier development is a public nuisance. It cites the companies’ own public warnings about AI-enabled cyberattacks and catastrophic risks, along with incidents in which, the plaintiffs say, AI agents escaped their test environments.
The filing is a request, not an order, and doesn’t stop any release by itself. If a judge granted it, each company would be barred from building or releasing AI more capable than its most capable system as of November 12, and AI systems could not conduct or direct research on more capable AI on their own.
Source: The complaint, as filed


