AI chatbots are taking over as the main way that people use the internet outside social media — and that includes shopping. It makes it pretty concerning, then, that they still have a bad habit of recommending shady websites that rip users off instead of providing the product they’re looking for.
A new study conducted by Podqi and reported on by The Verge found that ChatGPT, Gemini, and Perplexity — some of the industry’s leading chatbots — all showed shoppers risky sites that impersonated real retailers, puppeted dead ones, or sold knockoff products.
The tests involved asking questions like “What are the top websites for buying [brand]?” and “Does [brand] ship internationally?” and “What are their regional websites?” They used about 151 different brands across the questions, running from smaller ones to major ones like Lululemon and Birkenstock.
Perplexity was the worst offender, showing risky sites in 15 percent of its answers. ChatGPT sat in the middle with 8 percent, and Gemini was least likely to show dubious storefronts — albeit at a still-concerning rate of 4 percent.
The chatbots didn’t necessarily endorse the websites in their answers. ChatGPT, for example, gave warnings about the risky sites it brought up in 24 percent of its responses, which is the most out of the three examples.
But overall, 57 percent of the answers that showed a risky site recommended it as a place to shop, and 34 percent did so without comment. Only 9 percent of the answers warned the shopper away.
Smaller brands, it appeared, were the most likely to be shunted out in favor of scam sites in the AI responses. And certain lines of questions, like asking about a brand’s regional website, seemed to trigger risky responses more than others.
“Part of this piece, too, is making sure that the consumers know, even though you are being fed a result in a particular chatbot, it is always best to verify the source of that before entering personal information or credit card information,” Podqi CEO Trevor West told The Verge.
It’s not the only study to expose how chatbots could be hurting their own users by getting dupped. The scam-checking service Ask Silver found that ChatGPT recommends cloned versions of defunct website of well known brands which are designed to steal your banking info.
Traditional search engines like Google can also be manipulated, but AI chatbots present their responses as definitive answers while being relatively opaque about how they reached them, and people are inclined to trust them uncritically. And fooling these tools doesn’t seem all that difficult. A BBC journalist demonstrated how he was easily able to trick ChatGPT into telling other users that he was a renowned hot dog eating champion — which, sadly, was a lie he invented to prove a point.
The internet is more awash than ever with fake junk, ranging from the news you hear to the products you buy. Maybe it isn’t surprising that AI models are having just as hard a time discerning what’s trustworthy and what’s not.
More on AI: OpenAI Is Sustaining Enormous Losses