The dangers of letting AI agents do your shopping

Vass Bednar is the managing director of the Canadian SHIELD Institute and co-author of The Big Fix.

Meta just launched its Muse AI agent; a personal assistant designed to handle tasks such as e-mail, online shopping and travel booking. Some people assign these bots to find and negotiate discounts on flights, or to locate the lowest possible price for something.

Apparently, Wall Street loves Muse, available so far only in the United States and Canada. Meta stock has surged 20 per cent since its launch. Muse was behind the Nasdaq index’s 2.26 per cent rise to a record close on Monday, some analysts have said.

Shopify has announced that it would use Meta’s Muse for agentic checkout, effectively embracing agents as legitimate online shoppers.

But 12 days after the launch, Amazon blocked Meta’s agent on its platform because of concerns over security, privacy and the violation of Amazon’s Conditions of Use. Amazon also sells its own shopping agents, and its growing advertising business still depends on human eyes landing on sponsored listings. After all, bots don’t click on ads.

As AI upends shopping, retailers tailor their pitches for chatbots

Moreover, a study of more than two million product listings by Productrise found that when the same product appeared in Google’s AI Mode and in regular Google search on the same day, the AI version showed a price about 22 per cent higher on average.

Google says it hasn’t verified the findings. But the more unsettling number in the study is that barely one per cent of the products in regular search results showed up in AI Mode at all. Traditional search would surface more options, and it seems like the AI version offers far fewer items, picked for reasons that aren’t disclosed.

And those picks can be easily gamed. Consider influencer Deana Burke’s fake deodorant experiment, where she got a fake brand into chatbot recommendations with very little effort. Her research demonstrates that AI shopping advice is more concentrated, less independent and easier to manipulate than it looks.

The economic conditions that gave us competitive couponing in the early 2000s are worse now, and the online environment is becoming more opaque and manipulated. A growing Reddit trend shares crumpled receipts from pre-2020. Past prices are admired like fossils or hieroglyphs – ancient moments that we can’t reclaim. People are literally building their own personal price indexes from old receipts. There has got to be a better way.

The bots can also cause a ruckus. In Australia this year, a man asked his AI agent to get him into a popular gym class. Instead of joining the waiting list, it found a hole in the gym’s booking system, booked him weeks ahead and bumped the person at the top of the list. Meta’s own director of AI alignment watched her agent delete her inbox while she typed “STOP.” An early agent tested by the Washington Post bought a dozen eggs without asking, for $31.

On Tuesday, major banks warned that agentic commerce creates unresolved problems around fraud, mistaken purchases, privacy, and who bears responsibility when the bot gets something wrong. In turn, some retailers are rewriting their terms of service so that the customer carries the cost of the agent’s mistakes. So now, a software program’s error is our problem.

Canada signs international statement on AI safety, joining calls for stronger oversight

Ottawa, for its part, has been asking what to do about all this. Its consultation on AI transparency, which includes figuring out how to track what AI agents do on our behalf, just closed.

Europe isn’t waiting. Britain’s competition regulator has said plainly that consumer law applies whether a sales practice is carried out by a person or a bot. The European Union is preparing a Digital Fairness Act, and its advisers are questioning a basic assumption of consumer law: that a human being makes the final decision on whether to buy something.

We need to ensure that AI agents are traceable and verifiable when they buy and sell on behalf of users. Mandatory requirements of agent ID and verification of their authority to act as a user’s proxy are minimum requirements to guarantee that we continue to have trackable and stable commercial systems.

We should regulate these agents before they go rogue and write the rules themselves.