Shopping AIs Can Spot Suspicious “Made in USA” Contradictions, but Buyers May Never See the Warning
A Columbia Law School investigation found that Amazon and Walmart shopping assistants could identify contradictions between “Made in USA” claims and origin details in tested listings, yet the platforms did not consistently surface those warnings. The audit demonstrates an accountability gap, but chatbot answers do not prove corporate intent.
Verified topics and entities
AI shopping assistants are being sold as tools that can compare products and simplify difficult choices. A new investigation found that assistants from Amazon and Walmart could also notice suspicious contradictions in country-of-origin claims, yet those warnings were not consistently placed in front of shoppers.
The 30-second summary
- What happened? Researchers tested Amazon’s Alexa for Shopping and Walmart’s Sparky, and found that the assistants could identify some listings where “Made in USA” language conflicted with origin information elsewhere on the page.
- Why does it matter? If a shopping AI can detect a red flag but the marketplace does not surface it, automation may make purchasing faster without making it more trustworthy.
- What is the catch? This was an investigative audit, not a representative measurement of all listings, and chatbot explanations cannot establish company motives.
KEY FACT
The report documented specific contradictory listings and reproducible chatbot behavior, but it did not estimate the platform-wide rate of false origin claims.
What the investigators found
The 46-page report from Columbia Law School’s Center for Law and the Economy, published July 30, 2026, examined Amazon’s Alexa for Shopping and Walmart’s Sparky. Researchers asked the assistants to find American-made products, evaluate suspicious claims and compare equivalent queries about products from other countries.
They also inspected listings for obvious contradictions. Examples included products using “Made in USA” in a title while another field said “imported,” or language saying an item was made domestically from imported material. The appendix lists four products of concern, three on Amazon and one on Walmart, rather than a random sample of either marketplace.
The team says small changes to query wording could sometimes bypass an Amazon refusal to answer “Made in USA” requests. Walmart’s assistant could rank products by how suspicious their claims appeared. These observations support the narrow conclusion that the assistants can sometimes process origin evidence already available in a listing.
Why an AI capability is not the same as a consumer protection
A chatbot noticing a contradiction does not mean the marketplace has a verified enforcement system. Reliable protection would require consistent rules, access to trustworthy supplier records, escalation to human review, correction of listings and a visible warning before a purchase.
This distinction is the most important part of the story. NewTqnia’s reading is that the report reveals an interface choice as much as an AI failure: platforms decide which model outputs become shopping features, which remain hidden and which trigger action. A capable model has little value to a buyer if its warning never reaches the product page.
The legal standard is also stricter than casual marketing language. Under the Federal Trade Commission’s rule, an unqualified “Made in USA” claim generally requires a product to be “all or virtually all” made domestically. In July 2025, the FTC sent letters to Amazon and Walmart identifying third-party listings that might violate those requirements and asking the platforms to monitor and correct misleading claims.
What Amazon and Walmart said
Amazon told Reuters that it displays country-of-origin information on product pages when the information is available, acts against sellers who violate its policies and is working to improve how Alexa surfaces the data. Walmart did not provide a comment before publication of the Reuters report on the investigation.
The Columbia authors argue that the tools’ own responses described commercial reasons for inaction. Those answers are notable, but they must not be treated as testimony from executives. A generative model can produce plausible explanations from patterns in its data without having access to an internal decision, policy meeting or business incentive.
Before we overstate the result
- The report is an advocacy-oriented investigation from a new policy center, not a peer-reviewed or independently replicated study.
- It provides selected examples and a single-category filter snapshot, not a representative sample that measures how common false claims are across either marketplace.
- Country-of-origin fields can be incomplete or supplied by third-party sellers. A contradiction is a reason for review, not final proof of fraud.
- Chatbot statements about why a platform behaves a certain way cannot establish corporate knowledge, intent or internal policy.
- AI assistants change frequently, so the same prompts may produce different results after model or interface updates.
What shoppers and regulators should watch next
For shoppers, the practical lesson is to check the product’s origin field, manufacturer information and supporting documentation rather than relying on a title or a chatbot summary. The independent Quartz account notes that the assistants found conflicts inside the platforms’ own information, which makes clearer presentation a realistic product choice rather than a distant research challenge.
The next useful evidence would be a large, independently reproducible audit across product categories, followed by data showing what happens after a suspicious listing is flagged. Regulators will also need to decide whether a marketplace that can algorithmically identify a warning has a responsibility to show it or investigate it.
AI shopping will not become trustworthy merely because the assistant sounds informed. The meaningful test is whether the system connects its analysis to visible evidence, correction and accountability before money changes hands.
Sources and citations4 sources
External references used to support the reporting in this article.
Published by
NewTqnia Technology Policy Desk
An institutional editorial team within NewTqnia
