Here's what's on our plate today:

• 🧪 Amazon's assistant won't answer made-in-USA questions, though it holds the information.

• 📰 Meta drops the camera; Muse shakes markets; data centers heat neighborhoods.

• 🛠 Weekend To-Do: read the Columbia report, check the FTC letters, read Ferguson's dissent.

Let’s dive in. No floaties needed.

The Laboratory

TL;DR

• The next fight over AI rules is about what regulators can see, not what they can forbid.

• The letter: On September 17, 2026, two senators asked the FTC to investigate whether Amazon's and Walmart's shopping assistants bury American-made products and miss fraudulent labels.

• The test: Columbia researchers found that the Amazon assistant refused to answer a question about American-made fishing reels, then answered the same question once the phrasing changed.

• The gap: A model can hold information, but the system built around it is not permitted to use it, a point that rules aimed at model builders never reach.

• The law: The FTC's Made in USA rule has carried penalties of $53,088 per day since 2021, and this year's enforcement actions targeted sellers rather than either platform.

• The objection: Ferguson, who would have to authorize an investigation, holds that existing law already covers deception by AI and that no separate regime is needed.

• The stakes: Governing how AI is used comes down to whether the instructions, filters, and model versions behind an answer can be used as evidence.

Why Amazon's AI struggles with made in USA labels

In the late 1950s, American record companies worked out that the cheapest way to sell a song was to pay the man who played it. Disc jockeys took cash, gifts, and songwriting credits from labels, then put the records on air as though the choice were their own. Listeners heard a recommendation, but what reached them was an arrangement, and that arrangement was almost never written down. Congressional hearings at the end of the decade gave the practice a name, payola, and ran into the difficulty that misleading an audience was already unlawful while proving any particular deal was close to impossible. The fix Congress settled on in 1960 was a disclosure duty rather than a fresh prohibition, requiring stations to announce that a play had been paid for, which turned a private instruction into a record somebody could check.

An AI shopping assistant stands where that disc jockey stood, answering in the voice of a guide while carrying instructions from the company that owns the shelf. A model can hold a piece of information perfectly well even if the system built around it is not permitted to use it. Capability belongs to the model, and permission belongs to the deployment, which means regulating the company that trains a model leaves untouched what that model may do once a retailer wires it into a catalog. On September 17, 2026, Senators Tammy Baldwin of Wisconsin and Rick Scott of Florida asked the Federal Trade Commission to investigate whether Amazon's Alexa for Shopping and Walmart's Sparky bury American-made products and fail to flag fraudulently labeled ones. The request went to Chairman Andrew Ferguson and the agency's only other sitting commissioner, and the FTC declined to comment while confirming it had received the letter.

The refusal disappeared when the phrase changed

Both assistants run on a large language model, a system trained to predict the next stretch of text, which answers fluently whether or not it holds the underlying facts. A retailer does not hand that model a blank page. It writes standing instructions that the model reads before every conversation and the shopper never sees, decides which parts of its catalog the model may consult, and adds rules that block certain replies. A blocked reply arrives as a refusal, and from outside, a refusal caused by missing data looks identical to one caused by a rule.

The Columbia Law School report, published in July 2026 and written by Erie Meyer and Zachary Harris, separated the two by rephrasing blocked requests. When asked for American-made fly-fishing reels, the Amazon assistant refused, saying it had no access to the information, while the same assistant answered a question about Chinese-made reels with pricing and sourcing analysis. Asked again about reels made “in USA”, it returned a full answer, which established that the information was present and something had been intercepting the phrase. The refusal recorded a rule about what the deployed system could say, and told the shopper nothing about the limits of what the model knew.

Why that rule exists is harder to establish, since the evidence for it runs through the assistants' own answers. Both asked why fraudulent labels go unflagged, described the choice as commercial, and Sparky called it "a business calculation, not a legal justification.” Those are model outputs rather than admissions, since a model asked why a company does something will construct a plausible account whether or not a real one is available to it. Amazon rejects the framing, says it shows origin information on product pages where available, and told reporters that “The suggestion that we intentionally withhold country-of-origin information from customers is incorrect.”

An FTC rule has governed these claims since 2021

The Made in USA Labeling Rule carries civil penalties of $53,088 per violation per day, and on July 8, 2025, the agency wrote to Amazon and Walmart naming third-party sellers making apparently deceptive origin claims and asking both platforms to monitor and correct them. Enforcement 2026 has since gone elsewhere, since a sweep announced in April 2026 produced three actions against sellers rather than platforms, while neither Amazon nor Walmart has been charged.

Amazon told investors it expects Rufus, its shopping assistant, to generate more than $10B in annual incremental sales, and Walmart has begun testing advertising formats inside Sparky, which merges the commercial and informational jobs of a single output. Advertising rules attach a label to a sponsored listing, while a sentence of advice reaches the shopper unmarked.

Ferguson has argued that the existing law is enough

Any investigation would have to be authorized by Ferguson, who has spent the past two years arguing against building a separate regime for AI. Writing in his dissent in the Rytr matter, he said the Commission should punish people who use AI to deceive no differently than if they had used “quill and parchment”, and that Congress has never granted the agency power to regulate AI as such. He also holds that liability for supplying the tools to deceive requires proof the supplier knew or had reason to know, and the Commission acted on that view when it set aside the Rytr order in December 2025.

On that reading, the senators' letter describes conduct that is already unlawful if proven, and the remedy for it is an investigation rather than a statute. It says nothing, though, about how an enforcer would establish what the assistant had been told to do. The Columbia researchers ended up interrogating a chatbot about its owner's commercial reasoning, a method the authors defend as corroboration rather than proof, used because no other window into the system was open to them. State law is moving in the same direction, since Colorado replaced its 2024 AI Act in May 2026 with a narrower disclosure regime that drops the duty of care and the mandatory impact assessments.

The 1960 disclosure rule worked because the instruction and the man carrying it out were the same person, and announcing the payment left a record that outlived the broadcast. An assistant's instruction sits in a file the shopper cannot open, the regulator cannot demand without first opening a case, and the company can rewrite between one query and the next. Governing how AI is used, rather than who builds it, comes down to whether that file survives as evidence, along with the retrieval scope, the filter rules, the ranking criteria, the model version in service, and the log of what changed after launch. The FTC has no deadline to decide whether to look, which leaves open what would still be found.

Weekend To-Do

• Read "Made in America, Hidden by AI": Go straight to the transcripts, where the same question about fishing reels gets refused once and answered once.

• Check the FTC's July 2025 warning letters: See which sellers the agency named on Amazon and Walmart, then note that April's sweep charged sellers only.

• Read Ferguson's Rytr dissent: It is the clearest statement of why the chairman thinks quill-and-parchment law already covers everything in the senators' letter.

Friday Poll

Headlines You Actually Need

• Meta's camera-free glasses: Meta is expected to unveil smart glasses without cameras at Connect, following complaints about covert recording, UK venue bans, and a proposed California class action.

• Muse splits the market: Meta's Muse agent logged 2.8M downloads in two weeks and dragged banks, insurers, and booking stocks down as investors priced in comparison shopping.

• Data centers are heating neighborhoods: A single AI facility can consume over 50 megawatts and vent nearly all of it as waste heat, with plumes carrying it over 300 feet.

Meme Of The Day

The Toolkit

• Langfuse: Open-source LLM observability that traces every prompt, retrieval, and response, so you can audit what actually happened.

• PromptLayer: Prompt management workbench that version-controls templates and logs every request, so each change to an instruction leaves a record.

• Browse AI: No-code robot that monitors any web page and alerts you the moment listings, prices, or product claims change.

Rate This Edition