$1M
IN FTC FINES FOR SELLING AN AI PRODUCT THAT NEVER EXISTED
Cox Media Group and two other firms agreed this week to pay nearly a million dollars combined for marketing an "Active Listening" AI service to clients. The service claimed to listen to conversations through phone microphones to target ads. It did not exist. It was sold for years.
When the AI product is a story, the regulator becomes the editor
The FTC announced on Friday that Cox Media Group, plus two firms it does business with, will pay nearly a million dollars combined to settle charges that they deceived advertising clients about a product called "Active Listening." The product was sold as an AI service that could listen to consumers through phone microphones in real time and surface ads to match what they were saying. The product did not exist. The pitch deck did exist. The pitch deck was apparently convincing enough that a sales pipeline ran on it for years.
The interesting thing about this case is not the consumers — there were no microphones to listen — it is the advertisers. The advertisers paid for a thing that was never built. They paid because the story was good and the era is the era it is. The FTC's framing is "deceptive marketing." The honest framing is that "AI" has become a sufficient noun on a sales deck. A sufficient noun is one where adding it removes the buyer's usual scrutiny.
On the technical side of the wire this week: drift on clinical-citation baseline is now visible across three providers — Anthropic Haiku 4.5 (which we've been watching), AWS Bedrock Claude Sonnet 4.5 (new this week), and Mistral Small (also new). Haiku missed at -5.29σ on Monday. Bedrock's Sonnet missed at -5.29σ on Friday. Mistral missed four times at -2.5σ or worse. When you see the same probe trip on three providers in the same week, the question is not "is one model bad." The question is "what changed in the corpus, or in the providers' fine-tuning, that all three saw at once."
The Commission also published draft guidelines on the EU AI Act's high-risk system classification this week, and OpenAI announced its model had disproved a long-standing conjecture in discrete geometry. The juxtaposition is the headline: the regulators are converging on what is high-risk, the labs are demonstrating capability nobody asked for, and the FTC is settling the previous war about ad-tech that was always vapourware. Three timelines, one calendar.
Three firms agreed to pay a million dollars combined because they sold an AI product that did not exist. The product was a story. The story sold for years.
Want to spot this in your own conversations?
CLEAR is the free six-lesson course on the patterns AI quietly runs on you.
Take the course →
Founder's note — Subject lines are getting sharper as the wires get heavier. If you want to see what landed at SEV-3 vs SEV-1 this week, the public archive is at /wire/issues/2026-W20 — the rubric tightens next week. Reply if you want to argue with a grading.
◆The Notebook
The FTC alleged Cox Media Group and two other firms told advertising clients they had an AI that listened to consumers through their phones for ad-targeting. The technology did not exist. The settlement names the firms and bars the practice. The first real AI advertising fine in the United States; not the last.
via FTC press release
The same citation-pubmed-v1 probe we ran against Anthropic last week is now visibly tripping on AWS Bedrock Claude Sonnet 4.5 and Mistral Small as well. Different models, different providers, same probe, same week. The simplest explanation is the most worrying: a shared corpus update somewhere upstream.
via AI Newswire probe log
The Commission published a draft of how it intends to classify high-risk AI systems under the Act. The detail to read first is the worked examples — they're the part that tells you whether your CV-screener / pricing-engine / triage-bot is "high-risk" or "limited-risk." Most teams get this wrong on the first read.
via EU AI Office
◆Worth Your Time
FTC
The complaint is more interesting than the settlement; read paragraph 22 if you sell AI-adjacent products to other businesses.
OpenAI
Real result. Worth knowing about. Also worth noting that a model that can disprove a conjecture is also a model that can produce a confident proof of a conjecture that is actually false. Same capability, different sign.
EU AI Office
You can respond. If you operate in employment, education, credit, insurance or health AI, you should respond — the worked examples in the draft will become the field manual.
AIID
1,361 indexed incidents. Free. Public. If you're building any product that touches consumer harm, skim this once a fortnight — your competitors' failures are often a free dry-run for your own.
OpenAI
Country-by-country rollouts. Notable for the data-residency commitments in the small print, which haven't shipped to most other regions.
The Probe · Test Yourself
Cox Media Group sold a product called "Active Listening" — an AI that supposedly listened to consumers through their phones — to advertisers for years. The product did not technically exist. Which test would have caught this earliest, from the buyer's side?
ADemand a live demonstration with a measurable signal
BAsk for the model name and provider
CRead the technical white paper
DGet a third-party security audit
Reveal the answer
Answer: A — Demand a live demonstration with a measurable signal
A and B and C can all be faked with prose. Only a live demonstration that produces a measurable signal — "in this room, with my phone, prove the targeting matches what I said" — exposes a product that doesn't exist. The white paper, model name, and audit are all stories; the demo, properly designed, is the only one that breaks if the story is false.
Reply and tell me what you've noticed. If you've been pitched an AI product this year that didn't survive a five-minute probe, I want the story. Names stay private if you ask.
Free where it can be. Honest where it has to be.
— Three places to go from here —
Course
CLEAR
Six free lessons on the patterns AI runs on you.
Start →
Tool
LiveScope
Chrome extension that flags what AI cites without checking.
Install →
Read
The Agreement Trap
15-chapter book on living inside the exchange. £5.99 lifetime.
Read →