An "AI-powered" claim used to be a matter of taste — you either believed the pitch or you didn't. It isn't anymore. The Federal Trade Commission and the Securities and Exchange Commission have between them settled more than a dozen cases since March 2024 against companies that advertised AI capability they didn't actually have, and each settlement lays out, in public, exactly what question would have caught the fake before anyone paid for it. Those same questions work on the next pitch, whether or not a regulator ever gets to it.
Why this became a checkable claim, not an opinion
The SEC went first, on March 18, 2024, against two registered investment advisers. Delphia had told clients since 2019 that it used "machine learning" to analyze "the collective data shared by its members" and was "training our algorithm for pursuing ever better returns" — the firm had privately admitted, back in 2021, that it had never actually used any client data in its algorithm, and kept the public claim running two more years anyway. Global Predictions marketed itself as the "first regulated AI financial advisor," offering "expert AI-driven forecasts" nothing backed up. Delphia paid $225,000, Global Predictions paid $175,000 — $400,000 combined — and neither firm had to admit wrongdoing, but both had to stop the specific claims.
The FTC opened a wider front six months later: Operation AI Comply, launched September 25, 2024, against five companies at once — an "AI robot lawyer" never tested against a real attorney, an e-commerce course promising AI-powered passive income, a writing tool marketed for generating fake reviews. The sweep has kept running since, with at least 13 settled cases and roughly $51 million recovered as of this piece, spanning consumer software, marketing services, and business-coaching programs.
One of those 13 is worth walking through end to end, because it shows exactly how a fake number survives contact with a real one. Workado sold an AI Content Detector, marketed as "98% accurate" at flagging AI-written text. The 98% wasn't invented — it came from an independent Norwegian study of the underlying model, tested on academic writing. Workado just never ran that test on the general, non-academic content its own customers actually fed it; when the FTC did, the real number came back at 53%. "Consumers trusted Workado's AI Content Detector to help them decipher whether AI was behind a piece of writing," said Chris Mufarridge, director of the FTC's Bureau of Consumer Protection, "but the product did no better than a coin toss.(The FTC's own language from its order against Workado — not this desk's characterization.)"
Run this before repeating an "AI-powered" claim
Every one of these cases collapses under the same short list of questions — asked before the purchase, the investment, or the headline, not after.
Four questions, in the order that catches a dressed-up claim
- Cox Media Group's "Active Listening" product claimed to analyze conversations picked up by a phone's microphone and use them to target local ads. The FTC found in May 2026 that the service never touched voice data at all — it resold email lists bought from data brokers, marked up, with an AI label on top. The tell wasn't a bad output; it was that nobody could name the specific step where a model did anything.
- Workado's "98% accurate" figure was real, just measured on the wrong content. A real number attached to the wrong use case reads exactly like a fabricated one from the outside — the only way to tell them apart is asking what it was actually measured on.
- A model's exact weights or training recipe can be a real trade secret. What it decides, what kind of data trains it, and how a claimed number was measured are not the same kind of secret — Global Predictions called itself the "first regulated AI financial advisor" and never had to show its work, because nobody separated the ownable question from the proprietary one.
- Search the company's name alongside "FTC" or "SEC" before trusting a pitch. The same handful of claim shapes — an inflated accuracy number, an untested "proprietary algorithm," a feature that doesn't do what the name implies — keep reappearing across unrelated companies once you know to look.
The clearest example of question one — what does the AI actually decide — closed in May 2026. Cox Media Group and two smaller partners, MindSift and 1010 Digital Works, had sold "Active Listening": a service claiming to analyze conversations picked up by a phone's microphone in real time and use them to target local ads, with customers told they'd opted in. On May 21, 2026, the FTC found the service never touched voice data at all — it resold email lists bought from data brokers, marked up, with an AI label on top. Cox Media Group paid $880,000; MindSift and 1010 Digital Works paid $25,000 each — $930,000 combined.
What was claimed, and what a regulator actually found
- Cox Media Group's "Active Listening" ad service (FTC, May 2026)
- Workado's AI Content Detector (FTC, 2025)
- Delphia's investment algorithm (SEC, March 2024)
- Global Predictions' robo-adviser (SEC, March 2024)
None of these four needed a whistleblower or a leaked memo. A customer or a regulator asked what the product actually did, on what data, and checked the number against reality. That's the whole method.
What the enforcement numbers actually cover
What the enforcement numbers actually cover
- $51 million · FTC · Operation AI Comply
- Recovered across at least 13 settled cases since the Sept. 25, 2024 launch
Includes: Civil penalties and consumer redress from that named sweep of deceptive-AI-marketing cases
Excludes: Any FTC AI action outside the named sweep, and any state attorney-general enforcement - $400,000 · SEC · Delphia + Global Predictions
- Combined penalty from the SEC's first AI-washing cases, March 18, 2024
Includes: Two investment-adviser marketing-rule violations
Excludes: Later SEC AI matters, including Presto Automation's 2025 case against a public company under a different statute - $930,000 · FTC · Cox Media Group + partners
- Combined penalty for the "Active Listening" case, May 21, 2026
Includes: The three companies named in that specific settlement
Excludes: Any customers' own losses from the underlying ad-targeting service, which the settlement doesn't separately quantify
None of these totals should read as the size of the problem — they're the size of what's been caught. The FTC has run roughly a dozen cases a year against a market of many thousands of products carrying an "AI-powered" label; the absence of a settlement next to a company's name is not the same as a passed check. Run the four questions regardless of what the search above turns up.
Where this check gets skipped
Four ways this check gets skipped when it shouldn't be
Run the same check on a benchmark chart and you're doing what this desk's guide to verifying an AI benchmark claim teaches; run it on a funding number and you're doing the valuation read. All three are the same instinct pointed at a different kind of claim: not distrust by default, just the question a regulator would already be asking. The dictionary has short entries for a few of the terms that come up along the way, if any need unpacking.
- "AI washing" — claiming AI capability a product doesn't actually have — is now a finable violation.
- The FTC and SEC have brought at least 13 cases since 2024, recovering roughly $51 million total.
- Ask what specifically the AI decides, on what data, and who tested the claimed number.
- Cox Media Group's "AI" ad targeter turned out to be resold email lists with no AI at all.
- Caveat: no open case doesn't clear a company — regulators have reached only a slice of the market.