FTC final order: the '98% accurate' AI Content Detector was an unmodified student model — its own published data showed 53.2% on real-world text

The problem

AI content detectors sell certainty: paste text, get a verdict on whether a machine wrote it. The stakes of that verdict are real — in the FTC’s own words, “The consequences of these tools’ assessments can be severe – such as a student wrongly accused of cheating or a journalist’s article rejected for publication.” Workado, LLC (formerly Content At Scale AI) sold exactly that certainty: its AI Content Detector, advertised at “98% accuracy,” free to try and $49/month to use without limits from September 2023 to August 2024.

On 21 August 2025 the Federal Trade Commission issued a final consent order (Docket No. C-4822) settling charges that the 98% claim was false, misleading, or unsubstantiated. This is a verified negative: an independent federal regulator, in a signed public complaint and order, documented the gap between an AI product’s advertised accuracy and the only test data that actually existed for it.

What the FTC’s complaint alleges

Three layers (Workado settled by consent, neither admitting nor denying the allegations):

1. The advertised number. From at least November 2022 until on or about 21 June 2024, Workado’s website claimed the tool “will predict with 98.3% accuracy whether text was generated using AI technology such as ChatGPT, GPT4, Claude, Bard, or another generative AI technology,” and that it was “Trained on blog posts, Wikipedia, essays, and more.”

2. The actual model. Per the complaint, Workado “did not create, train, or fine tune the AI model used in its AI Content Detector.” The model is the publicly available open-source RoBERTa-academic-detector — built by “students in Norway who developed it for an undergraduate thesis to detect machine-generated academic text.” It was trained on academic research abstracts (human vs ChatGPT-generated) only: “The developers did not fine tune the AI Model using any non-academic content, such as Wikipedia entries or blog posts” — and never on Bard, Claude, or GPT-4 output at all. The model’s own Hugging Face card — captured three months before the FTC acted — says the same thing: a bachelor’s-thesis project from the Norwegian University of Science and Technology, with its ~98% figure measured only in-domain, on research abstracts.

3. The real-world number. Workado “relied solely on the publicly available testing results for the AI Model’s evaluation of academic content” to claim 98.3%. But the same developers’ published testing showed “the developers’ best result for the AI Model’s accuracy when evaluating a mix of human-created and AI-generated non-academic content was 74.5%” — and on the AI-generated non-academic text the product was actually sold to catch, the model succeeded “merely 53.2% of the time.” The complaint’s summary: for Workado’s real users, the detector “would do barely better than a coin toss.” FTC Bureau of Consumer Protection Director Chris Mufarrige put it the same way in the FTC’s 28 April 2025 press release announcing the case: “the product did no better than a coin toss.”

The order

Under final order C-4822 (issued 21 August 2025 after a 3-0 Commission vote), Workado, for any product that detects or purports to detect AI-generated content (text or images), must not make effectiveness claims unless they are non-misleading and backed at the time by competent and reliable evidence; must preserve the underlying testing data; must email eligible customers notice of the settlement; and must file sworn compliance reports. There is no monetary sanction — but each future violation of the order carries a civil penalty of up to $53,088.

Why this matters for buyers

This is the cleanest documented case of an AI product whose headline accuracy number was real — just measured on a different distribution than the one the product was sold for. The developers’ 98.4% in-domain result on academic abstracts became “98% accurate” marketing for your blog copy, where the honest published number was 53.2%. When a vendor quotes one accuracy figure, the question this case teaches is: measured on what, by whom, and does that match what you’re buying it for? That is precisely the verification gap TIN exists to close.

Verification record

All critical figures originate with the independent regulator: the FTC’s administrative complaint (Matter No. 232-3092) and final Decision and Order (Docket No. C-4822), both fetched live from ftc.gov this session and byte-identical to their publication-week Wayback captures (SHA-1 digests match the Internet Archive’s CDX records exactly). The underlying model’s Hugging Face card, captured 2025-01-24 — before the FTC action — independently confirms the provenance narrative. Workado settled without admitting or denying the allegations; every finding above is attributed to the FTC record.