FTC final order: the '98% accurate' AI Content Detector was an unmodified student model, and 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 FTC found the stakes of that verdict are real, warning that “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” (source). Workado, LLC, formerly Content At Scale AI, sold exactly that certainty: its AI Content Detector was advertised as able to “predict with 98.3% accuracy whether text was generated using AI technology such as ChatGPT, GPT4, Claude, Bard, or another generative AI technology” (source).

What was built

The Federal Trade Commission’s administrative complaint (Matter No. 232-3092) sets out three layers, which Workado settled by consent without admitting or denying the allegations (source).

First, the advertised number. From at least November 2022 until on or about 21 June 2024, Workado’s website claimed the tool would “predict with 98.3% accuracy whether text was generated using AI technology,” and that it was trained on “a vast amount of material, including blog posts and Wikipedia entries” (source). The FTC’s April 2025 press release summarised the marketing as a “98 percent” accuracy claim (source).

Second, the actual model. Per the complaint, Workado “did not create, train, or fine tune the AI model used in its AI Content Detector,” which is a publicly available open-source model its developers named the RoBERTa-academic-detector (source). Those developers “were students in Norway who developed it for an undergraduate thesis to detect machine-generated academic text,” and trained it on abstracts of scholarly articles rather than the blog posts and Wikipedia entries Workado’s marketing described (source). The model’s own Hugging Face card confirms the same origin, a bachelor’s thesis at the Norwegian University of Science and Technology reporting an accuracy of 98.4% measured only on research abstracts (source).

Third, the real-world number. The complaint states 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 that on AI-generated non-academic text the model correctly detected it “merely 53.2% of the time” (source). The complaint’s conclusion was that for Workado’s real users the detector “would do barely better than a coin toss” (source). FTC Bureau of Consumer Protection Director Chris Mufarrige used the same phrase in the FTC’s 28 April 2025 press release: “the product did no better than a coin toss” (source).

The outcome

What was claimed: 98.3% accuracy. Workado advertised that figure on its website from at least November 2022 until on or about 21 June 2024, and from September 2023 charged $49 per month for unlimited use (source).

What the record showed: 53.2% on the text it was sold to catch. The model’s best accuracy on mixed non-academic content was 74.5%, and its detection of AI-generated non-academic text was 53.2%, which the FTC called “barely better than a coin toss” (source).

The sanction: no money, binding conduct rules. Under final Decision and Order C-4822, issued 28 August 2025 after a 3-0 Commission vote, Workado may not make effectiveness claims about any AI content-detection product, for text or images, unless the claim is non-misleading and backed at the time by competent and reliable evidence (source). The order also requires Workado to preserve its testing data, email eligible customers notice of the settlement, and file sworn compliance reports one year after the order and for the following three years (source). There is no monetary relief, but each future violation of the order carries a civil penalty of up to $53,088 (source).

How this was verified

Every critical figure originates with the independent regulator and was re-fetched live from ftc.gov on 2026-08-11. The 98.3% claim, the 74.5% and 53.2% real-world figures, the “coin toss” line, the $49 subscription, and the model’s provenance are all quoted from the FTC’s administrative complaint (Matter No. 232-3092) (source). The sanction terms, the 3-0 vote, and the $53,088 penalty are from the final Decision and Order (Docket No. C-4822) and the FTC’s 28 August 2025 press release (source). One date was corrected during this remediation: an earlier version of this page gave the final order’s date as 21 August 2025, but the FTC record and the announcing press release both date the approved final order to 28 August 2025, and the page now matches. The model’s identity and thesis origin are independently confirmed by its Hugging Face card, a developer-authored source (Tier 3) that predates the FTC action and is used only to corroborate provenance, not the FTC’s legal findings (source). Workado settled without admitting or denying the allegations, so every finding above is attributed to the FTC record (source).

Related case files

The same pattern, a regulator finding that an AI capability claim outran its evidence, runs through the FTC’s DoNotPay order, where the “world’s first robot lawyer” could not do what it advertised, and the FTC’s IntelliVision order, where unsubstantiated accuracy and zero-bias claims for facial recognition drew a 20-year consent order. The wider phenomenon of a company overstating what its AI can do is the subject of the SEC’s Delphia AI-washing penalties, a different regulator sanctioning the same gap between claim and proof.

Sources

Tier legend: Tier 1 primary (the regulator stating its own findings or a party’s own record); Tier 3 weak (developer-authored model card). Figures checked live on 2026-08-11.

1. FTC, “Complaint” (Matter No. 232-3092), In the Matter of Workado, LLC, f/k/a Content At Scale AI (Tier 1, primary regulatory pleading; source of the 98.3%, 74.5%, 53.2%, “coin toss”, model-provenance, and advertising-period figures). https://www.ftc.gov/system/files/ftc_gov/pdf/ContentatScaleAI-Complaint.pdf
2. FTC, “Decision and Order” (Docket No. C-4822), issued 28 August 2025 (Tier 1, primary regulatory order; source of the conduct remedies and competent-and- reliable-evidence standard). https://www.ftc.gov/system/files/ftc_gov/pdf/ContentatScaleAI-DecisionandOrder.pdf
3. FTC, “FTC Approves Final Order against Workado, LLC, Which Misrepresented the Accuracy of its Artificial Intelligence Content Detection Product,” 28 August 2025 (Tier 1, regulator’s own release; 3-0 final-approval vote, $53,088 penalty, compliance-report terms). https://www.ftc.gov/news-events/news/press-releases/2025/08/ftc-approves-final-order-against-workado-llc-which-misrepresented-accuracy-its-artificial
4. FTC, “FTC Order Requires Workado to Back Up Artificial Intelligence Detection Claims,” 28 April 2025 (Tier 1, regulator’s own release; Mufarrige “coin toss” quote, “98 percent” / “53 percent” summary). https://www.ftc.gov/news-events/news/press-releases/2025/04/ftc-order-requires-workado-back-artificial-intelligence-detection-claims
5. FTC, “Content at Scale AI” case page (Tier 1, regulator case record; docket and status). https://www.ftc.gov/legal-library/browse/cases-proceedings/2323092-content-scale-ai
6. Hugging Face, “andreas122001/roberta-academic-detector” model card (Tier 3, developer-authored; corroborates the model’s thesis origin and academic-abstract training data only, not the FTC’s findings). https://huggingface.co/andreas122001/roberta-academic-detector