Buyers of “AI detectors” often need a blunt answer: was this essay, blog, or marketing page written by a model? On the FTC’s account, Workado sold that answer with a number that belonged to a different classroom.
Our Reported write-up on the Workado order holds the full consent-order record. This post pulls the control lesson: match the evaluation set to the claimed use case before you put a percentage on the tin.
What the record says
On 28 August 2025 the U.S. Federal Trade Commission gave final approval to a consent order against Workado, LLC (formerly Content at Scale AI). The Scottsdale, Arizona company must stop advertising the accuracy or efficacy of its AI content-detection products unless it has competent and reliable evidence that those products are as accurate as claimed. After public comment, the Commission voted 3–0 to approve the final order and respond to two comments. Workado neither admits nor denies the complaint’s allegations except as needed for jurisdiction.
The August finalisation release skips the arithmetic. The April 2025 complaint and Decision and Order supply it. The complaint alleges Workado claimed the detector would predict with 98.3% accuracy whether text came from tools such as ChatGPT, GPT-4, Claude, or Bard, and promoted a 98% rate in ads and on-site copy. It further alleges Workado did not build, train, or fine-tune the model: the detector used a public Hugging Face model, RoBERTa-academic-detector, trained by independent developers on scholarly research abstracts — not on Wikipedia entries or blog posts, despite marketing that described a wide range of everyday material.
Workado, the complaint says, did not test whether that accuracy held for the marketing and other plain-language text users typically submitted. It relied on the developers’ published academic-content results. Those same published results, per the complaint, showed a best accuracy of 74.5% on mixed human and AI non-academic content, and correct detection of AI-generated non-academic text only 53.2% of the time. Attachment A to the order (the customer-notice email) restates the 98% claim and that the FTC said the company lacked proof.
The final order bars efficacy claims without backing evidence, requires retaining that evidence, requires email notice to eligible customers, and requires compliance reports one year after issuance and yearly for three years more. Case materials live on the FTC’s Workado / Content at Scale AI page. The April announcement first floated the proposed order.
This package stays on Workado only — separate from the Gemini × Irregular CTF blog and the OpenAI DSEWiki agent-spam blog.
Key learnings
Accuracy is a claim about a distribution. A score measured on scholarly abstracts does not transfer to blogs and marketing copy just because the product page says “content.” Expert-it-isnt: the system was an academic classifier while the pitch sold a general detector.
Borrowed models still need your eval. Using a public Hugging Face checkpoint does not license reusing the authors’ academic metrics for a different domain. If you did not test the customer’s text type, you do not have their accuracy number.
Marketing copy is part of the product. Describing training on Wikipedia and blogs while shipping an abstracts-trained model is, on the FTC’s theory, the deception — not a footnote.
“Neither admits nor denies” still changes what you can say next. The order’s evidence and notice duties are the operational takeaway for anyone shipping detectors, regardless of how the caption reads.
What to try
- Define the claim domain before the metric. Write one sentence: “This % is for [genre] submitted by [user type].” If ads say something broader, either widen the test set or narrow the claim.
- Hold out a use-case eval set. Keep a frozen set of real customer-like texts (blogs, essays, marketing) with labels you trust. Report accuracy on that set, not only on the upstream paper’s abstracts.
- Publish the mismatch. If academic accuracy is 98-ish and non-academic is mid-70s / ~50 on AI text — as the complaint cites from the developers’ own published non-academic figures — show both, or do not advertise the higher number for the lower domain.
- Re-test when the generator mix shifts. GPT-4 / Claude / Bard-era text is not a permanent stand-in for next year’s models. Schedule re-runs when the claimed target tools change.
- Keep the evidence pack with the ad. The order’s “competent and reliable evidence at the time made” bar is a good internal rule even outside the FTC: store the eval script, split, and date next to the screenshot of the claim.
- Separate model card metrics from product metrics. Upstream Hugging Face cards describe the authors’ experiment. Your product page describes your users’ risk. Do not paste one into the other.
Closing
- Before you print a detector %, can you point to a test set that looks like the text your buyers will paste?
- If the best number you have is from scholarly abstracts, what are you willing to claim — and refuse to claim — about everyday writing?