Unverified. We have not been able to confirm the details of this account.
Reported The Expert It Isn't Cringe
FTC finalises Workado order over AI detector accuracy claims built on an academic model
On 28 August 2025 the FTC finalised an order against Workado (formerly Content at Scale AI) over alleged misrepresentation of its AI Content Detector’s accuracy; the complaint ties 98%/98.3% claims to an academic-only model.
“AI detector” accuracy percentages sold without published test sets matching the claimed use case.
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), requiring the Scottsdale, Arizona company to stop advertising the accuracy or efficacy of its artificial-intelligence content detection products unless it has competent and reliable evidence that those products are as accurate as claimed. Following a public comment period, the Commission voted 3–0 to approve the final order and to respond to two comments. The April 2025 complaint and proposed order had already set out the core allegation: Workado marketed an AI Content Detector to people trying to tell generative-AI text from human writing, while claiming the tool had been developed on a wide range of material including blog posts and Wikipedia entries.
The detail the August finalisation release omits — and the complaint and Decision and Order supply — is the accuracy arithmetic and the model underneath. The complaint alleges Workado claimed the detector would predict with 98.3% accuracy whether text was generated by tools such as ChatGPT, GPT-4, Claude, or Bard, and promoted a 98% accuracy rate in ads and on-site copy. It further alleges Workado did not build, train, or fine-tune the model: the detector used a publicly available Hugging Face model named RoBERTa-academic-detector, trained by its independent developers on scholarly research abstracts (human-written and ChatGPT-generated), not on Wikipedia entries or blog posts. Workado, the complaint says, did not test whether the same accuracy held for the marketing and other plain-language text its users typically submitted, and instead relied on the developers’ published academic-content results. Those same published results, the complaint states, showed a best accuracy of 74.5% on a mix of human-created and AI-generated non-academic content, and correct detection of AI-generated non-academic text only 53.2% of the time — figures the complaint uses to call the 98% claim unsubstantiated and deceptive. Attachment A to the Decision and Order (the required customer-notice email) restates that Workado had claimed a 98% accuracy rate and that the FTC said the company lacked proof.
The final order prohibits efficacy representations about AI content detection products unless they are not misleading and are backed by competent and reliable evidence at the time made; requires retention of supporting evidence; requires email notice to eligible customers; and requires compliance reports one year after issuance and yearly for the following three years. The Decision and Order states that Workado neither admits nor denies the complaint’s allegations except as specifically stated, admitting facts only as needed for jurisdiction.
The pattern is expert-it-isnt. A detector was sold as a general accuracy authority for everyday writing. On the FTC’s account, the working model was an academic-text classifier doing what it was trained to do — score research abstracts — while the marketing described a different use case.
What to sit with
Worth arguing about
- Should every accuracy ad name the test set and the kind of writing in the same breath as the percentage?
- When is rebadging a public model smart product work — and when is it a mismatch dressed as proof?
What’s at stake
- Workado sold an AI detector with eye-catching 98% accuracy lines aimed at everyday writing, not lab papers.
- The FTC’s complaint says the model underneath was trained on academic abstracts — and published scores on ordinary text were much lower.
What this gets right
- The final order pushes customer notices and forces future accuracy ads to rest on real evidence at the time.
- Buyers get a paper trail instead of a percentage with no test set attached.
What stays messy
- People who needed a “general” detector may have trusted a number built for a different kind of writing.
- Neither-admit-nor-deny framing still leaves the mismatch story as allegation on the liability form.
Easy to miss
- The cheerful finalisation press release skips the percentages — you need the complaint and order for the real numbers.
One thing to try What if you refused any detector percentage that did not ship with the test set for your kind of text?