The AI Detector That Performed “No Better Than a Coin Toss”

As generative AI tools like ChatGPT proliferated, so did a new market for software claiming to detect whether a piece of writing was produced by a human or a machine. Workado, LLC, formerly known as Content at Scale AI, marketed its “AI Content Detector” with a specific, confident number: 98.3 percent accuracy. According to the Federal Trade Commission, that number bore little resemblance to how the tool actually performed once independently tested.DOCUMENTED

The FTC finalized a consent order against Workado in April 2025, prohibiting the company from making unsupported accuracy claims about any AI detection product going forward.DOCUMENTED

What the complaint alleges

According to the FTC's complaint, Workado advertised that its AI Content Detector was developed using a wide range of material — including blog posts and Wikipedia entries — specifically to make the tool more accurate for average, everyday users checking ordinary writing.DOCUMENTED The complaint alleges that in reality, the AI model powering the detector was trained or fine-tuned to effectively classify only academic content, meaning its advertised broad accuracy did not extend to the general, non-academic writing many everyday users were actually checking.DOCUMENTED

Christopher Mufarrige, Director of the FTC's Bureau of Consumer Protection, put the gap in blunt terms: “Consumers trusted Workado's AI Content Detector to help them decipher whether AI was behind a piece of writing, but the product did no better than a coin toss,” he said, adding that “misleading claims about AI undermine competition by making it harder for legitimate providers of AI-related products to reach consumers.”DOCUMENTED

The gap between the advertised number and the tested result

The complaint cites independent testing data showing the AI model's best accuracy result on a mix of human- and AI-generated text reached only 74.5 percent — well below the advertised 98.3 percent — and dropped to just 53.2 percent specifically for non-academic content, a result barely better than random chance for a binary human-or-AI classification.DOCUMENTED That gap between an advertised near-perfect accuracy figure and a real-world result close to a coin flip is precisely what elevated the case from an ordinary overstated marketing claim to a Section 5 deception violation.

Why AI accuracy claims get the same scrutiny as any other advertising

The FTC's action against Workado reflects a broader principle the agency has applied consistently to emerging AI products: a claim about what an AI tool can do is subject to the same substantiation requirements as any other advertising claim, regardless of how novel or technically complex the underlying technology is.REVIEWED A company advertising 98.3 percent accuracy needs competent and reliable evidence supporting that specific figure across the range of uses the marketing implies, not just for a narrower subset — academic writing — that happened to be what the model was actually built to handle.

Terms of the settlement

Under the final order, Workado is prohibited from making any representation about the effectiveness of any AI content detection product unless the claim is not misleading and the company possesses competent and reliable evidence to support it at the time the claim is made.DOCUMENTED The company must retain that supporting evidence, email eligible past customers about the settlement, and comply with recordkeeping and compliance reporting obligations for up to 20 years.DOCUMENTED

The tool's real-world accuracy on non-academic writing was 53.2 percent — barely better than flipping a coin — against an advertised claim of 98.3 percent.

Why the case matters

For anyone relying on an AI detection tool — teachers checking student work, employers screening writing samples, or consumers simply curious whether a piece of text was AI-generated — the Workado case is a reminder that accuracy claims for this fast-growing category of software deserve the same skepticism as any other advertised statistic, and that a tool trained and tested primarily on one narrow category of writing may perform dramatically worse outside that category, regardless of the headline number in its marketing.

Why this case may shape the broader AI detection market

As schools, employers, and publishers increasingly rely on AI detection tools to screen submitted writing, the accuracy of those tools carries consequences well beyond a single dissatisfied customer — a student wrongly flagged for AI-generated work, or an employer wrongly clearing a candidate's plagiarized application, both depend on a detector performing close to its advertised accuracy. The Workado settlement puts other AI detection vendors on notice that a specific, quantified accuracy claim will be tested against real-world performance across the full range of content the marketing implies it covers.

What buyers of AI detection tools can ask before purchasing

Prospective business or institutional buyers of any AI content detection product can reasonably request the specific testing methodology and dataset behind an advertised accuracy figure, and can ask whether that testing covered the same type of content the buyer actually intends to screen. A vendor unable to answer that question in specific terms is asking a buyer to trust exactly the kind of unverified claim the FTC found unsupported in Workado's marketing. Until that kind of transparency becomes standard across the industry, treating any single accuracy figure as provisional, rather than definitive, remains the safer default for institutional buyers. The burden of proof, as this case shows, belongs with the vendor making the claim, not the consumer relying on it. That reversal of the usual burden protects the many institutions now relying on these tools to make real decisions about real people's work.

Have documents relevant to this story? Reach us through our tips channel.