How to Actually Evaluate an AI Detector's Accuracy Claims

Quick answer

Every AI detector puts a headline accuracy number on its homepage. Almost none of those numbers mean what they look like they mean. A new University of Chicago study gives the field its first rigorous framework for comparing detectors honestly. Instead of one accuracy percentage, ask what a tool's false-positive rate is at the decision threshold you'd actually use — then decide how much risk of a false accusation you're willing to accept. Separately, the Federal Trade Commission just forced one detector vendor to stop advertising a "98% accurate" claim built on a completely different type of text than the one its customers were actually checking. Together, these two developments spell out exactly what to ask before trusting any detector's marketing page, including ours.

The problem: "accuracy" is not one number

A detector makes two kinds of mistakes. It can call real human writing "AI" (a false positive), and it can call real AI writing "human" (a false negative). A single "accuracy" percentage on a vendor's homepage blends both mistakes into one number, which hides the tradeoff that actually matters to you. A detector tuned to catch almost every AI-written text will also flag more real human writing by accident. A detector tuned to almost never falsely accuse someone will also let more AI-written text slip through. Neither setting is wrong — they're different policy choices, and a marketing page that reports only one blended percentage doesn't tell you which choice the vendor made.

What the Chicago Booth research actually measured

Brian Jabarian and Alex Imas, economists at the University of Chicago's Becker Friedman Institute, built a dataset of roughly 2,000 human-written passages spanning six formats: blog posts, consumer reviews, news articles, novels, restaurant reviews, and résumés. They generated AI versions of that same content with four popular language models, then tested three commercial detectors (Pangram, GPTZero, Originality.ai) and one open-source model (RoBERTa) across four length categories — from full-length résumés down to passages under 50 words.

The results varied sharply by tool. All three commercial detectors kept false-positive rates under about 1% on medium and long text. The open-source RoBERTa model performed close to random guessing in several tests — the researchers concluded it's "unsuitable for high-stakes applications." Among the commercial tools, false-negative rates (AI text missed as human) ranged from roughly 0-2% for one tool up to 10-40% for another, depending on which language model generated the text. Short passages under 50 words were the hardest case for every tool.

The paper's real contribution isn't a leaderboard, though. It's a proposed standard the authors call a "policy cap." An institution first decides its own tolerance for false accusations — say, no more than 1 in 200 pieces of real human writing flagged as AI — and then asks which detectors can hit an acceptable catch rate without exceeding that cap. That turns the question from "which detector is most accurate" into "which detectors are honest about their tradeoffs, and which one fits the risk my institution can actually live with."

When a marketing number gets tested — and fails

The Chicago Booth paper is an academic evaluation, but a separate, unrelated case shows what happens when a detector's own advertised number goes unchecked by anyone — and nothing about that number was fabricated, either. In 2025, the Federal Trade Commission ordered Workado, LLC — the company formerly known as Content at Scale AI — to stop advertising its AI Content Detector as "98 percent accurate." The FTC's complaint laid out exactly where that number came from: a study measuring the tool's performance on academic writing. Workado then used the same 98% figure in marketing aimed at general, non-academic text — reviews, emails, everyday writing — where the FTC found the tool's actual accuracy was closer to 53%, which the agency's Bureau of Consumer Protection director called "no better than a coin toss." The real number came from a real study; the problem was applying it to a use case it was never tested on, without telling anyone the difference mattered. That's the exact failure the Chicago Booth policy-cap framework is designed to catch — Jabarian and Imas measured detectors across six separate text genres precisely because performance on one genre says almost nothing about performance on another. Workado's settlement now requires yearly compliance reports to the FTC for four years running, a rare case where a detector's claims got tested by an outside party with subpoena power, not just a competing vendor with an incentive to disagree.

What to actually ask before trusting a detector's claim

If you're an educator, an HR team, or an editor deciding whether to lean on a detector's score, four questions matter more than the headline accuracy percentage on any vendor's site. Start with the false-positive rate at the threshold you'd actually use — not the vendor's best-case number, but the rate at whatever confidence level triggers a real consequence for a real person. Then check what text length and genre it was tested on: a detector benchmarked on 1,000-word essays can behave very differently on a 100-word cover letter or a block of source code, since STEM and code text carry documented, separate bias problems of their own. Ask whether it holds up against paraphrasing and "humanizer" tools — the Chicago Booth paper specifically tested robustness to humanizer software, and some tools' false-negative rates collapsed under paraphrasing attacks while others didn't. And ask whether the number was produced by an independent test or only by the vendor, since a vendor's own marketing page is not a neutral source even when the underlying study is real, as the Workado case shows.

We don't ask you to take a single number on faith. Our own detection methodology and results sit openly on our accuracy page — that transparency is the point, not a specific score to chase.

FAQ

What is a "false positive rate" in AI detection? A false positive occurs when a detector identifies real human writing as AI. According to a Chicago Booth study, commercial detectors keep such false positives below roughly 1% for medium- and long-length texts. Yet, if that threshold is lowered even more, more AI‑generated content would pass undetected. This adjustment reflects the same tradeoff outlined above.

What is a "policy cap," and why does it matter more than accuracy? A policy cap is a threshold an institution sets in advance for how many false accusations it's willing to tolerate, then uses to compare detectors on equal footing. It matters because two detectors with similar-sounding "accuracy" claims can produce very different real-world outcomes once you decide how much risk of a false accusation you can live with.

Does a high accuracy claim mean a detector is trustworthy? Not by itself. The FTC's case against Workado shows a real, measured accuracy number can still be misleading if it was tested on different content than what it's marketed for. Ask what the number was actually measured on before trusting it.

Should I ever rely on a single detector score to make a final decision about a person? No. Every credible source in this space — including detector vendors' own guidance — treats a score as a signal to investigate further, not standalone proof. Pair a detector result with other evidence (drafts, revision history, a conversation) before acting on it.

Keep reading

Read how we think about what a detection score actually means before leaning on one — the same framework explains why detectors struggle with technical writing (a documented STEM/code bias), and for the wider false-positive picture, see whether AI detectors can be wrong.