AFP journalists tested three online AI detectors—JustDone AI, TextGuard, and Refinely—and found they often labeled human-written text as AI-generated, then offered paid services to remove or “humanize” the traces.

The findings highlight how misinformation ecosystems and even credible fact-checkers relying on AI can produce errors, underscoring the need for corroborating evidence beyond automated detections.

AFP warns that false positives can be exploited to discredit authentic content and damage reputations, feeding into broader misinformation risks and a so-called “liar’s dividend.”

Cornell University says it has no established relations with the detector companies, underscoring a disconnect between claimed institutional use and reality.

phrases torturées

Detector providers acknowledge their systems are not 100% accurate and can operate offline or without internet, suggesting limitations rather than guarantees of results.

Debora Weber-Wulff describes the products as scams that push for text “humanization” and often produce questionable results.

Universities like Cornell deny formal relationships with these detectors and doubt their usefulness for AI-misuse, noting that fact-checkers sometimes use AI detectors only as a supplement to traditional methods.

Some institutions dispute claims of using detectors at all, highlighting ongoing uncertainty about detection efficacy and current method limits.

JustDoneAI, TextGuard, and Refinely reportedly operate offline or with limited accuracy, hinting at scripting or non-technical analysis behind outcomes.

Experts like Debora Weber-Wulff call these tools scams or a “liar’s dividend” mechanism that undermines trust in AI verification and can discredit legitimate content.

Warnings remain that unreliable detectors threaten AI-content verification and could erode trust in online information.