AI detection tools are broken. The ones that claim to detect AI generated text are wrong often enough to be useless, the ones that claim to detect AI generated images are wrong in ways that are getting people in real trouble, and the ones that claim to detect AI generated code are barely better than random. The truth is that the detection problem is not a tooling problem, it is a fundamental property of the medium.
Turnitin’s AI detection tool flagged 10% of papers as AI generated in its first year. By 2025 the false positive rate was high enough that the company settled a class action lawsuit for $1.6 million. OpenAI killed their AI text classifier in 2023 because the accuracy was too low. The image detectors are getting better but they are still being used in contexts where the false positive cost is a job, a scholarship, or a deportation. The code detectors are even worse. None of this is a tooling gap. All of it is a fundamental impossibility gap.
Why detection does not work for text
Text detection is the original problem and it is the one most people have given up on. The fundamental issue is that language models produce text that is statistically indistinguishable from human text at the level of the features the detectors look for. Perplexity, burstiness, token distribution, all of them are statistical properties that the models have been trained to match. The detectors look for signatures that the models no longer have. Even the latest models that claim to add invisible watermarks to AI generated text are working at a layer the detector does not have access to. The text detection market is dead. The vendors selling it are selling snake oil.
Why detection is dangerous for images
Image detection is in worse shape than text detection because the consequences of false positives are worse. A false positive on a text detector is a teacher marking a paper wrong. A false positive on an image detector is someone losing their job, getting expelled, or getting deported. The current image detectors are trained on the outputs of the models that existed when they were trained, which means they are increasingly out of date as the models improve. They are also biased against certain types of images, certain types of people, and certain artistic styles. Using them in any high stakes context is malpractice.
What to do about it instead
Three moves if you are responsible for the AI detection question in your organisation. Stop trying to detect AI generated content and start focusing on outcomes. The student who is using AI to write their paper is not the same problem as the student who is not learning the material. Address the underlying problem rather than the symptom. Invest in provenance over detection, which means the cryptographic signing of original content, the timestamping of human work, and the documentation of the human process. Assume the detection tools are unreliable and build your policies around that assumption, because the tools are not going to get better fast enough to matter.
The bottom line
AI detection tools are broken at a fundamental level, the false positive cost is real, and the time spent on detection is better spent on outcomes and provenance. Stop buying detectors. Start designing for the world where AI generated content is undetectable.
Sources & Further Reading
All claims in this article are sourced from primary documentation, vendor advisories, and reputable security researchers.
Spotted an error? Email the editor. Corrections are issued with a visible correction note.
Editorial standards. Every article on humanrequired.org is reviewed by a human editor before publication. AI may assist with drafting or research; final editorial control is human. Read the full standards.
