A teacher who suspects that a term paper was written not by a student but by a large language model (LLM - a neural network trained to generate coherent text) can use an AI-text detector. But trusting such a tool completely is risky: researchers from the Massachusetts Institute of Technology (MIT) and Cornell University have shown that detectors are themselves artificial intelligence models and can be fooled.

A detector makes mistakes when the text being evaluated differs greatly from the text it was trained on. This is called distribution shift - a situation in which new data do not resemble the training data. In addition, people change their writing style over time, and one LLM is replaced by another.

In a new paper, the scientists applied the method of test-time adaptation (TTA). It allows an already released model to continue training on new unlabeled data, adjusting its internal parameters on the fly, without full retraining from scratch.

In experiments, a commercial detector recognized only about 27% of disguised texts generated by AI. The TTA approach reached an accuracy of 93%. In tests simulating natural language change over time, the detector with TTA was accurate in 85% of cases versus 73% for traditional methods.

The authors note: the method will be useful to teachers, editors, publishers and hiring specialists who need to distinguish human-written text from text created by a neural network. It can also help in cybersecurity - for example, in identifying phishing emails or search engine optimization fraud.

"There are many high-stakes situations where people really care about the authorship of a text - from scientific publications to short story contests. We have shown that test-time adaptation makes a detector much more robust, even if AI users are constantly trying to circumvent it," says study co-author Manish Raghavan, a professor at the MIT Sloan School of Management.

The lead author of the work is Kevin Ren, a graduate student at Cornell Tech, and the senior author is Nikhil Garg, an associate professor at Cornell Tech. The results will be presented at the NeurIPS 2026 conference (Conference on Neural Information Processing Systems), which will be held December 6-13 in Sydney, Atlanta and Paris. The paper is available on the arXiv preprint server.

The researchers warn: TTA will not always be an effective solution. Future LLMs may generate more natural text, and there will be fewer verifiable human texts for training. "It is not at all clear whether this problem is solvable in principle. Big open questions remain about how to think about science, literature and art in a world where it is impossible to distinguish human content from AI content," Raghavan concludes.

Source: Tech Xplore