Key points

  • A commercial detector flagged about 27 percent of disguised AI text; the adapted detector caught more than 93 percent.
  • When human writing shifted over time in the tests, the adapted detector was 85 percent accurate, versus 73 percent for traditional methods.
  • The method, test-time adaptation, updates a deployed model using unlabeled data without retraining it from scratch.
  • The work will be presented at the Conference on Neural Information Processing Systems.

Researchers at MIT and Cornell University report that AI text detectors are much harder to fool when they keep learning from unlabeled user writing after release. In their tests, a commercial detector flagged only about 27 percent of AI-generated text disguised to pass as human. Their adapted detector caught more than 93 percent. According to an MIT press release, the work is led by Cornell Tech graduate student Kevin Ren and will be presented at the Conference on Neural Information Processing Systems.

Detectors are AI models themselves, trained on text already known to come from people or from machines. They can falter when new text looks different from what they learned on, a problem called distribution shift. That can happen when human writing habits change, when new large language models (LLMs) appear, or when people run AI output through humanizer software designed to slip past detectors.

The team's answer is test-time adaptation, which lets a model already in use adjust its internal workings as new data arrives, without starting training over. To test it, the researchers built datasets that mixed human and AI writing, including LLM-written passages meant to evade detection. They hid the labels from the model and then graded it against the true answers. In experiments that imitated slow changes in human language, the adapted detector reached 85 percent accuracy, compared with 73 percent for traditional methods.

The researchers caution that the approach may not always be effective as future LLMs write more naturally and verified human text gets harder to find. Co-author Manish Raghavan of MIT says it is not entirely clear whether the problem will be solvable.

Why it matters

The researchers say the method could help educators, editors, publishers and hiring managers spot AI writing passed off as human. It could also help cybersecurity experts fight phishing emails and search engine optimization fraud.

What we don't know yet

The approach may become less effective as future LLMs produce more natural-sounding text and verified human writing becomes harder to obtain. Adapting to slow, subtle changes in human writing is also harder than adapting to deliberate evasion. The team plans to build better benchmarking tools.

Terms in this story

Test-time adaptation
A technique that lets an AI model already in use adjust itself to new, unlabeled data instead of staying frozen after training.
Distribution shift
A mismatch between the data a model was trained on and the data it meets later, which can make it less accurate.
Humanizer
Software that rewrites AI-generated text so that detection tools judge it to be human-written.

Sources and verification

  1. Helping AI text detectors keep up with changing conditions Massachusetts Institute of Technology · Nov 18, 2026
Fact-checked against the source on November 18, 2026: 26 factual claims verified.
Written in our own words: longest phrase shared with the source is 11 words.
Related image from Wikimedia Commons (CC BY-SA 4.0), checked for relevance and that it is not AI-generated. It is not an image from the study.

DexScience articles are produced with AI assistance and an automated fact-check against the linked sources. Spotted an error? Tell us and we will correct it openly.