Washington, Nov. 20 (SANA) Researchers at the Massachusetts Institute of Technology (MIT) and Cornell University have developed a new method for detecting AI-generated text, even after it has been rewritten using tools designed to make machine-produced content appear human.
According to MIT News, the system uses a technique known as test-time adaptation, allowing it to learn from new material during analysis rather than relying solely on its original training data.
Unlike conventional detectors that examine texts individually, the new approach analyzes groups of passages to identify shared patterns. It draws on a phenomenon researchers call inference-time homogeneity, in which AI-generated texts exhibit similarities that become more apparent when examined collectively.
By adapting to previously unseen material, the system can identify AI-generated content from newer language models and recognize text that has been altered by paraphrasing tools.
In tests, the method detected 93.3% of AI-generated texts rewritten using so-called humanizer tools, compared with 27% for Pangram, an existing detection tool used as a benchmark.
The system also achieved detection rates of 87% to 94% when tested on language models outside its original training data, outperforming conventional detectors that had not been updated.
Researchers examined its ability to distinguish human writing from AI-generated content. While one conventional detector’s accuracy in identifying human-written texts declined from 85.9% to 73.4% across a 10-year period, the new system maintained accuracy of between 86% and 90%.
The researchers said the technology could help universities, publishers, news organizations and employers assess the origins of written material. Potential cybersecurity applications include detecting AI-generated phishing emails and identifying attempts to manipulate search engine results.
The findings are scheduled to be presented at NeurIPS, an international artificial intelligence conference, later this year.
Although the results suggest improved detection capabilities, the method does not guarantee error-free identification, and human verification remains important.
TS/ABD