AI text detector that adjusts after release reached 93% accuracy in the team's tests
In experiments by researchers at MIT and Cornell University, a detector that keeps adjusting after release reached 93% accuracy on disguised AI writing, while a commercial detector identified about 27% of it.
In the researchers' experiments, a commercial detector identified only about 27% of AI-written text that had been disguised to look human. A detector that kept adjusting itself after release reached 93% accuracy.
The work comes from researchers at MIT, in Cambridge, Massachusetts, USA, and Cornell University. They showed that detection tools, which are AI models themselves, can be fooled by disguised text. The tools can also fail when writing habits shift over time or when new large language models (LLMs), the AI systems that write text, appear.
The paper's lead author is Kevin Ren, a graduate student at Cornell Tech, and its senior author is Nikhil Garg, an assistant professor at Cornell Tech. Co-author Manish Raghavan is the Drew Houston (2005) Career Development Professor at the MIT Sloan School of Management. The research will be presented at the Conference on Neural Information Processing Systems.
Disguising text can be automated. People can run AI-written text through "humanizer" software that changes it to look human-written to the detector. Ren describes the cycle between detectors and humanizers: "It’s like rinse-and-repeat. The cycle can go on forever."
The team's method is called test-time adaptation, which means adjusting while the detector is in use. Instead of freezing the detector so it cannot change after release, the researchers let it adapt to new text that users upload.
Developers do not know which of those texts are human or AI-written, but the texts reflect current writing. With enough examples, the detector learns patterns that tell them apart, without being retrained from scratch.
To test it, the researchers built datasets with known amounts of AI and human text and withheld those labels from the model. They also used LLMs to make examples meant to evade a detector. Then they checked the results against the true answers.
A second test mimicked natural shifts in human language over time. MIT News gives a hypothetical example: a detector trained on 2010 data might be unfamiliar with a person's writing style in 2030 and wrongly mark it as AI-generated. In the experiments, the adapting detector was 85% accurate, against 73% for traditional methods.
The researchers caution that the method may not always work as future LLMs produce more natural-sounding language and verifiable human text becomes harder to find for training. Shifts in human writing are a harder problem for the method, Ren says, because they are slower and more subtle. Raghavan says: "It isn’t entirely clear if this will be a solvable problem."
MIT News says the method could help educators, editors, publishers and hiring managers flag AI text passed off as human, and help cybersecurity experts prevent email phishing scams. The researchers want to explore similar areas and build better ways to measure how well detectors work, to help engineers improve them over time.
Sources
- Helping AI text detectors keep up with changing conditions · news.mit.edu