Skip to content
AI.info

The Pulse

MIT and Cornell study tests adapting AI text detectors to new writing

Researchers tested a method that lets AI text detectors adapt to new, unlabeled writing after deployment. Their approach identified 93.3% of adversarially generated text in one experiment, compared with 27.0% for a commercial detector.

MIT and Cornell study tests adapting AI text detectors to new writing

An AI text detector that adapts to new writing identified 93.3% of adversarially generated text in an experiment, compared with 27.0% for the commercial detector Pangram, according to a study by MIT and Cornell Tech researchers. The work, reported by MIT News on November 18, examines how detectors can respond when the text they encounter changes after deployment.

Source arxiv.org

The paper, a preprint posted on arXiv, is by Kevin Ren, a Cornell Tech graduate student; Manish Raghavan, a professor at MIT; and Nikhil Garg, an assistant professor at Cornell Tech. The research will be presented at the Conference on Neural Information Processing Systems.

When the text changes

AI text detectors are typically trained on examples labeled as human-written or AI-generated. They can struggle when writing presented to them differs substantially from the material used in training. The researchers call the change a distribution shift.

One source of change is deliberate: people can use automated “humanizer” tools to alter AI-generated writing so a detector is less likely to flag it. The study also examined shifts caused by new language models and changes in human writing over time.

A detector trained on older writing, for example, may encounter unfamiliar topics or expressions in later human work and incorrectly classify it as AI-generated. New language models can also produce text unlike the examples a detector learned to recognize.

Learning after deployment

The team explored a method called test-time adaptation. Rather than keeping a detector fixed after release, the approach lets it adjust to new pieces of text submitted by users, without retraining from scratch.

Those incoming examples are unlabeled: the detector is not told which were written by people and which were generated by AI. The approach uses similarities among a sufficiently large batch of unlabeled examples to help distinguish the two, the paper explains.

For their tests, the researchers created datasets containing known amounts of human and AI text but withheld the labels from the adapting model. They also generated examples designed to evade detection, then tested the adapted model against data whose authorship was known.

Their experiments included data on scientific abstracts and the RAID benchmark. The researchers also studied text from language models that had not been seen during detector training.

Results and limits

In the adversarial-text experiment, the researchers found that the adaptive approach detected 93.3% of the generated text, while Pangram detected 27.0%. In a separate experiment designed to mimic a natural shift in human language over time, the adaptive detector was 85% accurate, compared with 73% for traditional methods, MIT News reported.

The paper also reports a test following the release of GPT 5.4 in which Pangram classified only 34% of its outputs as fully AI-generated or AI-assisted, while performing nearly perfectly on older models. The result concerns that experiment, not every use of the detector.

Adapting to changes in human writing is harder because those shifts occur more slowly and subtly, Ren told MIT News. The researchers caution that adaptation may not be enough if AI and human writing become more alike or if verified human-written text becomes harder to obtain.

The team plans to explore other strategic settings where test-time adaptation might help and to build better benchmarking tools for detection models. The project used API credits provided by Pangram to access its AI text detectors.

Sources

  1. arXiv
  2. MIT News

Related stories

Explore

More articles