AI detectors caught more disguised text when allowed to adapt, researchers find
MIT and Cornell’s experiments suggest detectors can learn from incoming text without new human-or-AI labels. The researchers still question whether detection will remain feasible.
Listen to this story
The audio brief
Story brief
3 key pointsA Cornell–MIT research team tested whether AI-text detectors can keep pace with changing writing by adapting to new, unlabeled passages after deployment. The approach performed better than a commercial detector on deliberately disguised AI text and improved results in simulated shifts in human writing, but the experiments do not establish broad accuracy in real-world use. The team says detection could become harder as AI writing grows more natural and plans to develop benchmarks for evaluating detectors as conditions change.
- 01
In the adversarial-text experiment, Pangram detected 27.0% of disguised AI passages, versus 93.3% for the adaptive approach.
- 02
In simulated natural-writing shifts, the adaptive detector reached 85% accuracy, compared with 73% for traditional methods.
- 03
Adaptation uses similarities among incoming passages without requiring labels identifying each passage as human- or AI-written.
An AI-text detector need not stay frozen while the writing it judges keeps changing. In a November 18 research account, MIT described work with Cornell showing that detectors improved when allowed to learn from incoming text. The approach tackles both disguised AI writing and shifts that can cause human prose to be wrongly flagged.
The study, “Hitting a Moving Target,” examines a weakness in detectors trained on a fixed collection of human and machine writing: their lessons can age badly. Its preprint first appeared in June and was revised in October. The work is slated for presentation at the Conference on Neural Information Processing Systems, or NeurIPS.
Learning without an answer key
Conventional detector training starts with examples whose origins are known. Developers gather human-written and AI-generated passages, label them, and teach a model to distinguish the two. Once deployed, however, that detector encounters writing unlike its training material. Researchers call this a distribution shift—a change in the data the model must judge.
The paper identifies three sources of that change: tools that rewrite AI output to evade detection, newly released language models, and gradual changes in human writing after widespread AI use. These are not all deliberate attacks. A detector can also mistake unfamiliar human language for a machine’s work.
Test-time adaptation changes what happens after deployment. Instead of leaving the detector fixed, it adjusts the model using newly submitted passages, without requiring developers to supply a human-or-AI label for each one. The detector can adapt to unfamiliar inputs without being retrained from scratch.
The authors’ method draws on similarities among the unlabeled samples arriving during use. With enough examples, the detector learns patterns that help separate AI-generated from human-written text. Those incoming passages supply contemporary writing examples, even though their individual origins are not provided to the model.
Two tests of resilience
To test the approach, the researchers built datasets containing known amounts of human and AI writing, but withheld those labels during adaptation. They also generated examples intended to evade a detector. After adaptation, they evaluated the model against data with known answers.
The strongest contrast involved deliberately disguised AI output. The paper’s abstract reports that commercial detector Pangram identified 27.0% of the researchers’ adversarial text, compared with 93.3% for their adaptive approach. These figures measure the share of disguised AI text caught in that experiment, not accuracy across all human and AI writing.
MIT reports 85% accuracy for the adaptive detector in experiments designed to mimic natural changes in human language.
Traditional methods reached 73% accuracy in the same simulated-shift experiments described by MIT.
The project used API credits supplied by Pangram to access its detectors. Lead author Kevin Ren, a Cornell Tech graduate student, says adapting to human writing changes is harder: those shifts happen more slowly and subtly than deliberate attempts to disguise AI text.
Better detection, with an uncertain horizon
The researchers see potential uses for educators, publishers and hiring managers trying to identify AI writing presented as human work. But they caution that adaptation may not remain effective as language models produce more natural-sounding text and verifiable human writing becomes harder to obtain for training.
“It isn’t entirely clear if this will be a solvable problem,” MIT co-author Manish Raghavan said. The team wants to build better benchmarks for improving detectors as conditions change—making continued evaluation, rather than a single fixed test, part of its next research task.
Sources
- arxiv.orgHitting a Moving Target: Test-Time Adaptation for AI Text Detection under Continual Distribution Shift
- news.mit.eduHelping AI text detectors keep up with changing conditions
Reader comments
Newest comments first. Replies stay oldest first.