The target keeps moving
A strong score on yesterday's writing is not a guarantee that an AI text detector will recognise tomorrow's. MIT highlighted research on 18 November that examines three changes: new language models, deliberate rewriting of AI output and shifts in human writing. The paper was first posted in June and revised in October; this is an explainer prompted by the new research account, not a claim that the paper appeared this week.
The problem is called distribution shift. A detector learns patterns from examples labelled human or AI. If later submissions differ from those examples, the boundary it learned may no longer fit. A rewritten AI passage may look human to the detector; unfamiliar human writing can create the opposite problem.
What adapting a detector changes
Kevin Ren, Manish Raghavan and Nikhil Garg tested a method that adjusts a detector using new text whose authorship is not labelled. Rather than treating each submission in isolation with a frozen model, it uses patterns across a group of incoming texts alongside training examples.
In the paper's adversarial experiment, Pangram detected 27.0% of the disguised AI text, compared with 93.3% for the researchers' test-time approach. These are detection rates for that AI-text condition, not a statement that the new method correctly judges 93.3% of all writing. The tests used research datasets and particular rewriting methods.
Why a score is not a verdict
MIT's account says adaptation could help tools keep up with changing inputs, but also reports the authors' warning that future models may make detection harder. Obtaining recent text that is verifiably human-written is itself a challenge.
The useful lesson is to ask what a detector was tested on, what its score measures and how often it labels human work incorrectly. An improvement in one experiment does not establish reliable performance in every classroom, newsroom or language. This study offers a way to investigate changing writing patterns, not a universal proof of authorship.
Sources
- Ren, Raghavan and Garg: Hitting a Moving TargetFull research paper, methods, adversarial tests, measurement definitions and limitations; arXiv v2 dated 14 October 2026.
- MIT News: Helping AI text detectors keep up with changing conditions18 November research account and authors' cautions about future detection and verified human training text.



