MIT, Cornell Boost AI Detector Accuracy to 93% With Test-Time Adaptation

AI text detectors that teachers, editors and hiring managers lean on are getting a hard-won resilience upgrade, with MIT and Cornell results published on Wednesday.

In experiments, a commercial detector caught only about 27 percent of AI-generated text that had been disguised to look human-written. The new method, built on a technique called test-time adaptation, hit 93 percent on the same challenge. MIT News reported the results on Wednesday, and the team has posted the paper on arXiv.

Test-time adaptation, or TTA, lets an already-deployed model keep learning from fresh, unlabeled data — new pieces of text uploaded by real users — without a full retrain. The model adjusts its inner workings on the fly, which is exactly what a detector frozen at release time cannot do.

“There are many high-stakes contexts where people really care about the identity of the author, from scientific publishing to short story competitions. We’ve shown that test-time adaptation allows your AI detector to be much more robust, even with continual adversarial behavior by AI users,” said Manish Raghavan, the Drew Houston (2005) Career Development Professor at the MIT Sloan School of Management, who also holds a shared position in MIT’s Schwarzman College of Computing and Department of Electrical Engineering and Computer Science and is a principal investigator in the Laboratory for Information and Decision Systems.

The stakes are practical, not academic. The MIT and Cornell team demonstrated that these detection tools — which are themselves artificial intelligence models — can be deceived by AI-generated text that has been rewritten to pass as human work. Detection systems may also miss AI text because writing styles naturally evolve over time or because newly released LLMs appear on the scene.

Detectors trained on a frozen world

Every commercial detector is itself an AI model, trained on text the developer has labeled as human- or machine-written. The failure mode the researchers identified is a classic distribution shift: the model degrades when the text it evaluates looks meaningfully different from its training corpus.

Mit movingtarget press
Photo: news.mit.edu

Two forces drive that shift. One is slow and organic — a human writing in 2030 will reach for topics and analogies nobody used in 2010, so a detector trained on 2010-era prose may flag genuinely human writing as machine-made. The other is fast and deliberate: adversarial users actively gaming the model.

Once a detector ships, the evasion industry mobilizes. Automated “humanizer” tools rewrite LLM output to slip past the classifier, developers fold that humanized text into the next training run, and the wheel turns again.

“After the company releases a model, a bunch of ‘humanizers’ pop up to try to beat it, so the company tries to incorporate this humanized text into their training so the detector can identify it, and so on. It’s like rinse-and-repeat. The cycle can go on forever,” said Kevin Ren, a graduate student at Cornell Tech and the paper’s lead author.

Adapting to text nobody labeled

The core insight is that the incoming traffic is the training data. Users uploading essays, cover letters and articles to a detector are unknowingly supplying a live sample of contemporary human and AI writing — precisely the distribution the model was built to classify.

Story on The Morning Science
Photo: themorning.science

That data arrives unlabeled; the developer has no ground truth on which uploads came from an LLM. TTA sidesteps the labeling problem entirely, letting the model extract patterns from the stream itself and update its parameters in place rather than waiting for the next manual retrain-and-ship cycle.

The team stress-tested the approach by building datasets mixing AI-generated and human-written text, withholding the labels, then generating adversarial examples with LLMs specifically designed to evade detection. After adaptation, the models were evaluated against ground-truth data.

27 percent versus 93

The adversarial results are the headline. Against disguised text, the baseline commercial detector managed roughly 27 percent; the TTA-equipped model cleared 93 percent.

Both margins matter for different buyers. The disguise number is what protects publishers, hiring managers and cybersecurity teams hunting AI-assisted phishing and SEO fraud. The drift number is what keeps a detector from falsely accusing honest writers as language evolves.

“TTA has the potential to really help deployed AI-detection models in terms of being able to proactively address this iterative ‘gaming’ behavior by adversarial AI users,” Ren said.

Where the fix runs out

The researchers flag two limits. Adapting to slow, subtle shifts in human writing proved harder than adapting to adversarial evasion. And the ceiling is structural: as future LLMs produce more natural-sounding language, verifiable human text — the raw material any detector needs — gets scarcer.

Next on the calendar is a formal airing at the Conference on Neural Information Processing Systems, NeurIPS 2026, running Dec. 6–13 across Sydney, Atlanta and Paris. The team also plans to explore other strategic domains where TTA could apply and to build better benchmarking tools for engineers tuning detectors dynamically.

Raghavan’s senior co-author on the paper is Nikhil Garg, an assistant professor at Cornell Tech. Whether the detection problem stays solvable at all is, by the researchers’ own account, an open question rather than a settled one.

“It isn’t entirely clear if this will be a solvable problem. There are some big, open questions about how you should think about science, writing and art in a world where it is infeasible to tell the difference between human-generated content and AI content,” Raghavan said.

Photo of author

Technology Desk

Technology Desk at Archyde.

England Beat Australia by 97 Runs to Seal ODI Series 2-1