Over the past few years, AI-generated text has emerged as a huge problem across industries. The sheer scale at which AI content is being produced has led not just to a number of ethical problems, but has also become a huge headache for academia, the publishing industry, and related domains in general.
Now, there are plenty of AI text detectors out there, such as GPTZero and Pangram, that claim to reliably sniff out text generated using a large language model, but their accuracy remains a white-hot debate. The biggest problem isn’t just reliability, but the false positives that have landed many people in trouble, despite their work being fully human-written.
AI detectors themselves have to contend with a bunch of problems. The first is that they’ve been trained on samples of human and AI-generated material, and they don’t update themselves in real time. That leaves them vulnerable to AI-generated material cooked up by newer and more advanced large language models that increasingly write in a more human-like tone.
“It isn’t entirely clear if this will be a solvable problem”
Another major hurdle is human language itself. Research suggests that the way we talk is influenced by what we’re exposed to, whether it’s entertainment like films or social media. Interestingly, there’s also growing evidence that, owing to regular AI interactions, we’re increasingly using words and phrases favored by large language models. Broadly, the combination of all these factors raises the question of whether an AI detector that’s reliable today will be as effective tomorrow.

And let’s not forget the biggest challenge of them all: humanizer tools. There are plenty of third-party humanizer tools out there that are widely used. To make matters worse, it’s fairly easy to create your own. Research has shown that fine-tuning a model to match a specific author’s style can help its output largely evade detection.
Researchers at Cornell Tech and MIT have now applied a technique called test-time adaptation (TTA) to AI text detection, aiming to solve the static nature of existing detector tools. The key premise of TTA is that the detector doesn’t rely solely on what it learned during training. Instead, it also learns from the text it’s asked to check.
For example, when it’s given a large batch of material to test, such as a pile of essays, it looks for patterns across the whole batch rather than judging each piece in isolation. That works because AI-generated text tends to share a certain uniformity when you see a lot of it together, something the researchers call “inference-time homogeneity.” The detector then combines what it already knows from labeled examples with this new, unlabeled batch to work out which style looks human and which looks too uniform.
Additionally, the TTA approach continuously retrains its own decision boundaries using the material it’s given, so that every new batch of text serves not only as a task, but also as fresh training material that makes it better at telling human and AI-generated text apart.
During testing, the TTA approach proved a lot more effective than Pangram, a commercial detector that previous research found to have the lowest error rates among popular tools, and which the researchers used as their benchmark. When the two were fed disguised AI text, essentially AI-generated text that had been passed through a humanizer, Pangram caught only 27% of it, while TTA’s detection rate stood at an impressive 93.3%.
New AI models posed a similar problem. When OpenAI’s GPT-5.4 was released in March, the researchers found that Pangram flagged only 34% of its output as fully or partially AI-written, though its performance improved significantly after an update in May. TTA, meanwhile, held up well across AI models it hadn’t been trained on, detecting AI text from them roughly 87% to 94% of the time.
Finally, the researchers tested how a conventional detector handles human writing as language naturally shifts over the years. A standard detector trained only on older writing correctly recognized human text 85.9% of the time at first, but that fell to 73.4% a decade later. The TTA-based system held up far better, staying between roughly 86% and 90%. That means far fewer false positives, the very errors that land innocent writers in trouble.
“This method could help educators, editors, publishers, and hiring managers flag LLM-generated writing that people try to pass off as human-written. It could also be used by cybersecurity experts to prevent email phishing scams or identify search engine optimization fraud,” says the team behind the research in a press statement. The research group will present their findings at the Conference on Neural Information Processing Systems (NeueIPS) conference later this year.