Test-Time Adaptation Boosts AI Text Detection

Researchers at Cornell Tech and MIT report a detector that learns from unlabeled batches of text to flag disguised AI writing more accurately than an existing tool.

Updated on Nov. 19, 2026 in Artificial Intelligence

Test-Time Adaptation Boosts AI Text Detection

Live Poll

Do you trust AI detection tools to accurately tell human writing from AI writing?

Researchers at Cornell Tech and MIT applied test-time adaptation to AI text detection, and the method outperformed Pangram at identifying humanizer-disguised AI text. The detector caught 93.3% of disguised AI text, compared with 27% for Pangram.

Why it matters

Existing detectors are trained on static samples of human and AI-generated text and do not update in real time, which can leave them less reliable as new models appear. A detector that adapts to the batch it is given could make it harder for disguised AI writing to pass as human.

The detector learns patterns across a whole batch of essays and combines knowledge from labeled examples with a new unlabeled batch to identify uniform, AI-like style. Its human-text recognition stayed between roughly 86% and 90% over time, while a conventional detector fell from 85.9% to 73.4% over a decade.

The players

Cornell Tech

Cornell Tech is a graduate and research campus of Cornell University in New York City focused on technology and its applications.

MIT

The Massachusetts Institute of Technology is a private research university in Cambridge, Massachusetts, known for its computer science and artificial intelligence research.

The details

Test-time adaptation learns from the unlabeled text batch it is asked to check, continuously retraining its decision boundaries using the material it is given. The approach targets inference-time homogeneity, the uniformity AI-generated text shows across large batches, and the findings are to be presented at the Conference on Neural Information Processing Systems later this year.

Timeline

  1. March: OpenAI released GPT-5.4, which Pangram flagged as fully or partially AI-written 34% of the time.

  2. May: Pangram received an update that significantly improved its performance.

  3. Later this year: Findings to be presented at the Conference on Neural Information Processing Systems.

The Tech Race

The work is positioned against Pangram, a commercial detector whose accuracy on disguised text trailed the new approach in the reported tests. It reflects an ongoing contest among detection tools to keep pace with rapidly updated AI writing models.

Readers who rely on AI detection tools for schoolwork, hiring, or publishing may see more accurate flags for disguised AI text in the future. The research suggests detectors that adapt over time could better keep up with new models than static tools.

The takeaway

Detection tools that adapt to the text in front of them may hold up better as AI writing models change. Readers should still treat any single detector's verdict as one signal rather than proof.

What happens next

The findings are scheduled to be presented at the Conference on Neural Information Processing Systems later this year.

Further reading

Explore more on Artificial Intelligence developments shaping text detection and generative models.

Source note: This article includes information reported by Digital Trends.

Live Poll

Do you trust AI detection tools to accurately tell human writing from AI writing?