Artificial Intelligence1 min read

Helping AI text detectors keep up with changing conditions

By · Published by Everything Blog

In short

Traditional detection systems frequently degrade over time because language models evolve and users actively obfuscate machine text using humanizer tools. In controlled evaluations, standard commercial software caught just 27 percent of disguised synthetic passages, whereas this new test-time adaptation technique secured a 93 percent success rate. By processing contemporary human and machine samples as they are uploaded, the detector avoids the costly requirement of constant full retraining. Authored by Kevin Ren, Nikhil Garg, and Manish Raghavan, the research will be formally presented at the Conference on Neural Information Processing Systems.

Key points

  • On-the-fly model adjustments eliminate the ongoing cat-and-mouse dynamic where engineers…: On-the-fly model adjustments eliminate the ongoing cat-and-mouse dynamic where engineers must repeatedly retrain systems from scratch against new evasion software.
  • High-stakes sectors where authentic authorship is vital, including academic publishing, e…: High-stakes sectors where authentic authorship is vital, including academic publishing, education, and professional hiring, stand to benefit from the enhanced detection resilience.
  • The underlying adaptation technique offers protective applications beyond standard text v…: The underlying adaptation technique offers protective applications beyond standard text verification, extending to cybersecurity domains like phishing prevention and search engine optimization fraud detection.

Sources checked by our analysis pipeline: 1

A teacher who believes a student’s term paper may have been written by a large language model (LLM) can use a tool for detecting artificial intelligence-generated text to help confirm their suspicions. But the teacher may want to think twice before putting all their trust in the detector.

Researchers from MIT and Cornell University showed that such detection tools, which are AI models themselves, can be fooled by AI-generated text disguised to appear as if it were written by a human. Detectors can also fail to identify AI text due to natural shifts in writing habits over time or the emergence of new LLMs.

In a new paper, the researchers demonstrated how a special training technique makes an AI detector more accurate and resilient to deception, even as human and AI writing constantly changes.

They employed a method called test-time adaptation, which allows a deployed detection model to learn from new, unlabeled data by adjusting its inner workings on the fly. In this way, the model adapts to unseen inputs without being retrained from scratch.

In experiments, a commercial detector…

Artificial Intelligence