A recent piece from the journal Live Science highlights a persistent technical challenge in the world of generative AI: software designed to detect AI-written text is failing to keep pace with the technology it’s meant to police.
The “Training Data” Problem
According to the report, the core issue lies in how detection tools are built. Most current detectors are “learning-based,” meaning they are trained on vast datasets of known human writing, compared with known AI writing. They function by identifying statistical patterns, specifically looking for text that closely resembles the AI data they were trained on.
However, this reliance on training data creates a significant gap. As researcher Ambuj Tewari points out, when new, more advanced AI models are released, they generate text with different patterns and higher complexity than previous generations. Because this new output differs substantially from the detector’s older “training corpus,” the software frequently fails to flag it, resulting in a drop in accuracy.
A Game of Catch-Up
The result is a perpetual game of “cat and mouse” where detection tools are functionally obsolete the moment a new large language model (LLM) enters the market. Until detectors can identify AI writing based on universal characteristics rather than historical comparisons, identifying synthetic text will likely remain unreliable.
False Positives
Beyond missing AI text, these tools face growing scrutiny for flagging authentic human writing as artificial. Research suggests that non-native English speakers are disproportionately at risk, as their writing often employs simpler, more predictable sentence structures that algorithms mistake for machine logic. A piece from December 2025 on the site Proofademic outlines some of the consequences of relying on those tools.
Read the full story in Live Science on AI detection technology: Even AI has trouble figuring out if text was written by AI — here’s why