Un-AI-ing: Compliance, Evasion, and the Distortion of Research Writing in the Age of AI Detection

To protect research integrity from an influx of AI-generated content, academic journals have increasingly deployed AI detection tools. AI detection tools are designed to evaluate text and predict whether it was written by a human or a large language model. Rather than verifying factual truth or authorial honesty, they operate by measuring statistical predictability, primarily analyzing text through metrics like perplexity (word predictability) and burstiness (sentence structure variation). While journals adopt these tools as an efficient, scalable gatekeeping defense against academic fraud, this algorithmic surveillance has triggered a troubling counter-behavior known as un-AI-ing. Un-AI-ing is the deliberate modification of text specifically to evade algorithmic detection thresholds. Because formal, peer-reviewed scientific prose inherently relies on highly standardized and predictable language, authentic human writing is frequently misclassified as AI-generated, forcing authors to alter their work. This creates a dangerous systemic paradox divided into two behaviors. (1) Dishonest actors engage in opportunistic un-AI-ing, using AI humanizers or other maneuvers to artificially disrupt text predictability and easily launder synthetic content into literature. (2) Conversely, honest researchers, disproportionately non-native English speakers, are forced into compliant un-AI-ing. They must systematically degrade their clear, well-edited prose into awkward phrasing simply to bypass false positive thresholds. By policing metrics rather than merit, journals are not catching fraud, they are manufacturing marginalization. To restore true accountability, academic publishing must abandon automated gatekeeping and return to context-sensitive, disclosure-based authorship policies that judge the integrity of the scholar, not the algorithmic conformity of the text.