The 1D Collapse in AI Text Detection: Why Supervised Detectors Fail on Formal Writing and How to Fix It — Technical Report

We present a geometric analysis of supervised AI text detectors revealing that every tested detector collapses its 768-dimensional representation to effectively onedimension, and that this single dimension aligns with text formality rather than AI authorship. We demonstrate this 1D collapse across 8+ detectors from 5 independentlaboratories spanning 5 architecture families (RoBERTa, ELECTRA, DeBERTa, DistilRoBERTa, ALBERT), with formality alignment cosines ranging from 0.73 to 0.99. Thecollapse is pre-training-intrinsic, task-specific to AI detection, flat during training,and concentrates in transformer layers 8–12. We provide a mechanistic explanation:collapse occurs whenthecosinesimilarity betweenthebetween-classaxisandtheformality axis exceeds 0.6 and formality is a confound rather than the target, correctlypredicting collapse/no-collapse for 5/5 binary classification tasks. We show that existing post-hoc fixes (Kuznetsov et al., EMNLP 2024) cannot resolve the entanglementbecause the formality confound and the detection signal share the same principalcomponent, confirmed via LEACE concept erasure. We propose two complementarytraining interventions: (1) adversarial formality training via gradient reversal, whichreduces formality alignment from cos=0.98 to cos=0.45 and improves EvoBench TPRby+5.6percentagepointsovercross-entropy on identical data, and (2) AI-contrastivetraining, which pushes AI text representations together across task domains, achieving 49.6% Ghostbuster TPR without any Ghostbuster training data. Applied to external published detectors, our training recipe improves OpenAI’s detector from 0.326to 0.806 mean AUROC (+0.480) and Hello-SimpleAI from 0.659 to 0.743. Formalityscore predicts detector FPR with Pearson r=0.991 across 7 human populations (validated on held-out MAGE with 1.2% error), and cross-architecture early warning fromour ELECTRAmanifoldpredictsDeBERTa’sfailures with Spearmanρ=−0.800. Allexperiments ran on a single RTX 3080 consumer GPU in approximately 120 GPU-hours