Searcharxiv⌕ Search
Home/Archive/arXiv

arXiv · 2605.21653

Amplifying, Not Learning: The Price of Out-of-Distribution Generalization in AI-Text Detection

Alexander Smirnov
View original source ↗Read source document ↗Export metadata ↓

Abstract

AI-text detectors gate decisions in education, hiring, and publishing, yet they flag the most fluent, formal human writing as machine-generated: they rate the median formal-native human essay as 99.5% likely AI while clearing genuine high-temperature AI at 10.5%. Deployed detectors share it (chatgpt-detector-roberta flags 56% of formal essays at a 1% false-alarm rate). This is not a calibration bug but the signature of one mechanism: a fine-tuned detector does not learn an AI-versus-human boundary, it amplifies an inherited typicality axis (predictability under a language model) that pre-exists fine-tuning, rescaling it rather than constructing one. Decomposing the detector into this inherited reading and a fine-tuned residual, the inherited part carries the bulk of cross-generator transfer and produces the over-flagging of formal humans, while the residual is generator-specific and does not transfer; a frozen-representation probe fit on ~25 labels per class matches the fully fine-tuned detector on unseen generators (cross-generator AUROC 0.893 vs 0.831). This yields a no-go: because the axis that transfers is the axis that over-flags, no training objective, threshold, concept-erasure, or ensemble we test removes the harm while preserving cross-generator detection, across three architectures, fine-tuned decoders, and zero-shot perplexity detectors. A closed-form, training-free operator relocates and diagnoses the bias (reviving a dead deployed detector, true-positive rate 0 to 0.904 at a 1% false-alarm rate) but, consistent with the no-go, is AUROC-neutral on cross-generator detection: it moves the bias, it cannot erase it. The mechanism is not specific to English prose: the over-flag and the decomposition replicate in Chinese and in code. Detector unfairness is a predictable, structural property of the detection paradigm, the price of out-of-distribution generalization.

Explore related subjects

cs.LGcs.AIcs.CL

Keep this discovery

BibTeXRIS

Alexander Smirnov. 2026-05-20. Amplifying, Not Learning: The Price of Out-of-Distribution Generalization in AI-Text Detection. https://arxiv.org/abs/2605.21653

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

AUC Maximization from Biased Positive-unlabeled Data with Confidence

Maximizing the area under the receiver operating characteristic curve (AUC) is a standard approach to imbalanced binary classification. Although positive and negative data are required for maximizing the AUC, negative data are often difficult to collect in some real-world applications due to privacy concerns or the need for specialized expertise to annotate them. Thus, AUC maximization from positive and unlabeled (PU) data has been attracting attention. Existing methods assume that labeled positive data are unbiased samples from the true positive distribution. However, this ideal assumption is often violated in practice. In this paper, we propose a method to maximize the AUC from biased PU data. To address the bias, our key idea is to exploit {\it confidence}, i.e., the probability that an instance is positive, associated with the small number of labeled positive data. We derive an estimator of the AUC risk using biased PU data with confidence, enabling AUC maximization under such bias. We further show that the rewritten AUC risk induces a Bayes-optimal AUC ranking even when the available confidence is any strictly increasing transformation of the true posterior probability. We experimentally show the effectiveness of our method on eight real-world datasets.

cs.LG↗
arXiv2026

Measuring the Value of World-Model Updates: A Counterfactual Utility Protocol for Continual Adaptation

Continual world models must decide whether new data justify changing the model. Fixed replay schedules and prediction-error triggers specify when to update, but neither reveals the value of an individual update: one deployment run cannot show how the same model would have performed at that moment had it held its parameters. We introduce the fork ledger, which branches a deployment stream at pre-registered decision points into matched update and hold continuations under common random numbers. It evaluates both continuations on the same episodes and records $\Delta R = R_{\mathrm{update}} - R_{\mathrm{hold}}$. Always applying one fixed update mechanism lowers return on all three simulated control tasks: CartPole ($-144.0$; checkpoint-bootstrap $95\%$ CI $[-185.4,-116.1]$, against a converged return near $650$), Walker ($-82.8$; $[-101.1,-61.7]$) and Cheetah ($-18.6$; $[-29.0,-6.6]$). Divergence is an outcome of applying the update, so the estimand counts every attempted fork; restricted to the $693$ of $720$ that did not collapse, CartPole and Walker are unchanged in sign ($-113.4$ and $-82.1$) and Cheetah becomes unresolved ($-3.9$; $[-17.5,+13.0]$). The task is the unit of inference: each contributes $240$ attempted forks over five pretrained checkpoints crossed with two drift directions. The ledger makes counterfactual utility observable for a fixed mechanism, allowing triggers to be judged by the updates they select rather than by surprise detection alone.

cs.LG↗
arXiv2026

When More Is Not Better: Component Anti-Synergy in a P300 Speller

P300 brain-computer interface (BCI) spellers can provide hands-free communication for people with severe motor impairments. Modern pipelines combine multiple individually promising components, often assuming that 'more-is-better'. We tested this assumption using a four-component full-factorial experiment varying the inclusion of Euclidean Alignment (EA), xDAWN spatial filtering, subject calibration, and language model priors on a public P300 dataset. Performance was evaluated using accuracy, repetitions, and information transfer rate (ITR) with mixed-effects models. Results show that the value of components is conditional rather than additive. Calibration was the strongest singular contributor, while EA compensated for its absence in zero-calibration settings. Adding independently useful components could also reduce performance, revealing component anti-synergy. Contrary to conventional wisdom, LM support was not universally beneficial: its effect depends strongly on the strength of the underlying EEG pipeline, while results from a larger LM showed a similar pattern. Together, these findings challenge maximal 'all-on' pipeline design and highlight the value of selecting spatial and language-support components according to the quality of available EEG evidence.

cs.LG↗