Attention Capture Is Not Detection: A Two-Stage Account of How Humans Miss Localized AI Image Edits

A marketplace review photograph is a document: platforms approve refunds on it, and generative models drove the cost of forging one to zero. We study that detection problem, so we build a detector and attach an attribution map as its evidence, then measure what that pair delivers on 186,527 images under controls designed to change our conclusions when something is wrong. Compression history, not synthesis, drives naive evaluation: our strongest model reaches 0.9999 PR-AUC (area under the precision-recall curve) on a product-disjoint split, yet falls to 0.7254 once we re-encode synthetics into the real class's format, while five public detectors move by at most 0.07. Aligning one class relocates the cue rather than removing it, and the repaired model then assigns native files a median probability of synthesis of 0.0004. One identical final encode for both classes repairs that, and a three-seed factorial credits the encoding change with the whole gain (+0.176 +- 0.009 PR-AUC). That encode equalises the last stage only: forensic features alone still separate the classes at 0.7145 against a base rate of 0.254. For evidence we test maps causally, against controls that never consult the detector. Whether an attribution ranking exists at all depends on whether the detector reacts to the image. On our first-fix detector, which calls 96 of 100 edited frames real, no map beats a random one. On the detector we selected, twelve of seventeen maps clear that control on edited images and eight on generated ones; perturbation leads both axes and no gradient-CAM variant shows a positive advantage. The trivial controls never clear it, and on generated images the centre prior is worse than random. Our ensembled regional map clears both axes and takes the top pixel AP at 12.4 s per map against 44.9 for occlusion. Clearing a detector-blind control is not yet a faithful explanation, and we demonstrate none.

Large language models are increasingly used to edit human-written text rather than generate entire texts from scratch. Conventional AI-text detectors mainly distinguish human-written from fully AI-generated text, while recent methods for AI-edited text typically provide only a text-level label or editing-degree score. We introduce MixDetect, a word-level framework for localizing and quantifying AI editing. MixDetect separately predicts whether each word has been edited and, conditional on editing, how substantial the edit is, allowing editing scope and editing intensity to be estimated separately. During training, source--edited pairs are aligned to construct word-level supervision, while inference requires only the input text. Experiments show that MixDetect accurately localizes AI-edited words, reflects differences in editing intensity, and reveals different scope--intensity patterns across editing degrees and operations. The overall AI editing magnitude increases under additional AI editing, decreases when AI-generated text is edited by humans, and remains nearly unchanged under ordinary human-to-human editing. The aggregated text-level predictions also perform well on binary and ternary AI-text classification and remain effective under domain and generator shifts. These results show that AI editing can be analyzed beyond a single authorship label or editing-degree score by identifying both where AI editing occurs and how substantial the edits are.

Detecting AI-generated images is only half the task: a deployed detector must also justify its verdict, yet existing detectors inherit three failure modes from their training data: real and fake images collected from different sources invite provenance shortcuts, supervised explanation corpora teach templated rationales, and a static forgery corpus leaves the decision boundary standing still while generators keep moving. We introduce \methodname{}, an adversarial reinforcement learning framework that pits two heterogeneous models against each other. A diffusion image editor learns to edit real photographs into fake counterparts of those same photographs that fool the current detector, while a reasoning MLLM learns to expose them with a verdict grounded in free-form reasoning. Both rewards are shortcut-proof by design: the attacker is credited only when its edit is faithfully executed, and the defender only when its verdict is correct. As the two models alternate, each round's attacker regenerates a harder training pool aimed at the current detector's blind spots, so the detector must generalize rather than memorize any fixed artifact distribution. Although the explanation is never rewarded, its quality rises round over round as a side effect of accuracy-only training. A detector trained within this loop improves monotonically across rounds on each of three external benchmarks.