SafeIMG
SafeIMG evaluates whether today’s detectors can recognize AI-generated images when they imitate visual evidence used in high-risk decisions.
Southeast University
The Hong Kong Polytechnic University
Wuhan UniversityModern generators can now produce images that appear to document real events, identities, transactions, and private records. In these settings, a detection error can affect public judgment, financial trust, identity verification, legal interpretation, or personal reputation.
Can a detector identify AI-generated images when they imitate evidence used in high-stakes real-world decisions?
Generator gap
Many established benchmarks were built with GANs or early diffusion systems. Strong performance on those sources does not guarantee generalization to images produced by current models such as GPT Image 2.
Scenario gap
Natural scenes, artworks, and scientific figures are valuable test domains, but they do not directly measure the harms caused when synthetic images imitate disaster reports, transaction proofs, private communications, or identity endorsements.
Diagnostic gap
A single score hides large differences across risk domains, evidence types, and dissemination conditions. SafeIMG diagnoses category-level failures, explanation alignment, and robustness after compression, cropping, screenshots, and re-encoding.
Annotation emphasis. Beyond visible face and text artifacts, annotators prioritize subtle commonsense conflicts—temporal, physical, quantitative, and contextual errors that may remain difficult even as generation quality improves.
Select a category to explore its evidence-like images. The gallery loops automatically and pauses when you inspect it.
Five traceable stages connect the safety question to sample-level evidence and human explanations.
Organize images by evidentiary function and safety consequence.
Specify content, media forms, and risk focus for every category.
Build detailed 5W+1H scenarios with category-specific constraints.
Generate realistic evidence-like images with GPT Image 2.
Locate and explain implausible evidence.
Detection accuracy (%) on SafeIMG. Sort the original paper data by any category to expose capability gaps.
Strong visual understanding does not automatically translate into reliable synthetic-image detection.
Best VLM overall: 49.5%AI-generated Images Challenge Visual Trust in High-risk Scenarios
@misc{wang2026aigeneratedimageschallengevisual,
title={AI-generated Images Challenge Visual Trust in High-risk Scenarios},
author={Yi-Zhi Wang and Yichen Xiao and Linan Yue and Weibo Gao and Yichao Du and Pengfei Fang and Shimin Di and Min-Ling Zhang},
year={2026},
eprint={2607.22745},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2607.22745},
}