AI-generated images fail to fool humans in safety-critical scenarios

A new benchmark called SafeIMG has been developed to assess the reliability of AI-generated images in high-risk scenarios, such as those impacting public safety and personal reputation. Researchers found that current specialized detectors and vision-language models (VLMs) are not effective at identifying these synthetic images, with the best VLM detecting only 49.5% and the top detector identifying 33.1%. Human evaluators performed significantly better, achieving 81.7% accuracy, and were able to identify local artifacts, commonsense conflicts, and physical inconsistencies that current AI models struggle to detect or explain. AI

IMPACT Current AI image detection methods are insufficient for high-risk applications, necessitating further research into robust and explainable detection systems.

RANK_REASON The cluster contains a research paper introducing a new benchmark for evaluating AI-generated images. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• GPT Image 2
• SafeIMG
• vision-language models

• paper
• safety
• product

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →