A Multi-Domain Benchmark for Detecting AI-Generated Text-Rich Images from GPT-Image-2

A new benchmark for detecting AI-generated text-rich images from OpenAI’s GPT Image 2 reveals that existing detectors struggle with structured visual content, according to a paper submitted to arXiv in 2026 [1]. The dataset spans 8,602 images across six categories where text and layout are central [1]. The benchmark, described in a June 2026 preprint, targets a gap in current evaluation methods. Most existing benchmarks focus on object-centric images and provide limited coverage of scenarios where textual semantics and layout organization are central [2]. The new collection includes commercial posters, infographics, academic posters, receipts, tables, and UI screenshots [1]. Text-rich images often contain privacy-sensitive, transactional, or decision-relevant information, making reliable detection important for digital trust and content authenticity [2]. Researchers evaluated five representative AI-generated image detectors in a zero-shot setting [1]. Performance proved highly domain-dependent: methods that performed well in some categories often failed on others [1]. Even the strongest conventional detector showed severe sensitivity to JPEG compression [1]. The team also conducted an exploratory evaluation with a multimodal vision-language model, which showed promise but exhibited limitations on structured formats [1]. The findings arrive as generative image models have become widely available. Since the 2020s, generative AI has enabled the creation and modification of images, audio, and video from text prompts [5]. OpenAI’s GPT Image series, introduced in March 2025 as a successor to DALL-E, is native to ChatGPT and available through an API [9]. GPT Image 1, based on the GPT-4o architecture, replaced DALL-E 3 in ChatGPT that same month [8]. The models can generate images in specific artistic styles and edit existing visuals with precision [9]. A separate 2026 study underscored the detection challenge. Researchers released AIForge-Doc v2, a paired dataset of 3,066 document forgeries made with GPT-Image-2 [7]. Human inspectors, tested through a public two-alternative forced-choice site, achieved an accuracy of 0.501 — indistinguishable from chance — when asked to identify AI-edited receipts [7]. Three computational judges performed only modestly better, with a document-specific detector reaching 0.585 and a generic forensic tool reaching 0.599 [7]. The same GPT-Image-2 model used as a zero-shot self-judge scored 0.532, and its performance did not improve across five prompt strategies [7]. AI content watermarking has emerged as one approach to address synthetic-media traceability. Modern watermarking schemes embed imperceptible signals into AI-generated content, evaluated along three axes: quality, detectability, and robustness against modifications [6]. The benchmark authors argue that text- and layout-aware detection methods are needed for modern AI-generated images [1]. The dataset has been released for further research [2].