TextFake: leaderboard
TextFake (2026) asks whether a text-rich image, such as a screenshot, document or news page in one of 28 languages, is real or AI-generated; real images are paired with fakes from four text-rendering image generators and every image is re-encoded as JPEG. The board is overall accuracy in percent on the balanced 20,000-image set with one fixed zero-shot REAL or FAKE prompt; chance is 50 and higher is better.
Metric: Overall accuracy (%) over all 20,000 images at classifying text-rich images (screenshots, documents and news pages in 28 languages) as real or AI-generated; the balanced set pairs real images with fakes from four text-rendering image generators, every image JPEG-re-encoded; VLM APIs answer REAL or FAKE to one fixed zero-shot prompt; chance 50; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 3 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro (Preview) | 77.9 | #113 |
| 2 | Claude Sonnet 4.6 | 74.3 | #161 |
| 3 | GPT-5.4 | 67.6 | #123 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=textfake · How It Works · Data refreshed daily, snapshot 2026-11-13.