Our numbers, measured.
Every number on this page is a result we measured for TrueEye 2.1. Each set is scored at a decision point of p = 0.5.
How accurate is Perspicax?
In testing, Perspicax caught up to 99% of AI images, depending on the model: 92.1–93.9% of ChatGPT images, 93.6–98.3% of Gemini images and 91.0–98.9% of Seedream images. It called 99.4% of original camera photos real, so fewer than 1 in 100 were wrongly flagged. Every number on this page comes from our own result files.
Share of AI images caught, newest models
Each glass column is 100% of a model's images; the colour fill is the share caught. The table below has the exact numbers.
- OpenAI
- ByteDance
- Other labs
Images from the apps people use
Ranges across each app's newest models.
- 92.1–93.9%of ChatGPT images caught in testingGPT Image 2.5 Sunburst, GPT Image 2.5 Flare, GPT Image 2ChatGPT images
- 93.6–98.3%of Gemini images caught in testingImagen 4, Nano Banana 2.1, Nano Banana ProGemini images
- 91.0–98.9%of Seedream images caught in testingSeedream 4.5, Seedream 5.0 Pro, Seedream 5.0 LiteSeedream images
Public test sets
Scored in each set's official protocol.
- 99.4%Camera photosOriginal camera photos correctly called real.
- 98.0%AIGenImages2026A 2026 benchmark of real and AI images. AUC 99.7.
- 97.8%ImageTimeAI images caught.
- 97.1%OpenFake, RedditReal and AI images posted on Reddit.
- 95.6%So-Fake-OODReal and AI images from social media.
- 95.1%OpenFakeIts core set of real and AI still images.
- 93.7%DailyBenchSeven current generators.
- 93.5%RealChainImages passed along real sharing chains.
- 91.8%Never-seen generators17 generators the model never saw in training, 31,334 images.
In the wild
88.3% of GPT Image 2 images posted on X were caught, after the platform had stripped their Content Credentials (n = 10,072, self-reported labels).
The newest models
About 1,500 images per model, collected after release and mostly re-saved as 1 MP WebP.
| Model | Made by | Ships in | Caught |
|---|---|---|---|
| GPT Image 2.5 Sunburst | OpenAI | ChatGPT | 93.9% |
| GPT Image 2.5 Flare | OpenAI | ChatGPT | 92.6% |
| GPT Image 2 | OpenAI | ChatGPT | 92.1% |
| Imagen 4 | Gemini | 98.3% | |
| Nano Banana 2.1 | Gemini | 95.0% | |
| Nano Banana Pro | Gemini | 93.6% | |
| Seedream 4.5 | ByteDance | Doubao / CapCut | 98.9% |
| Seedream 5.0 Pro | ByteDance | Doubao / CapCut | 96.7% |
| Seedream 5.0 Lite | ByteDance | Doubao / CapCut | 91.0% |
| Hunyuan 3.0 | Other labs | – | 99.9% |
| HiDream-O1 | Other labs | – | 99.8% |
| MAI-Image 2.5 | Other labs | – | 99.2% |
| Qwen-Image 3.0 | Other labs | – | 99.1% |
| FLUX 3 | Other labs | – | 96.6% |
| Grok Imagine 2.0 | Other labs | – | 96.2% |
How we measure
- Model: TrueEye 2.1, the detector behind every Perspicax scan. See the model card.
- A decision point of p = 0.5 unless stated. Results near it can come back inconclusive in the product.
- Real photos: about 99.4% of original camera photos are called real.
- We publish our own results only.
Before you believe it, check it.
Your first scan is free, no account needed. A free account gives you 2 scans every day.