Image editors refuse unsafe text far more than unsafe images
When a user sends an image that is not safe with a neutral instruction, Gemini and GPT refuse less than 1 time in 4.
Researchers made EditSafetyBench, a test of how well image editing systems see risk. It has 5,940 pairs of an image and an instruction, in 11 risk groups and 3 edit aims. Gemini and GPT refuse 87.1% and 81.6% of unsafe prompts, but only 24.2% and 22.4% of unsafe images with neutral instructions. The team also tested Qwen API and Qwen Local.
How it works
The test puts safe and unsafe images with instructions, and it counts how many times each system refuses. The authors name the gap a text and vision mismatch. Most systems look at the prompt text, not at the image.
The numbers
- Gemini refused 87.1% of unsafe prompts. GPT refused 81.6%.
- With unsafe images and neutral text, they refused 24.2% and 22.4%.
- GPT permits 70.4% of unsafe political edits. Gemini permits 34.8%.
Sexual content shows the strongest visual sensitivity. The authors think that earlier work on detection of such images helped.
What is not known
The source gives no refusal rates for the 2 Qwen systems. It gives no rates for the other risk groups. The authors tell that guardrails must read the image and the text together.
Insight: Safety checks that read only text leave the image open
In 2026, researchers hid a text command in an image and got 58.75% success on 4 closed models that read images. Here, image editors refuse far fewer unsafe images than unsafe prompts. Both results read as one gap: checks look at the text, but the risk can be in the image. A vendor that ships a check for both gets trust. Leaders must test the image path.
Sources
Posted