Back Home

AI 安全

Two Studies Examine Image Safety: Refusal Rates Don’t Capture the Full Risk

Combinations of text and symbols can change image-generation results, while the risk in an input image can affect editing decisions. But returning an image does not mean harmful content was successfully generated; the two studies must be interpreted separately.

zh-Hant

For developers integrating image-generation and image-editing tools, a safety review that records only “whether the model refused” may miss the most important question: what does the finished output communicate? Two studies submitted to arXiv on November 19, 2026, separately tested text combined with affirmative symbols and risk-bearing images paired with different editing instructions. Both observed a gap between input form and safety behavior, but they did not measure the same thing. Study 1, Study 2

Adding an affirmative symbol to a question restores the original meaning

The first paper, by Haoyang Xu and colleagues, tested generated images containing text, such as signs. It asked whether a system would still refuse if a harmful claim were recast as a question and paired with an affirmative element such as a check mark. The authors call this structure “deferred semantic commitment”: a question first removes endorsement, then another element restores an affirmative meaning. Paper abstract

In the authors’ 300-case challenge set, gpt-image-2 returned images for 16.3% of direct statements and 43.3% of questions paired with affirmative symbols. Using all cases as the denominator, the share that produced images the authors judged to endorse the original claim was 15.3% and 41.3%, respectively. The test counted not only whether an image was generated, but also whether the output preserved the original meaning. Results and study design

One limitation is that all 300 cases were selected: direct statements did not yield images in the initial test, though some succeeded when resubmitted. These percentages therefore describe a specific challenge set and should not be treated as an overall bypass rate for ordinary requests. The authors also state that they could not inspect the commercial system’s internal safety processes. Test scope

If the input image is risky, should a neutral instruction be allowed?

The second paper, by Ziqing Yang and colleagues, introduces EditSafetyBench. It pairs 660 images with nine fixed instructions each, producing 5,940 test cases across 11 risk categories. The instructions cover neutral editing, risk reduction, and harmful intent, allowing the researchers to compare how safe and unsafe input images are handled with identical text. Dataset design

The authors report that Gemini and GPT refused harmful-intent instructions at rates of 87.1% and 81.6%, respectively. For “risky image plus neutral instruction,” the rates were 24.2% and 22.4%. The test used the commercial APIs gemini-2.5-flash-image and gpt-image-1-mini with their default settings; it did not use the first study’s gpt-image-2. Abstract, Experimental setup

These results cannot be directly interpreted as a success rate for generating harmful images. The second study recorded only whether a request was allowed, refused, or resulted in an error; it did not assess the harmfulness of the output. An allowed request may still produce a safe result. The authors also acknowledge that whether editing an existing risky image by brightening or cropping it should be refused depends on the use case and platform policy; there is no answer that applies to every situation. Evaluation scope and discussion

Review the finished output and the purpose of the edit

Both papers are original test reports from research teams, not defect announcements confirmed by model providers; this article has not reproduced the experiments. They support the conclusion that how inputs are combined can affect observed safety behavior. They do not establish that all tools review text and images separately. Study 1, Study 2

For integrators, this article infers that a more useful review should consider the input image, the purpose of the edit, and the finished output, rather than target a single refusal rate. Obscuring sensitive content may reduce harm; generating a sign that preserves harmful endorsement is a different outcome. The two call for different criteria.

Sources

  1. Question, Then Affirm: Restoring Commitment Does Not Restore Refusal
  2. Question, Then Affirm: Restoring Commitment Does Not Restore Refusal
  3. Text-Aware, Vision-Blind: How Multimodal Image Editing Systems Handle Unsafe Inputs
  4. Text-Aware, Vision-Blind: How Multimodal Image Editing Systems Handle Unsafe Inputs