acceptodds
Withdrawn from ICLR 2027

Semantic Underspecification as a Safety Risk in Text-to-Image Generation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Existing text-to-image (T2I) safety research primarily focuses on harmful semantics that are explicitly expressed in prompts, deliberately encoded through adversarial manipulation, or already manifested in generated images. However, in everyday T2I use, ordinary users may unintentionally omit contextual information that appears irrelevant to their immediate description but is critical for safety. Such seemingly benign prompts can remain compatible with both benign and unsafe interpretations, allowing a T2I model to instantiate an unsafe scene despite the absence of any explicit harmful request or malicious intent. We identify and formulate this previously underexplored pre-generation safety risk as semantic underspecification, and construct a curated benchmark of 100 prompts across five representative safety risk categories to systematically study it. To mitigate this risk, we propose Risk-aware Semantic Completion (RSC), an iterative prompt repair framework that resolves safety-critical semantic gaps before image generation. RSC first interprets plausible unsafe readings of the current prompt and identifies the missing semantics that enable them, then introduces targeted benign context to disambiguate the scene. A semantic preservation check subsequently verifies whether the repaired prompt retains the user's explicitly specified content; failed candidates are returned for further repair. By combining risk interpretation, targeted completion, and preservation-aware repair, RSC suppresses unsafe interpretations without simply removing risky objects, altering user-specified actions, or broadly rewriting the original scene. On the benchmark, RSC achieves a 92.0% Risk Mitigation Rate, a 92.0% Semantic Preservation Rate.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Text-to-Image Generation, Semantic Underspecification, AI Safety

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_jgs9OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper studies semantic underspecification as a safety risk in text-to-image (T2I) generation, where seemingly benign prompts omit safety-critical contextual information and may consequently admit unsafe interpretations. It proposes Risk-aware Semantic Completion (RSC), an LLM-based preprocessing framework comprising risk interpretation, risk-aware completion, and semantic preservation check. The method aims to introduce benign contextual information while preserving explicitly specified prompt semantics. Experiments on a curated benchmark of 100 prompts across five risk categories report high risk mitigation rate, semantic preservation rate, and joint success rate to show the effectiveness of RSC.

Critical strengths
  1. A Meaningful Research Topic. The paper investigates a relevant safety problem in modern T2I models.
Critical weaknesses
  1. Limited Novelty. Similar ideas have been explored in prior studies on safety-aware prompt rewriting and intention disambiguation. The novelty of the proposed RSC framework is not sufficiently established.

  2. Insufficient Experimental Evidence. The experimental design does not adequately account for the stochastic nature of T2I generation, raising concerns about the validity of the reported improvements.

Questions for authors
  1. My major concern is the novelty of this work. How RSC's RI–RC–SPC pipeline differs technically from VALOR's [1] risk-aware rewriting? What's the new methodological contribution RSC offers? [1] Zhao et al., Value-Aligned Prompt Moderation via Zero-Shot Agentic Rewriting for Safe Image Generation (AAAI 2026).

  2. The benchmark contains only 100 risk-positive prompts constructed by weakening explicitly harmful prompts. How does the framework distinguish genuinely safety-critical underspecification from ordinary benign ambiguity? What evidence supports its ability to avoid unnecessary modifications to benign prompts?

  3. How is "partially preserved" defined, and why is it counted as successful semantic preservation? Since only 46% of cases are classified as fully preserved, how would the conclusions change under a stricter preservation criterion?

Reviewer_kqsDOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies safety risks caused by missing context in text-to-image prompts. It proposes Risk-aware Semantic Completion, which identifies plausible unsafe interpretations, adds benign context, and checks whether the revised prompt preserves explicitly specified content. Experiments examine whether these repairs produce safer images while retaining the requested scene.

Critical strengths
  1. The paper identifies a concrete challenge: seemingly ordinary descriptions can leave purpose, authorization, or relationships unspecified, allowing different interpretations.
  2. The proposed prompt rewriting framework is straightforward to integrate with existing generators. Experiments across different repair models and image generators provide an initial assessment of its generality.
Critical weaknesses
  1. The main evaluation criteria are unclear. The formulas count cases labeled “risk eliminated” or “risk reduced,” but the paper does not explain how these labels are assigned. The full Grok evaluator prompt and scoring rubric are missing, and there is no human validation. It is therefore difficult to tell what qualifies as meaningful mitigation, as well as semantic preservation.
  2. The benchmark construction pipeline lacks details. The paper outlines generating harmful prompts, weakening them, and manually filtering candidates, but omits the construction LLM, instructions, review protocol, etc.
  3. The contribution beyond existing prompt rewriting is unclear. Prior methods such as [1] already rewrite prompts for safer generation while preserving semantic content or user intent. The paper neither adequately discusses nor compares against these methods.

[1] Wu, Zongyu, et al. "Universal prompt optimizer for safe text-to-image generation." Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024.

Questions for authors

Please see weakness part.

Reviewer_PkU6OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies semantic underspecification in text-to-image (T2I) generation, where seemingly benign prompts may lead to unsafe images because they lack safety-critical context. The authors propose Risk-aware Semantic Completion (RSC), a prompt-rewriting framework that identifies potentially unsafe interpretations, adds benign context to resolve ambiguity, and checks whether the original prompt semantics are preserved. The authors also introduce a benchmark of 100 prompts across five safety risk categories.

Critical strengths
  • I think the paper does a good job of explaining the problem of semantic underspecification in T2I generation. The authors also introduce a benchmark covering five safety risk categories, which could be useful for further research on this problem.
Critical weaknesses
  • My main concern is the limited novelty of the proposed method. RSC identifies unsafe interpretations, adds benign context, and checks semantic preservation, but these steps seem relatively straightforward. I find it difficult to see what is fundamentally new in this framework, which significantly weakens the paper's contribution.

  • The paper mainly compares RSC against safety detection methods rather than existing prompt-rewriting approaches. For example, VALOR [1] also addresses ambiguous intent and safety-aware prompt rewriting. Without such comparisons, it is difficult to assess whether RSC offers meaningful improvements over existing methods.

[1] Value-Aligned Prompt Moderation via Zero-Shot Agentic Rewriting for Safe Image Generation. AAAI 2026.

Questions for authors
  • Some examples in figure3 seem ambiguous, especially in terms of whether the original images are actually unsafe. Did the authors conduct any human evaluation to validate these results?

  • Could the authors compare RSC with existing prompt-rewriting methods rather than only safety detection baselines?

  • How does RSC perform on benign prompts without safety risks? Since the benchmark only includes prompts with potential risks, could RSC unnecessarily modify benign prompts and affect image generation quality?

  • The paper introduces a new benchmark, but its quality is not sufficiently evaluated. Could the authors provide additional validation, such as human evaluation, to show that the collected prompts actually represent meaningful safety risks?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.