Small Symbols, Big Risks: Exploring Emoticon Semantic Confusion in Large Language Models

TL;DR

Emoticon semantic confusion reveals that ASCII emoticons can be misgrounded as code or commands by LLMs, enabling dangerous consequences. The authors create a generate-and-fill pipeline to produce 3,757 test cases across 21 meta-scenarios and four programming languages, and evaluate six LLMs. They measure Confusion Ratio and Confusion Impact Ratio, finding an average CR of 38.6% and a majority of Level-2 executable misinterpretations, with many classified as High Harm. They demonstrate transfer to agent frameworks and show that prompt engineering alone does not reliably mitigate the risk, highlighting a pressing need for defense mechanisms and safer human-AI interaction designs.

Abstract

Emoticons are widely used in digital communication to convey affective intent, yet their safety implications for Large Language Models (LLMs) remain largely unexplored. In this paper, we identify emoticon semantic confusion, a vulnerability where LLMs misinterpret ASCII-based emoticons to perform unintended and even destructive actions. To systematically study this phenomenon, we develop an automated data generation pipeline and construct a dataset containing 3,757 code-oriented test cases spanning 21 meta-scenarios, four programming languages, and varying contextual complexities. Our study on six LLMs reveals that emoticon semantic confusion is pervasive, with an average confusion ratio exceeding 38%. More critically, over 90% of confused responses yield 'silent failures', which are syntactically valid outputs but deviate from user intent, potentially leading to destructive security consequences. Furthermore, we observe that this vulnerability readily transfers to popular agent frameworks, while existing prompt-based mitigations remain largely ineffective. We call on the community to recognize this emerging vulnerability and develop effective mitigation methods to uphold the safety and reliability of the LLM system.

Figures (8)

• Figure 1: LLM may consider emoticons that humans use to express emotions as part of the instructions, ultimately leading to catastrophic consequences (such as deletion of critical data)
• Figure 2: The overview of the data generation pipeline.
• Figure 3: The distribution of Confusion Rate (CR) across different models.
• Figure 4: The distribution of confusion cases for different contextual complexity levels across different LLMs.
• Figure 5: The distribution of CIR across different LLMs.