When Smiley Turns Hostile: Interpreting How Emojis Trigger LLMs' Toxicity
Fuente:
arXiv
Enregistré dans:
| Auteurs principaux: | , , , , , , , , |
|---|---|
| Format: | Preprint |
| Publié: |
2025
|
| Sujets: | |
| Accès en ligne: | |
| Tags: |
Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
|
| _version_ | 1866911153390616576 |
|---|---|
| author | Cui, Shiyao Feng, Xijia Wang, Yingkang Yang, Junxiao Zhang, Zhexin Sikdar, Biplab Wang, Hongning Qiu, Han Huang, Minlie |
| author_facet | Cui, Shiyao Feng, Xijia Wang, Yingkang Yang, Junxiao Zhang, Zhexin Sikdar, Biplab Wang, Hongning Qiu, Han Huang, Minlie |
| contents | Emojis are globally used non-verbal cues in digital communication, and extensive research has examined how large language models (LLMs) understand and utilize emojis across contexts. While usually associated with friendliness or playfulness, it is observed that emojis may trigger toxic content generation in LLMs. Motivated by such a observation, we aim to investigate: (1) whether emojis can clearly enhance the toxicity generation in LLMs and (2) how to interpret this phenomenon. We begin with a comprehensive exploration of emoji-triggered LLM toxicity generation by automating the construction of prompts with emojis to subtly express toxic intent. Experiments across 5 mainstream languages on 7 famous LLMs along with jailbreak tasks demonstrate that prompts with emojis could easily induce toxicity generation. To understand this phenomenon, we conduct model-level interpretations spanning semantic cognition, sequence generation and tokenization, suggesting that emojis can act as a heterogeneous semantic channel to bypass the safety mechanisms. To pursue deeper insights, we further probe the pre-training corpus and uncover potential correlation between the emoji-related data polution with the toxicity generation behaviors. Supplementary materials provide our implementation code and data. (Warning: This paper contains potentially sensitive contents) |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2509_11141 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | When Smiley Turns Hostile: Interpreting How Emojis Trigger LLMs' Toxicity Cui, Shiyao Feng, Xijia Wang, Yingkang Yang, Junxiao Zhang, Zhexin Sikdar, Biplab Wang, Hongning Qiu, Han Huang, Minlie Computation and Language Emojis are globally used non-verbal cues in digital communication, and extensive research has examined how large language models (LLMs) understand and utilize emojis across contexts. While usually associated with friendliness or playfulness, it is observed that emojis may trigger toxic content generation in LLMs. Motivated by such a observation, we aim to investigate: (1) whether emojis can clearly enhance the toxicity generation in LLMs and (2) how to interpret this phenomenon. We begin with a comprehensive exploration of emoji-triggered LLM toxicity generation by automating the construction of prompts with emojis to subtly express toxic intent. Experiments across 5 mainstream languages on 7 famous LLMs along with jailbreak tasks demonstrate that prompts with emojis could easily induce toxicity generation. To understand this phenomenon, we conduct model-level interpretations spanning semantic cognition, sequence generation and tokenization, suggesting that emojis can act as a heterogeneous semantic channel to bypass the safety mechanisms. To pursue deeper insights, we further probe the pre-training corpus and uncover potential correlation between the emoji-related data polution with the toxicity generation behaviors. Supplementary materials provide our implementation code and data. (Warning: This paper contains potentially sensitive contents) |
| title | When Smiley Turns Hostile: Interpreting How Emojis Trigger LLMs' Toxicity |
| topic | Computation and Language |
| url | https://arxiv.org/abs/2509.11141 |