Digital Skin, Digital Bias: Uncovering Tone-Based Biases in LLMs and Emoji Embeddings

Mingchen Li; Wajdi Aljedaani; Yingjie Liu; Navyasri Meka; Xuan Lu; Xinyue Ye; Junhua Ding; Yunhe Feng

cs.SIApr 8, 2026

Digital Skin, Digital Bias: Uncovering Tone-Based Biases in LLMs and Emoji Embeddings

Authors: Mingchen Li, Wajdi Aljedaani, Yingjie Liu, Navyasri Meka, Xuan Lu, Xinyue Ye, Junhua Ding, Yunhe Feng

Organizations: University of North Texas Denton, TX, United States · International Center for Artificial Intelligence Research and Ethics Riyadh, Saudi Arabia · Southern Methodist University Dallas, TX, United States · The University of Arizona Tucson, AZ, United States ·

Abstract

Skin-toned emojis are crucial for fostering personal identity and social inclusion in online communication. As AI models, particularly Large Language Models (LLMs), increasingly mediate interactions on web platforms, the risk that these systems perpetuate societal biases through their representation of such symbols is a significant concern. This paper presents the first large-scale comparative study of bias in skin-toned emoji representations across two distinct model classes. We systematically evaluate dedicated emoji embedding models (emoji2vec, emoji-sw2v) against four modern LLMs (Llama, Gemma, Qwen, and Mistral). Our analysis first reveals a critical performance gap: while LLMs demonstrate robust support for skin tone modifiers, widely-used specialized emoji models exhibit severe deficiencies. More importantly, a multi-faceted investigation into semantic consistency, representational similarity, sentiment polarity, and core biases uncovers systemic disparities. We find evidence of skewed sentiment and inconsistent meanings associated with emojis across different skin tones, highlighting latent biases within these foundational models. Our findings underscore the urgent need for developers and platforms to audit and mitigate these representational harms, ensuring that AI's role on the web promotes genuine equity rather than reinforcing societal biases.

Figures & tables

Emoji	Unicode
	U+270A
	U+1F44F
	U+1F468
	U+1F3C7

Table 1. Examples of skin-toned emojis

Pre-trained Model	# Tokens	# Skin-toned Emojis
FastText-crawl-300d-2M ( bojanowski2016enriching )	1,999,995	5 (642)
GloVe-Twitter-27B ( pennington2014glove )	1,193,514	0 (139)
FastText-wiki-news ( bojanowski2016enriching )	999,994	1 (147)
emoji-word2vec ( AdiShirsath2021 )	23,069	5 (930)
emoji-sw2v ( barbieri2018gender )	2269	834 (2269)
emojional ( barry2021emojional )	1816	0 (1816)

Table 2. Support for skin-toned emojis in pre-trained embedding models. The table lists the total vocabulary size (# Tokens) for each model and the number of supported skin-toned emojis, with the total number of supported emojis shown in parentheses.

Model	Avg. Tokens/ Emoji	Token Range	Most Common Count
Gemma-2-2B ( gemma_2024 )	3.96	1–9	3 (893 emojis)
Qwen-2-1.5B ( qwen2 )	5.97	1–13	6 (635 emojis)
Llama-3.2-1B/3B ( meta_llama3.2_1b_instruct ; meta_llama3.2_3b_instruct )	10.05	3–19	10 (550 emojis)
Mistral-v0.3-7B ( mistral7b_instruct_v0.3 )	$\sim$ 9.5	2–26	6 (490 emojis)

Table 3. Tokenization statistics for skin-toned emojis across LLMs. Columns show each model’s average tokens per emoji (Avg. Tokens/Emoji), minimum to maximum token range (Token Range), and the most frequent token count (mode), with the number of emojis at that count in parentheses.

Model
Gemma-2	1	1	1	1	1
Qwen-2	1	1	1	1	1
Llama-3	3	3	3	3	3
Mistral-v0.7	2	2	2	2	5

Table 4. Number of tokens for each of the five skin tone modifiers across different LLMs. The inconsistency in Mistral’s tokenization for the dark skin tone is highlighted.

Model	Default
Aggregated Representation Distance (Cosine Distance)
Gemma-2-2B	0.819	0.906	0.864	0.860	0.856	0.912
Gemma-2-9B	0.612	0.718	0.794	0.746	0.778	0.695
Llama-3.2-1B	0.740	0.662	0.653	0.668	0.641	0.694
Llama-3.2-3B	0.721	0.644	0.661	0.643	0.651	0.639
Llama-3.1-8B	0.701	0.568	0.577	0.570	0.588	0.578

Table 5. Semantic consistency between emojis and their textual descriptions across different LLMs and skin tones. All cell values represent the mean score for the given metric. For Cosine Distance and WMD, higher values indicate more semantic shift between the cross-modal representations.

Panel of Figure 1 — Figure 1. t-SNE visualizations (top row) and similarity shift heatmaps (bottom row) for hand gesture emoji embeddings across six models. Columns correspond to the same model as the caption shown. The t-SNE plots qualitatively illustrate clustering patterns of emoji variants, while the heatmaps depict pairwise cosine-distance shifts between skin-tone variants (higher values indicate larger semantic shifts).

Panel of Figure 2 — Figure 2. Heatmaps illustrating the pairwise Relative Norm Distance (RND) scores for the hand gestures emoji subdomain across different embedding models. Each cell represents the RND score between two skin tone groups, where color intensity indicates the magnitude of the bias relative to a set of neutral words.

Tone Pair (X vs. Y)	Emoji2vec	Gemma Gemma-2-2B	Llama Llama-3.2-3B	Mistral Mistral-v0.3-7B	Qwen Qwen-2-1.5B
n (roles)	7	82	82	82	82
default vs	-0.286	-0.537	-1.415	-0.634	-0.829
default vs	-2.000	-0.780	-1.268	-1.512	-0.683
default vs	-1.429	-1.073	-1.122	-1.707	+1.220
default vs	-0.286	+0.244	-1.024	-1.561	+0.439
default vs	+0.286	+1.073	-0.537	-1.902	+1.220

Table 6. WEAT score across person roles. Cells report the mean effect size

d

per tone pair after averaging across all roles for each model. The row n (roles) lists the number of available roles per model. For a tone pair X vs Y , positive values mean X is more associated with the “Good” emoji set (and Y with “Bad”); negative values mean X is more associated with “Bad” (and Y with “Good”).

Model	default						Avg KL
emoji2vec	0.418	0.167	0.160	0.165	0.169	0.170	0.0004
Gemma	0.163	0.164	0.163	0.164	0.170	0.175	0.0057
Llama	0.219	0.126	0.152	0.152	0.115	0.236	0.0821
Mistral	0.127	0.155	0.177	0.157	0.180	0.204	0.0308
Qwen	0.357	0.071	0.061	0.116	0.099	0.296	0.5310

Table 7. RNSB across person roles. Cells report across-role average negative share per tone; Avg KL summarizes average KL divergence (higher indicates more uneven distributions).

Tone Pair (X vs. Y)	Emoji2vec	Gemma Gemma-2-2B	Llama Llama-3.2-3B	Mistral Mistral-v0.3-7B	Qwen Qwen-2-1.5B
default vs	-0.114	0.188	0.196	0.066	0.148
default vs	-0.004	0.121	-0.080	0.253	0.143
default vs	0.399	0.191	0.045	0.128	0.159
default vs	0.038	0.096	0.055	0.119	0.116
default vs	0.001	-0.071	-0.216	0.233	0.239
vs	-0.054	-0.024	-0.512	0.455	0.090

Table 8. WEAT results for the Caliskan et al. benchmark ( caliskan2017semantics ) , treating skin tones as targets. Cells report the mean effect size for each tone pair, averaged across all standard benchmarks for a given model. For a tone pair X vs Y , a positive value indicates that skin tone X is more strongly associated with culturally positive concepts. A negative value indicates that X is more associated with the opponent concepts.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 3. A boxplot comparing the distribution of the number of tokens per skin-toned emoji across different LLMs. Each box shows the median, quartiles, and range, illustrating the variance in tokenization strategies.

Attribute A	Sample A	Attribute B	Sample B
flowers	aster, clover	insects	ant, caterpillar
instruments	bagpipe, cello	weapons	arrow, club
european american names 7	Brad, Brendan	african american names 7	Darnell, Hakim
european american names 5	Adam, Harry	african american names 5	Alonzo, Jamel
male names	John, Paul	female names	Amy, Joan
math	math, algebra	arts	poetry, art

Appendix

Table 9. WEFE attribute pairs with two samples per attribute.

Tone Pair	emoji2vec	Gemma	Llama	Mistral	Qwen
default vs	47%	7%	7%	13%	33%
default vs	27%	27%	13%	20%	33%
default vs	47%	13%	0%	27%	27%
default vs	40%	20%	0%	13%	27%
default vs	53%	20%	20%	7%	27%
vs	27%	20%	40%	60%	13%

Appendix

Table 10. Significance rate (% of benchmarks with

p<0.05

) per tone pair and model.

Demographic Bias in Language Models

Demographic bias in language models studies how identity cues shape model outputs, decisions, and representations, and how to detect and reduce resulting disparities.

See all 135 papers

Are LLMs Safe Beyond Text: Do Emojis Expose Gaps in Safety Evaluation
Aug 15, 2026M P V S GopinadhLLM Safety Adversarial Robustness of Language Models
Independent Researcher
Dual-Metric Evaluation of Social Bias in Large Language Models: Evidence from an Underrepresented Nepali Cultural Context
Mar 8, 2026Ashish Pandey, Tek Raj ChhetriSocial Bias in Language Models Cultural Bias in Language Models
Center for Artificial Intelligence Research Nepal, Sundarharaincha-09, Buddhabare, Nepal
Tracing the Latent Threads: A Mechanistic Study of How LLMs Represent and Operationalize Race and Ethnicity Cues
Jan 19, 2026Shiyue Hu, Ruizhe Li, Yanjun GaoLLM Interpretability Mechanistic Interpretability
University of Colorado Anschutz · University of Colorado Boulder · University of Aberdeen

Aug 15, 2026J/K move · Enter open · S save

Are LLMs Safe Beyond Text: Do Emojis Expose Gaps in Safety Evaluation

M P V S Gopinadh

Independent Researcher

LLM Safety Adversarial Robustness of Language Models LLM Safety Evaluation Adversarial Attacks on LLMs

Open paper AskPDF arXiv

Safety evaluations of large language models (LLMs) predominantly rely on text-based adversarial prompts, potentially overlooking vulnerabilities arising from alternative input representations. This work examines emoji-augmented prompts as a test case for this gap, evaluating 50 prompts across four open-source LLMs (Mistral 7B, Qwen 2 7B, Gemma 2 9B, Llama 3 8B). Results show substantial variation in robustness: Gemma 2 9B and Mistral 7B exhibit non-zero success rates (10%), Llama 3 8B 6%, while Qwen 2 7B shows complete resistance (0% success rate). A chi-square test (

χ^2 = 32.94, p < 0.001

) confirms significant differences in outcome distributions. These findings indicate that robustness is sensitive to input representation, and that evaluations restricted to standard text prompts may underrepresent model vulnerabilities.

Abstract

Demographic Bias in Language Models

Explore similar work

Are LLMs Safe Beyond Text: Do Emojis Expose Gaps in Safety Evaluation

Dual-Metric Evaluation of Social Bias in Large Language Models: Evidence from an Underrepresented Nepali Cultural Context

Tracing the Latent Threads: A Mechanistic Study of How LLMs Represent and Operationalize Race and Ethnicity Cues