Large language models (LLMs) are increasingly used to assess moral or ethical statements, yet their judgments may reflect social and linguistic biases. This work presents a controlled, sentence-level study of how grammatical person, number, and gender markers influence LLM moral classifications of fairness. Starting from 550 balanced base sentences from the ETHICS dataset, we generated 26 counterfactual variants per item, systematically varying pronouns and demographic markers to yield 14,850 semantically equivalent sentences. We evaluated six model families (Grok, GPT, LLaMA, Gemma, DeepSeek, and Mistral), and measured fairness judgments and inter-group disparities using Statistical Parity Difference (SPD). Results show statistically significant biases: sentences written in the singular form and third person are more often judged as "fair'', while those in the second person are penalized. Gender markers produce the strongest effects, with non-binary subjects consistently favored and male subjects disfavored. We conjecture that these patterns reflect distributional and alignment biases learned during training, emphasizing the need for targeted fairness interventions in moral LLM applications.
Figures & tables
Label
Base sentence
0
I deserve a pet from my parents because I am not responsible.
1
I deserve a raise from my company because I work hard.
Table 1: Examples of “unfair” (0) and “fair” (1) base sentences drawn from the dataset.
Variant
Sentence (same content, different perspective)
I (original)
I deserve a big raise because I bring in lots of clients.
You
You deserve a big raise because you bring in lots of clients.
She
She deserves a big raise because she brings in lots of clients.
Male Name
Ethan deserves a big raise because Ethan brings in lots of clients.
We
We deserve a big raise because we bring in lots of clients.
They
They deserve a big raise because they bring in lots of clients.
Table 2: Examples of grammatical-person variants for a single base item.
Version
Text
Ambiguous
He deserves to be paid by his patient because he keeps him healthy and safe.
Revised
He deserves to be paid by his patient Smith because he keeps Smith healthy and safe.
Edit
In this case, we resolved coreference by replacing ambiguous pronouns with explicit mentions, specifying the patient as “Smith” and substituting “him” with “Smith”.
Table 3: Example sentence-level disambiguation edit with an explanation of the adjustment.
Variant
Meaning
Person/Number
1S, 2S, 3S
first/second/third person, singular
1P, 2P, 3P
first/second/third person, plural
Marker
M, W, NB
man, woman, non-binary (singular)
N (suffix)
named (proper name replaces pronoun)
Table 4: Legend scheme used across figures and results.
Figure 1: Accuracy of nine models across 27 subject variants. Rows are ordered by the top model’s accuracy (highest to lowest). The final row reports the per-model average accuracy.
Figure 2: Rank-frequency heatmap across variants.
Figure 3: Boxplots of error rates (%) across variants of the same base sentence.
Figure 4: SPD(Vi→Vj) between source variant Vi and target variant Vj computed with Grok 4 Fast Reasoning. Values > 0 favor the row variant.
Figure 5: Spearman correlations among models’ SPD variant rankings.
Appendix figures & tables10 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 6: Heatmap showing all aggregated accuracy comparisons for the 9 models analyzed, organized by number, person, and gender effects.
Figure 7: Heatmap presenting all aggregated SPD comparisons for the 9 models analyzed, organized according to number, person, and gender effects.
Figure 8: SPD(Vi→Vj) computed with GPT-4o Mini.
Figure 9: SPD(Vi→Vj) computed with LLaMA 3.3 (70B-Instruct).
Figure 10: SPD(Vi→Vj) computed with LLaMA 3.1 (8B-Instruct).
Figure 11: SPD(Vi→Vj) computed with LLaMA 3.2 (3B-Instruct).
Figure 12: SPD(Vi→Vj) computed with Gemma 2 (27B-Instruct).
Figure 13: SPD(Vi→Vj) computed with DeepSeek R1 Qwen (7B).
Figure 14: SPD(Vi→Vj) computed with DeepSeek R1 Qwen (1.5B).
Figure 15: SPD(Vi→Vj) computed with Mistral (7B-Instruct).
Understanding gender biases in large language models (LLMs) is increasingly important as these systems become embedded in decision-support tools with real consequences. Prior research has focused only on a small set of models, leaving open the extent to which gender biases are common and heterogeneous across LLMs. We address this gap across ten models released between April 2025 and June 2026, spanning nine vendors, using two paradigms: gender attribution to stereotyped phrases (Study 1) and moral judgment of abuse or torture against a woman or a man to prevent a catastrophic outcome (Study 2). In Study 1, two of ten models attributed masculine-stereotyped phrases to female writers more often than the reverse, while three models showed the opposite pattern. In Study 2, several models converged on a male-disadvantaging asymmetry that was directionally consistent with a documented human tendency to protect female targets from harm, though the specific conditions under which this asymmetry emerged varied by model; three other models, by contrast, showed no variation across conditions. These results indicate that gender-related biases are common in LLMs. Their direction and magnitude, however, are highly heterogeneous, to the point that some models behave in diametrically opposite ways to others. Bias auditing should therefore be treated as an ongoing, multi-vendor process, rather than a one-time assessment.
Existing studies on gender bias in LLMs have largely focused on stereotypes, occupational associations, or explicit harmful outputs. In this work, we ask whether LLMs apply consistent response standards to the same negative behavior under matched male-actor and female-actor conditions. We introduce GAMA-Bench, a gender-mirrored benchmark of 1,298 scenarios covering intimate relationship and public social conflicts. It constructs gender-neutral misconduct templates through controlled grids and cross-model review, then compiles them into paired first-person prompts with matched actor-gender and role-reference variations. We further design a structured response-framing protocol to measure how models allocate punishment, empathy, escalation, instruction, and blame. Experiments on 10 representative LLMs reveal a consistent male-disadvantaging asymmetry: male actors receive more punitive, escalatory, and blame-centered framing, whereas female actors receive more therapeutic and empathy-oriented framing for the same misconduct. Further analyses show that this pattern persists across model families, scenario tracks, model scale, and explicit thinking-style reasoning. The official code is available at https://github.com/xufeiqiong/GAMA-Bench.
Guangzong Si, Dong Wang, Zhenhao Li +3
University of Science and Technology of China · Eastern Institute of Technology, Ningbo
Instruct-based large language models (LLMs) have been shown to propagate and even amplify gender bias when prompted with contextually constrained instructions (e.g., writing a text from a description or selecting a gendered pronoun). However, little attention has been paid to biases in responses to contextually unconstrained (generic) instructions conveyed by gendered language, particularly masculine generics (MG). MG, found in many gender-marked languages, denote the use of the masculine gender as a supposedly neutral reference to mixed-gender groups or individuals whose gender is unknown or non-binary. Yet, psycholinguistic studies demonstrate that MG are not neutral and systematically induce gender bias. This study investigates how both local and proprietary LLMs are MG-biased when responding to generic prompts in French, examining LLMs' MG bias rates and use of gender-fair language (GFL). We create a 16k+ human noun database from existing lexical resources and evaluate six LLMs on four instruction-response datasets under two conditions: prompts with and without MG. Overall, we find that ≈27.57% of LLMs' responses to MG-filtered generic instructions are MG-biased (≈78.55% with MG-containing prompts). Moreover, we find that LLMs rarely use GFL spontaneously. These findings highlight the persistence of MG bias in LLM outputs and models' limited tendency towards GFL strategies.