GenderBench 1.1 Results

What is GenderBench?

GenderBench is an open-source evaluation suite designed to comprehensively benchmark gender biases in large language models (LLMs). It uses a variety of tests, called probes, each targeting a specific type of unfair behavior.

What is this document?

This document presents the results of GenderBench 1.1, evaluating various LLMs. It provides an empirical overview of the current state of the field as of May 2025. It contains three main parts:

• Final marks - This section shows the marks calculated for evaluated LLMs in various categories.
• Executive summary - This section summarizes our main findings and observations.
• Detailed results - This sections presents the raw data.

How can I learn more?

For further details, visit the project's repository. We welcome collaborations and contributions.

Final marks

This section presents the main output from our evaluation. Each LLM has received marks based on its performance with various probes. To categorize the severity of harmful behaviors, we use a four-tier system:

• A - Healthy. No detectable signs of harmful behavior.
• B - Cautionary. Low-intensity harmful behavior, often subtle enough to go unnoticed.
• C - Critical. Noticeable harmful behavior that may affect user experience.
• D - Catastrophic. Harmful behavior is common and present in most assessed interactions.

---

Harms

We categorize the behaviors we quantify based on the type of harm they cause:

• Outcome disparity - Outcome disparity refers to unfair differences in outcomes across genders. This includes differences in the likelihood of receiving a positive outcome (e.g., loan approval from an AI system) as well as discrepancies in predictive accuracy across genders (e.g., the accuracy of an AI-based medical diagnosis).
• Stereotypical reasoning - Stereotypical reasoning involves using language that reflects stereotypes (e.g., differences in how AI writes business communication for men versus women), or using stereotypical assumptions during reasoning (e.g., agreeing with stereotypical statements about gender roles). Unlike outcome disparity, this category does not focus on directly measurable outcomes but rather on biased patterns in language and reasoning.
• Representational harms - Representational harms concern how different genders are portrayed, including issues like under-representation, denigration, etc. In the context of our probes, this category currently only addresses gender balance in generated texts.

---

Comprehensive table

Below is a table that summarizes all the marks received by the evaluated models. It is also possible to categorize the marks by harm. The marks are sorted by their value.

Categorize by harm

	Outcome disparity	Stereotypical reasoning	Representational harms
claude-3-5-haiku	🟩🟩🟩🟩🟩🟨🟧	🟩🟩🟩🟩🟩🟨🟨🟥🟥	🟧🟥🟥
gemini-2.0-flash	🟩🟩🟩🟩🟩🟨🟧	🟩🟩🟩🟩🟩🟩🟨🟥🟥	🟧🟧🟧
gemini-2.0-flash-lite	🟩🟩🟩🟩🟩🟨🟧	🟩🟩🟩🟩🟩🟩🟧🟥🟥	🟨🟨🟧
gemma-2-27b-it	🟩🟩🟩🟩🟩🟩🟥	🟩🟩🟩🟩🟩🟨🟨🟥🟥	🟧🟧🟧
gemma-2-9b-it	🟩🟩🟩🟩🟩🟨🟥	🟩🟩🟩🟩🟩🟩🟨🟥🟥	🟧🟧🟧
gpt-4o	🟩🟩🟩🟩🟩🟧🟥	🟩🟩🟩🟩🟩🟧🟧🟧🟥	🟧🟧🟧
gpt-4o-mini	🟩🟩🟩🟩🟩🟨🟧	🟩🟩🟩🟩🟨🟨🟧🟥🟥	🟧🟧🟧
Llama-3.1-8B-Instruct	🟩🟩🟩🟩🟩🟩🟨	🟩🟩🟩🟩🟨🟧🟧🟧🟥	🟩🟧🟧
Llama-3.3-70B-Instruct	🟩🟩🟩🟩🟩🟩🟧	🟩🟩🟩🟩🟩🟧🟧🟥🟥	🟧🟧🟥
Mistral-7B-Instruct-v0.3	🟩🟩🟩🟩🟩🟨🟧	🟩🟩🟩🟩🟧🟧🟧🟧🟥	🟨🟨🟧
Mistral-Small-24B-Instruct-2501	🟩🟩🟩🟩🟩🟩🟧	🟩🟩🟩🟩🟩🟨🟧🟧🟥	🟧🟧🟧
phi-4	🟩🟩🟩🟩🟩🟨🟧	🟩🟩🟩🟩🟩🟧🟧🟥🟥	🟨🟧🟧

	All
claude-3-5-haiku	🟩🟩🟩🟩🟩🟩🟩🟩🟩🟩🟨🟨🟨🟧🟧🟥🟥🟥🟥
gemini-2.0-flash	🟩🟩🟩🟩🟩🟩🟩🟩🟩🟩🟩🟨🟨🟧🟧🟧🟧🟥🟥
gemini-2.0-flash-lite	🟩🟩🟩🟩🟩🟩🟩🟩🟩🟩🟩🟨🟨🟨🟧🟧🟧🟥🟥
gemma-2-27b-it	🟩🟩🟩🟩🟩🟩🟩🟩🟩🟩🟩🟨🟨🟧🟧🟧🟥🟥🟥
gemma-2-9b-it	🟩🟩🟩🟩🟩🟩🟩🟩🟩🟩🟩🟨🟨🟧🟧🟧🟥🟥🟥
gpt-4o	🟩🟩🟩🟩🟩🟩🟩🟩🟩🟩🟧🟧🟧🟧🟧🟧🟧🟥🟥
gpt-4o-mini	🟩🟩🟩🟩🟩🟩🟩🟩🟩🟨🟨🟨🟧🟧🟧🟧🟧🟥🟥
Llama-3.1-8B-Instruct	🟩🟩🟩🟩🟩🟩🟩🟩🟩🟩🟩🟨🟨🟧🟧🟧🟧🟧🟥
Llama-3.3-70B-Instruct	🟩🟩🟩🟩🟩🟩🟩🟩🟩🟩🟩🟧🟧🟧🟧🟧🟥🟥🟥
Mistral-7B-Instruct-v0.3	🟩🟩🟩🟩🟩🟩🟩🟩🟩🟨🟨🟨🟧🟧🟧🟧🟧🟧🟥
Mistral-Small-24B-Instruct-2501	🟩🟩🟩🟩🟩🟩🟩🟩🟩🟩🟩🟨🟧🟧🟧🟧🟧🟧🟥
phi-4	🟩🟩🟩🟩🟩🟩🟩🟩🟩🟩🟨🟨🟧🟧🟧🟧🟧🟥🟥

Executive summary

This section introduces several high-level observations we have made based on our results. All the data we used to infer these observations are in the figures below.

---

🙈 Note on completeness

This benchmark captures only a subset of potential gender biases - others may exist beyond our scope. Biases can manifest differently across contexts, cultures, or languages, making complete coverage impossible. Results should be interpreted as indicative, not exhaustive.

Converging behavior

All the LLMs we evaluated have noticeably similar behavior. If one model proves to be healthy for a given probe, others likely are too. If one LLM prefers one gender in a given probe, others likely prefer it too. This is not surprising, as we have seen a remarkable convergence of training recipes in recent years. Most AI labs train their LLMs using similar methods, data, and sometimes even outputs from competitors. In effect, the behavior of the LLMs is very similar.

LLMs treat women better

Historically, it was assumed that machine learning models might treat men better due to their historically advantageous position that is often reflected in training text corpora. However, when we directly compare the treatment for men and women, our probes show either equal treatment or women being treated better. In creative writing, most of the characters are written as women, in decision-making, women might have a slight edge over men, when asked about who is right in relationship conflicts, LLMs tend to take women's side. This overcorrection should be considered when deploying the LLMs into production.

Strong stereotypical reasoning

Using gender-stereotypical reasoning is a relatively common failure mode. LLMs tend to write characters with stereotypical traits, assign stereotypical statements to certain genders, agree with stereotypical ideas, and so on. Stereotypical associations with occupations are especially troubling, considering the usage of LLMs in professional settings. Mitigating this issue is extremely challenging, as stereotypes are deeply embedded in vast amounts of training data.

Decision-making deserves caution

Decision-making in everyday and business situations, such as hiring decisions or financial decisions, does not seem to be strongly affected by biases, but there are still cases when the results could be characterized as unfair. We recommend special caution in all use cases when the LLM is making decisions based on data that contain the information about gender. Fairness should always be monitored. Removal of gender-related personal information, such as names or pronouns, can also be considered as a mitigation measure.

What is missing

There are still noticeable gaps in our evaluation. GenderBench currently does not address several important verticals, such as multimodal processing, non-English languages, reasoning capabilities, or multi-turn conversations. These will be progressively covered in future releases.

Outcome disparity

This section shows the probe results for the outcome disparity probes. This includes differences in the likelihood of receiving a positive outcome (e.g., loan approval from an AI system) as well as discrepancies in predictive accuracy across genders (e.g., the accuracy of an AI-based medical diagnosis).

---

Stereotypical reasoning

This section shows the probe results for the stereotypical reasoning probes. Stereotypical reasoning involves using language that reflects stereotypes (e.g., differences in how AI writes business communication for men versus women), or using stereotypical assumptions during reasoning (e.g., agreeing with stereotypical statements about gender roles).

---

Representational harms

This section shows the probe results for the representational harms probes. Representational harms concern how different genders are portrayed, including issues like under-representation, denigration, etc.

---

Treatment of women and men

This section directly compares the treatment of men and women in situations when it can clearly be said that one or the other group is being preferred. In the probe below, negative values mean that the LLMs give preferential treatment for women, positive values mean preferential treatment for men.

---

Normalized results

The table below presents the results used to calculate the marks, normalized in different ways to fall within the [0, 1] interval, where 0 and 1 represent the theoretically least and most biased models respectively. We also display the average result for each model.

---

	DiscriminationTamkin.max_diff	DiversityMedQa.diff_success_rate	HiringAn.diff_acceptance_rate	HiringAn.diff_regression	HiringBloomberg.masculine_rate	HiringBloomberg.stereotype_rate	RelationshipLevy.diff_success_rate	Bbq.stereotype_rate	BusinessVocabulary.mean_diff	Direct.fail_rate	Dreaddit.max_diff_stress_rate	Gest.stereotype_rate	GestCreative.stereotype_rate	Inventories.stereotype_rate	Isear.max_diff	JobsLum.stereotype_rate	GestCreative.masculine_rate	Inventories.masculine_rate	JobsLum.masculine_rate	Average
claude-3-5-haiku	0.062	0.010	0.022	0.006	0.021	0.015	0.329	0.096	0.000	0.026	0.005	0.578	0.116	0.116	0.077	0.572	0.400	0.404	0.231	0.162
gemini-2.0-flash	0.023	0.023	0.003	0.017	0.044	0.000	0.312	0.013	0.000	0.046	0.007	0.687	0.106	0.000	0.059	0.571	0.257	0.160	0.202	0.133
gemini-2.0-flash-lite	0.007	0.001	0.001	0.000	0.041	0.011	0.277	0.033	0.000	0.037	0.013	0.535	0.176	0.105	0.078	0.747	0.068	0.283	0.109	0.133
gemma-2-27b-it	0.039	0.002	0.003	0.016	0.030	0.023	0.635	0.020	0.003	0.037	0.013	0.563	0.154	0.160	0.060	0.591	0.220	0.279	0.209	0.161
gemma-2-9b-it	0.043	0.001	0.024	0.001	0.010	0.011	0.543	0.011	0.004	0.030	0.008	0.477	0.132	0.097	0.067	0.604	0.262	0.294	0.193	0.148
gpt-4o	0.007	0.004	0.020	0.026	0.101	0.009	0.542	0.001	0.000	0.052	0.010	0.238	0.287	0.279	0.021	0.624	0.169	0.205	0.195	0.147
gpt-4o-mini	0.020	0.003	0.011	0.002	0.061	0.000	0.379	0.075	0.003	0.085	0.009	0.415	0.227	0.153	0.029	0.593	0.294	0.294	0.211	0.151
Llama-3.1-8B-Instruct	0.078	0.015	0.001	0.017	0.023	0.044	0.126	0.207	0.018	0.017	0.011	0.108	0.232	0.280	0.071	0.842	0.259	0.313	0.078	0.144
Llama-3.3-70B-Instruct	0.010	0.002	0.027	0.022	0.024	0.008	0.290	0.041	0.022	0.042	0.009	0.641	0.195	0.271	0.062	0.648	0.340	0.313	0.188	0.166
Mistral-7B-Instruct-v0.3	0.008	0.009	0.005	0.011	0.057	0.014	0.443	0.238	0.000	0.053	0.002	0.143	0.270	0.284	0.078	0.801	0.100	0.188	0.095	0.147
Mistral-Small-24B-Instruct-2501	0.036	0.002	0.005	0.006	0.026	0.001	0.464	0.049	0.000	0.031	0.017	0.165	0.215	0.159	0.038	0.689	0.266	0.271	0.150	0.136
phi-4	0.024	0.002	0.008	0.020	0.057	0.002	0.272	0.017	0.000	0.031	0.008	0.416	0.338	0.320	0.030	0.747	0.143	0.277	0.124	0.149

Methodological Notes

• The results were obtained by using genderbench library version 1.1.
• Marks (A-D) are assigned by comparing confidence intervals to predefined thresholds. A probe's final mark is the healthiest category that overlaps with its confidence interval.