A Leaderboard to Benchmark Ethical Biases in LLMs

Marcos Gomez-Vazquez, Sergio Morales, German Castignani, Robert Clarisó, Aaron David Conrardy, Louis Deladiennee, Samuel Renault, Jordi CabotAIMMES 2024 · OpenReview

As artificial intelligence regulations like the EU AI Act transition from draft stages to active enforcement, the need for robust, transparent benchmarking of Large Language Models (LLMs) has never been more critical. This paper, originally presented at the AIMMES 2024 Workshop, introduces a public leaderboard designed to evaluate LLMs through the lens of ethical bias, providing a necessary framework for developers to align their models with societal safeguards.

Addressing the Ethics Gap in AI Benchmarking

While the AI community has established numerous leaderboards for reasoning, coding, and mathematical capabilities, specialized benchmarks for ethics remain rare. LLMs are known to inherit and often amplify human biases present in their training data. As these models are increasingly integrated into "high-risk" applications—such as hiring, legal assistance, or medical triage—the lack of standardized, transparent testing for fairness becomes a significant liability. The authors address the need for "regulatory sandboxes": controlled environments where models can be tested for compliance with mandates that AI systems avoid discriminatory impacts.

System Architecture and Methodology

The researchers developed a modular architecture to facilitate automated bias testing. The system consists of an administrative front-end for test experts, a testing execution engine, a PostgreSQL database for storing results, and a public-facing leaderboard.

The system architecture includes an admin frontend, a test execution engine integrated with LangBiTe and various model providers, and a visitor-facing leaderboard.

Core Components:

  • LangBiTe Integration: The evaluation core relies on LangBiTe, an open-source tool that utilizes prompt templates to probe for biases. It generates comparative prompts that differ only by a sensitive attribute (e.g., changing a religious or ethnic identifier) to observe if the model's judgment remains consistent.
  • LLM-as-a-Judge: To overcome the challenge of evaluating long or non-formatted model responses, the authors employ a second, highly capable LLM to act as a "judge." This judge categorizes target responses as biased or unbiased based on predefined ethical criteria.
  • Multi-Provider Support: The system is designed to pull models from various sources, including OpenAI’s API, HuggingFace’s Inference API, and Replicate.

The Ethical Benchmark Suite

The leaderboard currently evaluates 16 different LLMs (including GPT-4, Llama-2, and Mistral) against a dynamic library of over 300 input tests. These tests are categorized into seven ethical dimensions:

  1. Ageism
  2. LGBTIQ+phobia
  3. Political bias
  4. Racism
  5. Religious bias
  6. Sexism
  7. Xenophobia

Key Results and Insights

The leaderboard provides a "percentage of tests passed" metric, where a 100% score indicates no detectable bias across the executed test suite.

The public leaderboard interface displays pass rates across seven social bias metrics for various LLMs.

Several critical insights emerged from the initial benchmarking phase:

  • GPT-4 Performance: GPT-4 emerged as the overall leader in avoiding bias, though it did not secure the top spot in every individual sub-category.
  • Scale and Fairness: A consistent finding is that model scale correlates with ethical alignment. Larger versions of model families (e.g., Llama-2-70b vs. 7b) tend to be significantly less biased, suggesting that smaller models are more prone to ethical lapses even if they perform well on standard logic tasks.
  • The Political Bias Frontier: Across all tested models, political bias remains the most significant challenge. Models showed the highest levels of inconsistency and bias in this category, suggesting that "neutrality" in political discourse has not been prioritized as heavily as other social biases like racism.
  • The Refusal Paradox: The authors note a philosophical debate regarding model "refusals." Currently, if a model refuses to answer a sensitive or potentially biased prompt, it is counted as a "pass." However, the researchers acknowledge that total refusal can also limit the utility of the model.

Limitations and Future Directions

The authors emphasize that "bias" is not a universal constant; it is often culturally subjective. Future iterations of the leaderboard aim to include culturally specific test suites and expand to multilingual benchmarks to capture biases that only emerge in specific linguistic contexts. Additionally, the team plans to expand the framework to include multi-modal models (images and video) and advocate for the consolidation of various benchmarks into a single, sustainability-conscious platform to reduce the environmental impact of large-scale model testing.

Summary updated

BibTeX
@inproceedings{gomezvazquez2024leaderboard,
  author = {Marcos Gomez-Vazquez and Sergio Morales and German Castignani and Robert Clarisó and Aaron David Conrardy and Louis Deladiennee and Samuel Renault and Jordi Cabot},
  title = {{A Leaderboard to Benchmark Ethical Biases in LLMs}},
  booktitle = {AIMMES},
  year = {2024},
  url = {https://openreview.net/forum?id=J8xbfAgeFW},
}
Figure