BBVA and IBM Research Develop Dataset to Test Bias in Non-English Generative AI Models

BBVA and IBM Research Develop Dataset to Test Bias in Non-English Generative AI Models

(IN BRIEF) BBVA and IBM Research have created a dataset to evaluate biases in generative AI models in languages other than English, specifically testing for biases in Spanish. The dataset, based on IBM’s SocialStigmaQA (SSQA), examines how AI models respond to stigmatized scenarios involving race, gender, and other factors. Early tests showed greater bias in non-English languages. The dataset is available for the open-source community to further improve. This research, presented at NeurIPS, will help develop more equitable AI systems and is a part of BBVA’s commitment to responsible AI development.

(PRESS RELEASE) BILBAO, 6-Jan-2025 — /EuropaWire/ — BBVA, in collaboration with IBM Research, has developed a groundbreaking dataset designed to evaluate the presence of discriminatory biases in generative artificial intelligence (GenAI) models in languages other than English. This stress test, tailored to measure biases in Spanish, was presented at NeurIPS, the largest AI conference, and made available to the open-source community for further research. The dataset, based on IBM’s SocialStigmaQA (SSQA), helps identify biases related to gender, race, sexual orientation, disability, and more by testing how generative AI models respond to stigmatized scenarios.

The SSQA dataset includes around 100 ‘stigma’ conditions and 40 hypothetical situations, combining these elements to generate prompts for AI models. For example, a situation might involve a recommendation for a caregiver based on a stigmatized trait, and the model’s response is then benchmarked for bias. Early analysis revealed that non-English languages, including Spanish, displayed more pronounced biases than English-based tests, highlighting the importance of ensuring that AI systems are culturally and socially relevant across diverse linguistic regions.

Clara Higuera, a key author and data scientist at BBVA’s GenAI Lab, emphasized the significance of this work in advancing responsible and equitable AI practices. BBVA’s commitment to ensuring safe and inclusive AI is reflected in their collaborations, including those with OpenAI, and their ongoing analysis of AI model fairness. Researchers are also planning to expand this dataset further, with potential future versions to include data from sources such as the European Social Survey. Looking ahead, BBVA is considering creating a specialized dataset for the banking sector.

The research was showcased at the NeurIPS conference’s ‘Socially Responsible Language Modelling Research’ workshop, and the datasets in Spanish and Japanese are now available on GitHub and HuggingFace for global collaboration and enhancement. This project marks a significant step toward understanding and mitigating bias in generative AI.

Media Contact:

Tel.: 944 875 555

SOURCE: BBVA