⚠️ Our findings highlight the need for multilingual AI safety efforts. Focusing on English alone means we risk missing important harms in other languages.

Paper 📄: arxiv.org/abs/2506.03867
Data: 🗄️: github.com/JacquelineRo...

2️⃣ the larger the model, the more strongly these stereotypes are encoded

3️⃣ instruction finetuning does not consistently reduce gendered stereotyping in LLMs

We then use the new dataset to test 24 LLMs from a range of families and sizes. We show that:

1️⃣ LLMs display gender-stereotypical reasoning across languages, associating women with beauty, empathy and neatness, and men with leadership, strength, toughness and professionalism

Building on an existing expert-informed benchmark by @matus-pikuliak.bsky.social et al, we use translation tools, quality estimation metrics, and morphological heuristics to adapt the benchmark to the additional languages. This pipeline results in high quality data, validated by expert translators

Huge congrats to Swolkien & team for this effort, it's a huge step towards officialising and promoting use of Kabuverdianu in Cape Verde 👏

Kudos also to Cape Verde government and Ministry of Education, who are recognising and investing in the central role of creole language for their citizens 🇨🇻