Imperial researcher awarded Royal Society Career Development Fellowship for culturally aligned AI research

The fellowship will support research aimed at reducing cultural bias in artificial intelligence and improving the safety and reliability of large language models for African language communities.

Dr Shamsuddeen Muhammad, Google DeepMind Academic Fellow in Imperial's Department of Computing, has been awarded a Royal Society Career Development Fellowship to advance research on making artificial intelligence (AI) systems more culturally aware and better aligned with the needs of African language communities.   

As AI technologies become integrated into sectors such as healthcare, education, business and public services, concerns have grown about the tendency of many models to generate culturally biased responses. This issue can be particularly pronounced when users interact with AI systems in African languages, where models often lack the cultural understanding needed to respond appropriately to local contexts and practices.  

Imperial's mission is creating Science for Humanity, and humanity does not speak one language or live in one culture. What we learn in this project will help make AI safer and fairer for everyone." Dr Shamsuddeen Muhammad Google DeepMind Academic Fellow

The four-year fellowship will support a programme of research addressing an increasingly important challenge in artificial intelligence: ensuring that large language models (LLMs) can serve people from diverse linguistic and cultural backgrounds safely, fairly and effectively.  

Commenting on the fellowship, Dr Muhammad said: "A model that speaks your language but does not understand your world can mislead the very people it is meant to help. This fellowship will let me work alongside African communities to align large language models with the values, norms and everyday realities of African societies, building models that reason carefully about cultural context before they respond. The aim is that speakers of African languages can have AI systems that are safe, culturally aware, fluent and trustworthy.

Imperial's mission is creating Science for Humanity, and humanity does not speak one language or live in one culture. What we learn in this project will help make AI safer and fairer for everyone."

Advancing culturally aware AI

Large language models are typically trained on vast amounts of online data, much of which originates from Western languages and cultural contexts. While these systems have demonstrated remarkable capabilities, they can also generate responses that reflect cultural assumptions, stereotypes and biases that may not translate appropriately across different societies and languages.  

These challenges can be particularly acute when users interact with AI systems in African languages. In sensitive areas such as maternal health, mental health and financial advice, a lack of cultural understanding can affect the relevance and trustworthiness of AI-generated responses.  

To address these challenges, the project, Mitigating Cultural Biases of LLMs in African Languages with Deliberative Alignment, will combine advances in AI safety with expertise from local communities and domain specialists to create culturally grounded training data, evaluation tools and safety guidance for AI systems.  

A key part of the research is the use of deliberative alignment, an emerging approach that trains models to reason through safety guidance before generating a response. The aim is to help AI systems provide answers that are culturally appropriate and sensitive to different community contexts, particularly in low-resource language settings where existing AI safety approaches have received relatively little attention.  

The project will also create open resources for the wider research community, including culturally annotated datasets and tools for evaluating AI safety and cultural alignment across African languages. These resources are intended to support future research and help improve the development of multilingual AI systems.  

The expected outcome is a new generation of language models that can better serve African language communities while reducing harmful stereotypes and Western-centric cultural biases. Beyond their application in African languages, the methods developed through the project could contribute to broader efforts to make AI systems safer, more inclusive and more effective for users around the world.  

About The Royal Society Career Development Fellowship  

The Royal Society Career Development Fellowship is a four-year postdoctoral fellowship that supports researchers from underrepresented backgrounds in STEM. Alongside funding for ambitious research programmes, the scheme provides mentoring, training and networking opportunities to help fellows establish independent research careers in the UK.