Skip to main content

International Women's Day 2023

Debiasing AI for a fairer future

Generative AI has seen a huge surge in popularity. It has the potential to be truly revolutionary for industry and indeed for human society. However, the increased exposure to and use of AI has revealed one of its most pervasive flaws: bias. The bias displayed in most forms of AI reflects the bias that exists in society. If the data on which a model is trained doesn’t reflect diversity, then the model itself is confined to the same limited view or approach.

To celebrate International Women’s Day 2023, Freshfields convened an expert panel event, Machine Relearning: How To Debias The Tech We Live With, to discuss how the technology we use can be debiased, in order to help advance gender parity and drive better outcomes for society.

We were joined by Dorothy Chou, Head of Public Policy at DeepMind, Martina King, CEO of Featurespace, and Sana Khareghani, Former Head of the UK Government Office for AI. They discussed the impact of inherent bias in generative AI, the importance of diversity in the workforce developing this technology and considered the role of regulation. 

The panellists stressed the importance of looking at the people who are building technologies like generative AI and the culture of their organisations. They argued that data gaps exist across most areas in which AI is being used. Therefore, humans need to remain in the loop to monitor and correct potentially biased outcomes.

If the data that is being fed into the algorithm reflects existing inequalities, then there are things that we need to do to update how those models work in the world, in order for them to see a more accurate perspective based on how the world is changing.

Dorothy Chou, Head of Public Policy, DeepMind

To address these issues, the panellists suggested adding more data points, more granular information, and a wider range of representative data points to retrain models and reduce the risk that they reach biased conclusions. Synthetic data sets can also be used to evaluate past errors and create future outcomes that are more accurate and tolerable.

The discussion concluded with the idea that diversity in leadership roles is key to shaping upstream decision making and debiasing the technology that increasingly shapes our lives.

Watch the discussion in full: