Tackling AI‘s diversity problem

Improve the transparency and diversity of AI development

Google, Facebook and Microsoft have been named by the Future Today Institute as three of the nine companies driving the future of AI – but less than 4% of their staff in the US are black (compared to almost 13% of America) and their workforce gender spilt is still far from equal. Diversity isn’t just for hiring either, though; organisations seeking to drive change need to pursue inclusive policies across the board and to champion these values. 

It is crucial for organisations building AI systems to ensure that a diverse range of individuals directly contribute to the process – from the inception of the system, all the way through to auditing. Drawing on different perspectives is the only way that the important questions will be asked, yet the current tech landscape does a poor job of representing the overall population.

Improving the transparency of how AI operates is also key. If a user does not understand how the algorithm has reached a decision (or the manufacturer will not share this information), then relying on it is an act of faith. There needs to be an onus on responsible operators to run audits, and to monitor what their algorithms have learned and how they reach their goals.

AI: a reflection of diversity

The real answer to tackling AI ethics is to ensure that a diverse range of people – spanning race, class and gender – are involved in the building of algorithms; while the organisation actively strives to consider the consequences that these systems may have for society. 

“The industry has to acknowledge the gravity of the situation and admit that its existing methods have failed to address these problems,” Kate Crawford, a co-author of the AI Now Institute’s report, told The Guardian. “The use of AI systems for the classification, detection, and prediction of race and gender is in urgent need of re-evaluation.”

The proposed Algorithmic Accountability Act in the US, for example, calls for companies to be required to screen their algorithms for bias if they have more than 1 million users or $50m in annual revenue. While this may be difficult to negotiate in practice, it points to the growing prominence of the issue – and calls for regulation are likely to only grow louder.

But if businesses take steps to act more responsibly by making their AI more transparent to users and ensure that a diverse talent pool have been involved in the creation and testing of algorithms, they may avoid the threat of regulation.