The A-Z of AI

Bias

Incomplete data can lead to bias in AI.

The results of AI systems can be affected by data that amplifies existing biases found in the real world.

Typically, AI forms a bias when the data it’s given to learn from isn’t fully comprehensive and therefore starts leading it toward certain outcomes. Because data is an AI system’s only means of learning, it could end up reproducing any imbalances or biases found within the original information.

For example, if you were teaching AI to recognise shoes and only showed it imagery of trainers, it wouldn’t learn to recognise high heels, sandals or boots as shoes.

Bias makes it challenging to develop AI that works for everyone.

No AI system is complex enough, nor dataset deep enough, to represent and understand humanity in all its diversity. This can present profound challenges when you consider the potential AI has to influence the experiences of real people.

A job-matching AI shortlisting candidates for CEO interviews might learn to favour men, simply because it was given successful CVs to learn from, and historically there was a societal bias towards male candidates.

To ensure that AI systems are ethical and reduce the risk of bias, programmers design their systems and curate their data vigilantly. This is the only way to ensure that systems work well for everybody.