Addressing bias in AI face changing technology


At Change my Face, we use the power of artificial intelligence to generate facial simulations, ie visualise your future self, to improve health habits. However, with power comes great responsibility and we recognize that AI systems can be prone to bias, which can lead to discrimination in daily life. Bias in AI refers to the systematic and unfair favouritism or prejudice shown by AI algorithms towards certain groups of people, based on factors such as race, gender, or age.

This bias can have significant consequences in various aspects of daily life. Biased AI algorithms can inadvertently reinforce stereotypes and perpetuate existing inequalities but also in our work, creating facial simulations. When developing models using publicly available image datasets that primarily feature male Caucasians, like CelebA and FFQH, our algorithms may fail to accurately capture the diverse range of human appearances. As a result, an individual with Asian features may unintentionally be transformed to appear Caucasian as they age with lighter skin, or different shaped eyes. Many other cultural features may be overlooked and when testing for weight gain and diet generated imaging, we turned Barack Obama into a large, smiling Asian dude — not ideal.

To combat bias in AI and create more inclusive and equitable systems, we are making the most of the grant we received from the Inclusive Innovation Award (Innovate UK) to build less biased image datasets. Our goal is to build a brand new synthetic image dataset that adequately represents the cultural and gender diversity of our society purely using generated AI images and test these against existing datasets, which bypasses privacy and security laws. By including a wide range of ethnicities, ages, genders, and other relevant factors, we aim to mitigate and reduce the inherent biases present in existing datasets.

Additionally, we are implementing multiple tests and validation processes to identify and address bias within our algorithms. By analysing the outputs of our facial simulations across various demographic groups, we can identify any potential biases and refine our models accordingly. We aim to seek feedback from diverse user groups to ensure that our AI-powered facial simulations are inclusive, accurate, and respectful of individuals’ unique identities.

From this work, we’ve created Future Face, an interactive skin analysis tool to help consumers shop more easily for online skincare products. People with different skin types, ethnic origins, lifestyles and environments will have different needs for their skin and this kind of digital tool co-exists harmoniously with the new personalised skincare brands that are emerging post Covid.

We firmly believe that combating bias in AI requires a collaborative effort from all stakeholders. We actively collaborate with organisations, researchers, and experts in the field to develop best practices and standards for fair and unbiased AI systems. Through transparency, accountability, and ongoing research, we strive to build AI technologies that serve all individuals, regardless of their background.

With the support of the Inclusive Innovation Award grant, we are confident in our ability to reduce bias in AI, even in a small way, and pave the way for a more inclusive future. Together, we can harness the potential of AI-powered facial simulations to celebrate and acknowledge the diverse society we live in.