Fixing AI bias – using AI

Bias in computer code is not new; back in 1988, the UK Commission for Racial Equality found a British medical school guilty of discrimination because the programme it was using to pick out interviewees from job applicants was biased against women and people with non-European names.

Twenty years later, Amazon also had to stop using a hiring algorithm after finding it was favouring male applicants who used words like ‘executed’ or ‘captured’.

But the most visible form of bias in artificial intelligence is in imagery itself.

A 2018 Gartner report predicted that right up to 2030, 85 percent of AI projcts would still produce false results caused by bias – bias either built into the algorithms or biases in the AI team members.

When new AI imaging tools such as DALL-E, Midjourney and Stable Diffusion are asked to create artificial images of a chief executive (CEO) or a doctor, a firefighter or an astronaut, a founder or a president – the pictures they produce are of men.

When asked to create AI imagery of a nurse, an assistant or a flight attendant? Now the AI tools picture a woman.

When it comes to creating pictures of a CEO or a director, OpenAI’s DALL-E 2 generated images of white men 97 percent of the time.

We asked ChatGPT, another tool in OpenAI’s stable, why this happens.

AI tools are trained on billions and billions of fragments of data and images scraped from the internet. They use the file names, descriptions and other meta data uploaded along with the image to understand what is contained within the image. [Link to accessibility article]

If that image – and thousands or billions of others uploaded in the 40-year history of the internet – shows a male CEO, that is what AI presumes a CEO looks like.

AI is simply reflecting – and then amplifying – society’s prejudices and stereotypes around gender, race, culture and age.

OpenAI’s ‘risks and limitations’ document notes: “DALL·E 2 additionally inherits various biases from its training data, and its outputs sometimes reinforce societal stereotypes.”

Research scientist Sasha Luccioni also noticed that AI was more likely to create imagery of women as chief executives (CEOs) – or in 149 other professions – if ‘feminine’ adjectives such as ‘gentle’ and ‘supportive’ were added to the prompt, instead of ‘masculine’ adjectives such as ‘assertive’ and ‘decisive’.

A video posted to Reddit recently showed the images that Midjourney created of professors based on the prompt “A profile photo of a [insert department] professor.”

Midjourney created 22 images: 15 were male and seven female. The women professors it created operated in roles such as anthropology, art history and gender studies. However, the male professors it imagined worked in STEM (science, technology, engineering and maths) and business subjects.

Fixing the bAIs is a very visual representation of what’s wrong, and makes systemic issues and the unconscious bias very easy to understand. Diana Wilde, co-founder Aurora50

This is where Fixing the bAIs comes in. Aurora50 has worked to produce an AI tool that can fix its own bias.

Vast image bank of women in work

Fixing the bAIs is a vast image bank of women in various professions, to teach AI that women can be engineers, mathematicians, CEOs and more – and that men can work in professions such as nursing.

Gender is removed from the file names, tags and descriptions.

All of our images are royalty- and rights-free, and we encourage you to download and use them on the internet so that diversity in jobs becomes commonplace in public images.

Fixing the bAIs has won a clutch of international awards – bronze at the Epica Awards, three golds, a silver and a bronze at the Cresta Awards and three silvers at the Gerety Awards.