In this column, ‘Our World in AI’, we investigate how Artificial Intelligence sees the world. We use OpenAI’s DALL-E to generate a set of images for an aspect of society and analyse the result. Will AI shape a better place, or does it make biases worse?
Today’s prompt: “a corporate leader giving a presentation”
Before we look at the results, there are a few points to mention. First, DALL-E likes quite specific input. If we use only ‘corporate leader’, it takes much longer to generate output. Second, the images we use are simply the first 40 that DALL-E created, in that order, starting from the bottom. And last, the collection is publicly available here. Some faces and hands are pretty creepy up close.
Ok, so here’s what DALL-E produced for “a corporate leader giving a presentation” (Fig 1):
We have 32 men, and 29 of those are white. We also have eight women, and five are white. Ethnic diversity is higher among women, but we only have a small number of observations, and proportions may shift if we generate more images.
Let’s put some context around the data. Say we walk past a room where 100 corporate leaders are meeting. Then we will see 80 men and 20 women. In total, 85 people are white, and 15 have different ethnic backgrounds. The gender split is even among the ethnically diverse, but the white group is 85 per cent male. Overall, the single largest group of corporate leaders are white men at 72.5%.
So is this what the real world looks like? The United States and the United Kingdom did in 2019, according to this May 2020 report from McKinsey. Female representation on executive teams was 20 per cent, and ethnic minorities stood at 13 per cent. That’s very close to DALL-E’s numbers.
Of course, 2019 is now three whole years ago, and the hope is that improvements have been made since. We’ll keep looking for more recent data and update the article if we find any. In the final section of this column, we choose whether AI’s interpretation of society is leading, lagging, or live.
Today’s verdict: Lagging.
DALL-E accurately reflected the United States and the United Kingdom in 2019 when it generated images for “a corporate leader giving a presentation”. The verdict may seem harsh, but we take the optimistic position that recent data would show more progress.
Next week in Our World in AI: commuting.