-25
Gender Bias in Text-to-Image Generative Artificial Intelligence When Representing Cardiologists
www.mdpi.comIntroduction: While the global medical graduate and student population is approximately 50% female, only 13–15% of cardiologists and 20–27% of training fellows in cardiology are female. The potentially transformative use of text-to-image generative artificial intelligence (AI) could improve promotions and professional perceptions. In particular, DALL-E 3 offers a useful tool for promotion and education, but it could reinforce gender and ethnicity biases. Method: Responding to pre-specified prompts, DALL-E 3 via GPT-4 generated a series of individual and group images of cardiologists. Overall, 44 images were produced, including 32 images that contained individual characters and 12 group images that contained between 7 and 17 characters. All images were independently analysed by three reviewers for the characters’ apparent genders, ages, and skin tones. Results: Among all images combined, 86% (N = 123) of cardiologists were depicted as male. A light skin tone was observed in 93% (N = 133) of cardiologists. The gender distribution was not statistically different from that of actual Australian workforce data (p = 0.7342), but this represents a DALL-E 3 gender bias and the under-representation of females in the cardiology workforce. Conclusions: Gender bias associated with text-to-image generative AI when using DALL-E 3 among cardiologists limits its usefulness for promotion and education in addressing the workforce gender disparities.
So people are complaining that the depictions of cardiologists, when you ask an image generator to show you one, are too accurate? That 85% of the time they show you a man in a job where 85% of people are men?
You’d likely see the same thing if you asked one to show you a warehouse worker, another job that’s male-dominated.
If people want more women in these roles, push for it at the university level. Push for it in medical posters in hospitals. I don’t see how forcing the hundreds of AI models out there to be biased in favour of depicting women when their training material doesn’t have as many is an effective way of achieving this goal.
This just seems like a “we want to complain about this field being male dominated, and we’re sure to get headlines if we include the AI buzzword”
Because the biases in an AI model will shape the perception of people who may think about entering those fields more than a poster at a place where people have already entered those fields work at.
Likewise you can train it out of a bias, just feed it more content showing diverse workforces and it will start weighing them higher.
Representative of actually real world statistics is not “bias”.
People are not statistics.
There are huge biases at play for why gender differences exist in the world place and most other places.
Those biases carry on into the works we make, such as LLMs which are becoming hugely influential. To tackle these biases, we need to change how we view them. That means if the statistical average scientist is a white man, that we show more women and PoV to help them feel like they too can do this.
tl;dr: We don’t need to reflect society as it currently is, we should aim to show how it can be.
Statistics are people.
It cannot possibly, under any circumstances, be a correct, reasonable, or valid word choice to describe an “AI” that accurately models reality as “biased”. That word already has a meaning and using it in that manner is a lie.
Bias is an irrational departure from reality. You can want an AI to be biased towards diversity, but that is adding bias, not removing it.
This. I’d you don’t like the AI presents more men than women as cardiologists because its mirroring society, the problem is not the AI, the problem is the lack of non cis-male representation in real world cardiology.
The depiction aligning with reality is not a bias. Artificially altering the algorithm so that it shows more women for this prompt on the other hand is, unquestionably, adding a bias.
If you want to add a bias, fine. Biases aren’t always a bad thing, I can certainly see the argument for why you might want a 50/50 gender split for all AI prompts. But don’t pretend that what you’re actually advocating for here is correcting a bias, because it isn’t.
That is training in a bias. Because it’s not representative of reality.
Exactly. There are perfectly legitimate reasons to wish to bias learning to certain changes in output.
There are no legitimate reasons to do so without acknowledging that you are adding bias and being clear on the intent of doing so.
And even in cases where introducing a bias is desirable, you have to be very careful when doing it. There has been at least one case where introducing a bias towards diversity has caused problems when the algorithm is asked for images of historical people, who were often not diverse at all.