Midjourney is incredible. But you can see there are definite existing biases in its dataset.
I’ve been playing with Midjourney and I think it is incredible.
The TLDR of that: We have sites now where you can type in a phrase and AI generates images for you based on concepts. Problem is: every dataset is biased and comes with issues. The images probably aren’t and …will always be grade-B versions of actual art.
But they’re great tools. Need to have an illustration of… an engineer, sitting at their desk, while the clock ticks over their head?
There is of course some bias in its results (why was that engineer male and slightly bald) but you can generally fix a ton of that by adding keywords. The main thing is to be aware the bias exists (and will probably improve and change). But as a sampling, here are a number of base word searches, as of July 29th 2022. I believe they’re on V3, and I tried not to change the parameters beyond the baseline.
For each of the keywords in the images I ran the concept 5 times at the relaxed speed and just took a copy of the first results.
The images are worth perusing first.
As a baseline, I tried various forms of the most neutral word, and generally yes. All the images look human. [There is a conspicuous lacking of melanin. Just saying.]
Now, let's compare men.
And women!
Ok, now let's add some basic cultural concepts to see how that changes things.
Ok, let's try broaching race and see what happens there…taking the base words from the census, what happens with Midjourney?
Some words clearly come with more meaning and weight in them towards a person than others. I find it interesting that these seem to lean towards displaying women… but surprised that not a single person for black wasn’t…not only a white person but a kinda goth white person….
Some of this, though, is a part of what Midjourney is. Without the nouns, the adjectives/descriptors are useless. So we clearly can see that with the right adjectives, you can easily get people across cultures and races.
What happens when we add in adjectives/concepts? Does any bias exist there?
Again nothing terrible. Just a general lack of melanin in the pictures.
For most of these so far, the bias feels…underwhelming. It exists but isn’t harming things. As long as you’re aware, you can add some adjectives and concepts and get the specificity you want.
But let’s step into careers.
What if we try the sorta, classic psychological archetypes.
This is about where I petered out and started looking at more physical descriptors. But I think that could wait until the next article?
I’m half doing this just for my own purposes and thinking through the results the system gives me, so that when I’m making other images, they can think of that innate bias. But I think this is something the application creators need to consider too. These are the first quick 5 pictures of leaders.
And these are designers.
I’m not 100% what the fix is. Pictures aren’t always representing society. Now we’re making art based on pictures that didn’t represent society. Do we make the art lie? Do we skew the existing data for a society we want? Or do we just get creative in the meantime, until… everyone can be anything?
