Of bananas, racism and other AI horrors
In my last AI art creation blogpost, I attempted to depict the lead villain of my science-fantasy novel, only to discover that AI art generator, MidJourney, was determined to give him unwanted rhinoplasty.
In this next thrilling episode of 'AI art and other horrors', I will be attempting such awesome feats as 'getting an AI to draw bananas' along with the epic-for-all-ages challenge of 'visualising a biracial woman using AI'.
Let's start with the bananas. There's been a fair bit of discussion appearing on my Twitter feed about whether AIs should be producing art, whether AI will replace human artists, and so on. I think, with current technology, hailing 'the end of human artists' is a bit premature. And these bananas are proof.
I was having a discussion over lunch about racist AIs (we'll come to those in a moment) with the additional needs expert who helps with my older son. During the conversation, I decided to demonstrate what MidJourney did - and she provided the prompt 'ginger toddler with a banana, photorealistic' (our ginger toddler was, indeed, eating a banana at the time).
What came back was worthy of the best works of Stephen King. Ginger Toddler Eating a Banana (below) is a good example of why, for every beautiful AI-generated picture you see on the internet, the originating 'artist' probably generated thirty or more horror-movie-worthy images from the same text prompt.
Not only are these toddlers covered in strange rashes and creepily clown-like but, more disturbingly, they are eating caterpillars... Huge, fat, banana-ry caterpillars. And, at this point, we all had the burning question: "Can MidJourney draw bananas?"
I think MidJourney can draw bananas... but not if they're being held by a ginger toddler. But it's not 100% confident about the whole banana thing, which makes it less of a bona-fide artist.
After all, still life is pretty much the first thing you learn in art class.
And... onto the racism...
I promised you some AI racism. So, here it comes. You see, I tried using Artbreeder and MidJourney to visualise Syer Roth (aka Izadora lee Tigrane) the main character/protagonist of my science-fantasy novel. And, gentle reader, I found AI bigotry by the TETRABYTE.
First, a bit of background. Syer Roth is a 24-year-old fictional fighter pilot, who I can picture in my head in reasonable detail, and who lives in a fantasy world with approximately World War II-level technology. My mental image of Syer Roth looks a bit like Anya Taylor Joy in the Queen's Gambit and also a bit like Chandra Nalaar from Magic the Gathering. But not exactly like either of them, or any of the other reference pictures I'd found.
Aged 15, when I wrote the original science-fantasy novel, I thought I'd look better with ginger hair. So, Syer Roth has ginger hair (NB: it turns out that I'm a ginger gene carrier. Aged 40, I gave birth to an unexpectedly ginger son).
What's more interesting about Syer Roth's physical appearance (from an AI racism point-of-view, at least) is that my mental image of Syer Roth has dark eyes and olive skin. But, when you think of a 'ginger' or 'red-headed' woman in a fantasy novel, what stereotypically comes to mind is probably a Celtic-looking sorceress with green eyes and pale skin.
Only about 2% of the world's population have ginger hair, with the vast majority found in Ireland and Scotland. There's a whole sub-genre of fantasy inspired by Celtic mythology, such as the Katherine Kerr's Deverry Cycle, which I read as a kid, leading to a huge number of red-headed female characters, and the fantasy trope of red-haired queens, warriors and mages.
So, you're probably wondering why Syer Roth doesn't look Scottish or Irish. And the reason is that I REALLY loved Moorish architecture as a teenager. I was learning Spanish and there were loads of pictures of the Alhambra Palace in southern Spain on the classroom wall. So, I decided my fantasy setting had a Mediterranean climate and - by extension - the locals would tend to look southern European as well.
Now, until I started generating AI art, I never considered that Syer Roth's skin colour was 'a problem'. I live in London, which is among the most racially-diverse cities in the world, and where only 37% of the population are white British.
South-east London, where I live, is home to many descendants of the Windrush Generation of British-Caribbean immigrants who came to the UK after World War II. I'm simplifying the history a lot here, but my understanding is that many of the Windrush Generation worked, at least initially, as bus drivers, or in other traditionally working-class occupations, such as construction or manufacturing - often due to racial discrimination.
Over the last 70 years, these post-war migrants have assimilated, often into the working class (at least, they did in south-east London, which is a relatively poor area, and the area I know best). As people do, they married and had children, often with working-class white people living locally (it's a lot more complicated than that, but let's stick with this over-simplification). As a result, the working class in south-east London is now mostly biracial and it's much easier to find jerk chicken in a greasy-spoon cafe than pie and eels.
The genetics of gingerness
Some white working-class people have Scottish or Irish ancestry, meaning they carry copies of the genes for ginger hair. Ginger hair is a recessive trait, which means you need to inherit the genes from both parents. There are now known to be several genes associated with being ginger but - all you need to know, for now - is that you can be a ginger carrier without being ginger yourself.
Also, you can have dark skin and be ginger at the same time. This is because red hair and a freckly complexion is caused by the overexpression of a type of pink/red melanin called pheomelanin. If you simultaneously express large amounts of brown/black melanin (eumelanin) AND overexpress pheomelanin, you can have dark skin AND red hair and freckles, at the exact same time.
Anyway, over the last seventy years or so, biracial Londoners carrying a single copy of the genes for ginger hair had families with other ginger carriers. Until, eventually, it's become common for biracial south-east Londoners to be simultaneously black and ginger. In fact, being black and ginger is so common, there are similar numbers of noticeably biracial ginger kids and stereotypical white redheads at my son's school.
Knowing that NOT ALL RED HEADS HAVE WHITE SKIN AND GREEN EYES, I blithely tried to create a brown-eyed, olive-skinned, ginger-haired Syer Roth with the Artbreeder AI. And, dear reader, I succeeded. But, boy, was it a hassle...
I won't go into all the trials of doing this, it's in my previous blogpost. Let's just say there was a lot of turning the 'middle-eastern' slider UP TO ELEVEN.
Having created Syer Roth as I wanted her, including her longish face (her dad is a WWI-era test pilot who I always imagine as skinny with a large moustache), I decided to make the image more like a photograph. So, I plugged the Artbreeder image into MidJourney and asked for the image to be made more 'photorealistic'.
And, here, folks, I wanted to scream... Because...
You guessed it.
MidJourney gave her green eyes. It also lightened her complexion and widened her face. Basically, it did everything possible to turn her into a conventionally-attractive, sparkly Celtic fantasy sorceress without expending any originality, creativity or conscious effort.
I don't think the AI is being deliberately racist (AIs don't have brains). I think it's generating what I call 'consensus images'. It's been trained on a lot of images scraped off the internet, and it's picking up the racial biases of what's commonly available on the web.
Now, the MidJourney AI can draw sort-of biracial people (admittedly, very creepy-looking ones), but I have to specifically ask for them. MidJourney doesn't assume multiracial people in south-east London by default, and the ethnic diversity of the people on display here is shallow (no one here looks British-Chinese-Jamaican, for example).
[As an aside, the children are standing in what looks like a Victorian slum. Most of south-east London just doesn't have narrow streets with small, red-brick terraces. The houses are typically newbuild flats, inter-war semis, or Victorian townhouses with bay windows. So, at best, MidJourney was trying to draw Brick Lane, north of the River Thames...]
But, then, learning the ability to generate local, specific images is not the way art AIs like MidJourney are trained. They're trained to link keywords with images featuring that keyword, so they can generate something resembling the original keyword when it's entered as a text prompt. They don't know anything about south-east London or the people living there.
Furthermore, if the AI was originally trained on keyword-labelled images primarily from America, where interracial marriage remains less common, they're going to have great difficulty creating faces that don't look like ethnic archetypes.
Nor does MidJourney have the ability to generate statistically-unlikely images. If 90% of the images of pilots on the web are male, then the AI will give my female pilot (i.e. Syer Roth) five o'clock stubble.
The most damning thing, though, is that if you enter 'toddler' as a MidJourney prompt (and nothing else), you get 50% ginger toddlers. Yet, as previous, only 2% of the world's population have red hair, and less than that also have pale skin and blue/green eyes. So, MidJourney isn't even reflecting reality here.
Toddlers of West African heritage, folks. They exist.
The programmers of AIs like MidJourney are aware of these biases, but it's not something I expected to encounter in casual use. I assumed getting art AIs to Nazi-out required deliberate effort: maybe by asking one to draw a Third Reich commemorative plate... or maybe a pro-colonialist travel poster.
Social issues apart, it's just a little (no, maybe a lot) frustrating when you're trying to create original science-fiction and fantasy art. Syer Roth is ethnically Usfani. She's not from an Earth ethnic group and I wish AIs would stop trying to make her archetypically Celtic. It's not as silly as the ridiculous bananas, or the murder toddlers, but it is downright useless from a creative point-of-view.
Further reading on south-east London
Finally, if you are interested in science fiction (as I hope you are), and want to find out more about south-east London, may I heartily recommend reading The Upper World by Femi Fadugba, which combines quantum theory and time travel, and is set in Peckham.
You may also want to watch Attack the Block, which helped propel a young John Boyega to Star Wars fame. This is a great film - absolutely hilarious (although Peckham is actually pretty gentrified)!