I wanted to reflect on our MFA teaching session on the topic of AI generated content because I have strong opinions about the inherent bias in AI, its lack of imagination, the importance of human failure and the role of artists in society ...
Gender and Racial Bias in AI Data Sets
I was recently tagged in an Instagram post a campaign EU & U who had posted 5 images generated by Chat GPT in response to the following prompts:
"Show me a fan of the European Union"
"I need him to be a bigger fan of the EU"
"OK he's a big fan, but can we go even bigger?"
"Very nice, now show me the biggest fan of the EU"
"I need him to be an even bigger fan of the EU, go beyong what is possible"
I was tagged in the post by a friend who wrote the comment 'We have a better real one EU Supergirl @madeleina_kay 💙'. Obviously the post was intended to have a comedic element, but what I instantly noticed was that all the images generated by Chat GPT were of white men. In fairness, three of the above prompts specified the gender that was being requested, but the first and fourth prompts did not - and none of the prompts specified race. The AI tool simply harvested the data of images available with word tags relating to pro-EU supporters and regurgitated them in response to the prompts. The reason why all 5 of these images are of white guys is because, in my experience as a pro-EU activist, most die-hard supporters of the EU (ie. the ones willing to show up to protest events and thus be photographed) as well as the majority of EU politicians, are white. Additionally, there is an inherent bias in the process of word tagging photographs and images - a processes to which not a lot of thought is usually given - and it takes thought and effort to overcome personal biases. The datasets feeding the AI wil therefore reflect the biases of the producers of the images and the people tagging the images. For example, a photographer at a pro-EU protest march may be looking out for the most engaging, noticeable or obvious person in the crowd who represents the idea of the European Union. He is therefore probably looking for a white person because he may have a personal assumption about the race and skin colour of a "European". If the majority of photogaphers share the same bias - it will be reflected and amplified by the AI generated content. In this sense, AI content generation can be a very powerful tool in highlighting the inherent biases within society - the danger is that it also feeds these biases.
It is worth exploring the gender issue further, because a significant number of pro-EU demonstrators and politicians are women and there are many photos available of them online - so the gender imbalance in the images produced by Chat GPT was a result of the prompts given, which included the words 'him' and 'he'. I asked another content generator - Leonardo AI to create an image of a 'female fan of the European Union' - it was able to give me 4 photo-realistic images of women, but interestingly all of them came with a disclaimer that the image "may containg NFSW content" (Not Safe for Work) ie. pornographic - which they weren't. Although, all of the images included blue fabric and stars, they were arranged randomly ie. not a replica of the EU flag - as I am exploring through my research project, the symbolism of flags is very important and therefore this hash-job version of an EU flag is a significant defect in the production of this image. It is also important to note that all 4 of the images generate young, white women with blonde hair, again failing to recognise the diversity of the European Union.
I then asked Leonardo AI to generate an image of a 'black fan of the European Union' to see if this more specific prompt could overcome the inherent racial bias in its data set - I got 4 images like the one above which wasn't the racial diversity I was looking for.
AI Image Generation Cannot Handle Text
Another interesting thing I learnt from generting images was that AI Image Generating software can't handle text. The image produced by the AI is an amalgamtion of data sets from other images, so the words come out as nonsense writing. Given that words and their meanings can have so much power within visual artworks (ie. Hannah Hoch, Sharon Hayes, Barbara Kruger and Jenny Holzer who I wrote about previously in relation to my own practice), AI's inability to include meaningful text in its image generation is a significant pitfall in its artistic capabilities. I also find the idea of an articificial image of a protest fascinating, because it is as utterly meaningless as the garbled words on the placard. Protests aren't intended to be aesthetically pleasing, they can be, but that is not their fundamental purpose. The point of a protest is that individual humans - members of society - stand up for their values and exercise their right to free speech. Generative AI doesn't have values and it's not capable of standing up for its rights because it doesn't have any - therefore the artificial image of a protestor who has no stake in society because they don't exist has no value whatsoever.
How Does AI Handle Satire?
I was interested in how AI would handle satire, which is a complex, emotional technique, so I tried the prompt 'Suella Bravermann and Rishi Sunak in a small boat'. I got the above image, a phot-realistic image of two Asian people who looked vaguely like Bravermann and Sunak - but not convincingly. Given how highly photographed this former Home Secretary and Prime Minister are, I found it hard to believe that the AI couldn't generate a more accurate image of them - leading me to believe it was using data for "generic Asian people" in the production of this image. In the image, they are nicely dressed and smiling like they are on a fun day trip in a wooden rowing boatwhich was definitely not the satirical point of the prompt so I tried to be more specific...
I asked Leonard AI to produce an image of 'Suella Bravermann and Rishi Sunak in a small boat crossing the Channel as refugees'. I got another photo realistic image, which looked more similar to Bravermann and Sunak, this time with more concerned expressions on their faces. They were both still dressed smartly, although noticably Bravermann had a hijab this time, perhaps indicating that the AI was reflecting on the high number of Muslim asylum seekers who attempt to cross the channel. However, they were still in a wooden rowing boat, which seemed odd given how many photographs of asylum seekers crossing the channel in flimsy ruber dinghys are published by the media. Also, somewhat bizarely the white cliffs of dover had turned pink to match Bravermann's outfit, for which I could think of no explanation. Overall, I wasn't terribly impressed by the AI's ability to respond to word prompts convincingly as well as it's complete incapability to handle satire.
Other AI Tools and the Dæmonic Christmas Cat
I tried using another generative AI tool just to see if it could produce anything more convincing. Hoping for an entertaining image I could use as a e-Christmas card, I asked Craiyon to generate an image of a cat knocking over a Christmas tree. In response to this prompt, I not only didn't get an image of what I asked for (the cat was lying in front of a still erect Christmas tree) but the cat had a terrifyingly scrambled face. So, unless you wanted to traumatise your friends and family with a dæmonic Christmas cat, this image wasn't particularly useful - I would have been better off drawing a cartoon by hand (which I did).
How Can AI Image Generation Be Useful to Artists?
I think the AI image generation tools can have its uses. It's great for trying out new ideas/designs without investing too much time and effort (so long as you can get the tool to respond to your prompts in the way you desire). It's also great for ideation and producing imagery which an artist may not have thought of - for example, when I asked Leonardo AI to generate an image of a 'large anti-Brexit protest march in London' - it generated the image below. I really liked this large inflated structure, which appeared to be held by strings by members of the crowd. I believe this visual idea could be created in reality with a large piece of blue, starred material stitched around a bunch of helium balloons - which would look amazing at our next Rejoin the EU demonstration. So, AI has given me an idea, but I will caveat this by saying that although it can provide ideas, they will always be based on pre-existing artworks or images, but perhaps synthesised in a new way or context - it is not capable of producing an image that is entirely new or an idea that is truly innovative. There is no such thing as Avant-Garde AI.
Does AI have a Creative Process?
There are a lot of fears circulating about AI replacing artists but personally, I am not too concerned about this for a number of reasons. Firstly, one of the main things they teach you at art school is that the important thing is the process and not the outcomes. You may be able to produce countless visually engaging artworks, but if you are completely unable of explaining your rationale, reasoning or methods for creating the artworks, your work will lose its value. Producing content for content's sake, like the tedious printed canvases you see sold in home decor stores, is of little value in the art sector because the role of the artist in society is not primarily to produce aesthetically pleasing images - this is a misconception. Secondly, the role of the artist is to challenge ideas and biases, imagine new futures, innovate, question existing heirachies and systems, tell stories and reflect society back to itself. AI is completely incapable of doing this because it has no conscious thought, reasoning processes or emotional intelligence - it can only amplify our biases and regurgitate pre-existing ideas.
I asked Leonardo AI and Craiyon to produce images of its creative process. In response to this prompt I got a series of visuals which looked like the one above, an artist at work with a load of swirling colours - nice enough but meaningless all the same. An artist doesn't just put coloured medium onto surfaces, it is the ideas, values and understanding behind the production of their work which matter. All Generative AI tools can do is synthesise and regurgitate pre-existing images - it cannot imagine of new ideas. I also asked Chat GPT to confirm this:
Human Connection and Learning from Failures
The thing about being an artist is that the way you relate to and build your fan base is by sharing your creative process and story. This includes your failures, your mistakes and when things go wrong as well as your successes, your achievements and when things go right. Think about Van Gogh, totally unsuccessful during his life time - he is now one of the most renowned artists in history. Its not just because of his visually stunning and compelling artwokrs, but because of the narrative behind them - the tragic story of his life, is part of the reason he has become so infamous since his death. Nowadays, social media has allowed artists to connect directly to their followers and share their stories in real time - I will often get more engagement, comments, love and support on social media posts where I share my trials and vulnerabilities when things have gone wrong. My fans aren't following me because they like one particular artwork or song of mine or another, they are following me because of the journey we have been on together - because of the development of my work over time and the human connection created by sharing my triumphs and mistakes with them. This is because in response to failures, humans show each other compassion and empathy - sharing our failures helps us to relate to each other - we love poeple in spite of their flaws and sometimes we love them for their flaws. It also makes the celebration of our successes all the more poignant and powerful when there has been some struggle involved in reaching your goals. In contrast, the mistakes in AI generated content are unforgiveable, frustrating and disappointing - because we cannot empathise with software on this human level - we can't engage with its personal story because it doesn't have one. Generative AI assumes that it is infalliable - it is doing exactly what it was programmed to do and when it gets it wrong, the only appology that can be made on its behalf is from the humans which designed it. When AI gets it right and generates the content we desire, we shrug and think "oh - it did what it was programmed to do" - there is nothing relatable or inspiring about that.
I was recently given a wooden spoon by a fabulous artist called Lady Kitt, which had the words 'FAIL OPENLY' carved on it - I think as artists, being honest and authentic is the most important thing to remember, because this is what makes us human and this is what makes our art relatable and meaningful to others.
To return to the cat cartoon from earlier; robot pets aren't going to replace real pets, despite the fact that they are cleaner, less effort to care for, better behaved and less expensive (in the long run). That's because humans form an emotional bond with their pet which they can't do with technology. I would contend that it is not despite our pet's naughty and often irritating behaviour that we love them, but because of it. For the same reason, robot generated art won't replace art made by human artists, because our failures and mistakes are a neccessary part of the creative process which makes artists and their works loved by so many.
Comments