We need to talk about AI - A digital photographer’s perspective
Reading time - approx. 20 minutes
The conversation around AI has been far reaching, but appears to have begun going around in circles. Both sides of the conversation are rife with misinformation and a lack of understanding of what AI actually is and how it works. I’ve been dedicating a lot of time into researching and discussing the topic as my own livelihood appears at stake in favour of machines. Now I’m no computer scientist, but I have been using digital art technologies for over 15 years, and so this topic is incredibly important to me. I believe Generative AI is doing enormous amounts of harm, and I’m going to do my best, in my understanding, to explain why I think this is. I don’t think we’re even remotely close to a Skynet situation, the AI we have now is still heavy on the ‘Artificial’ and very light on the ‘Intelligence’, that’s not the kind of harm I’m talking about. I want to focus on corporate misuse, and public perceptions primarily. But first, a bit about what AI actually is.
There are numerous applications for AI and machine learning and they’ve been in use for a long time. Every time you’ve clicked on pictures of cars, letters, solved sums etc. to prove you are human, you’ve been training these models to recognise objects, text, and solve problems. Traditional Machine Learning models are how digital assistants such as Siri or Alexa answer questions quickly and efficiently for you. It’s how predictive text learns how you type, even the grammar checker I’m using right now utilises Machine Learning. These machines are given a huge amount of data, and use it to make accurate predictions and relay the wanted information from questions asked. I’m not worried about these models for the large part, we use many of them in our own day to day, and professional workflows.
Generative AI is a different beast altogether. When talking about GenAI, I’ll be talking about Stable Diffusion based models such as Midjourney and Dall-E, and Large Language Models (LLMs) such as a ChatGPT and DeepSeek. GenAI is a recent advancement that uses much larger datasets than traditional machine learning models. Devavrat Shah, Professor in Electrical Engineering and Computer Science at MIT said “The highest value they have, in my mind, is to become this terrific interface to machines that are human friendly. Previously, humans had to talk to machines in the language of machines to make things happen. Now, this interface has figured out how to talk to both humans and machines.” Whilst people have jumped on things like ChatGPT as a supposed fountain of knowledge, its function is essentially to take large amounts of complicated data and information, and to relay it in full, readable sentences.
This, as Shah says, allows us to speak to machines in our native tongue, rather than having to learn computational languages, improving the accessibility of computing in general. It is important to note that the models themselves lack the ability to fact check, discern nuance, or understand bias. Therefore the accuracy of the output is only as trustworthy as the data it has been trained on, and when a model has been trained on unfiltered and unvetted sources from across the publicly accessible internet, it’s likely to give inaccurate results. Indiscriminate scraping of internet results has also led to the inclusion of illegal materials such as CSAM and other sexual abuse material, as well as pornographic material in the training sets. Without total transparency on training data for GenAI, this currently holds massive potential for harm despite their promises to deal with this behind the scenes.
Many big businesses have applauded recent advancements and have expressed a keenness to move customer service interfaces to AI, rather than people. A supposed step up from the many automated bots we interact with today. This application of GenAI is decidedly dicey; negating the need for humans whilst we’re living under a capitalist model that requires humans to work to be able to afford to live, isn’t a viable situation. With the growing number of stories emerging about entire teams being laid off in favour of an AI program, people are rightly becoming afraid of being bumped out of careers they’ve spent years developing the skills for in favour of computers. In an ideal world, I personally believe the solution to this would be a Universal Basic Income (UBI) model, allowing the machines to take over menial human tasks, but still allowing the humans that did those jobs a chance to live a reasonably comfortable life. It’s argued that in a world with fewer jobs but still a livable income, people would be able to pursue passion projects, to volunteer in their community, to help raise young family members and more. However due to corporate greed, and highly stigmatised ideas around self-worth and work among the working class, we don’t seem any closer to that becoming a reality, so how has the situation gotten this out of hand so quickly, with seemingly no good solutions on the horizon?
In my opinion, this technology was released to the public far too soon. There was no safeguarding in place, no contingency plan for what people would do with it, and now we have a number of major issues in the wake of its release. Misinformation is spreading at an alarming rate, sped along by companies wanting to adopt new technologies as fast as possible with no care to the consequences, simply to stay ahead of the curve. ChatGPT comes with many of its own challenges, but in my experience, Google’s AI snippet is doing much more harm. As previously stated, large language models exist to put large amounts of information in full sentences. They cannot discern fact from fiction. Google has become (for better or worse) a largely trusted and respected source of information, and what used to be a snippet of the first result at the top of the page, is now an AI summary.
I have personally witnessed numerous searches of mine coming back with misinformation in the snippet because the model is simply regurgitating any bit of information it deems relevant. I have Coeliac disease and regularly have to search whether something is gluten-free or not, recently the AI summary informed me that puffed wheat is gluten-free, which it took from a website selling grains. Except puffed wheat is literally just wheat that’s been heated up, and is most certainly not gluten-free. And yet, it took this as fact. In a time where seldom few people are doing their own further research, and this coming from a source they trust (google), this could potentially be widely harmful in unimaginable ways.
As well as AI models being used to replace admin and tech teams (to varying success it seems,) and spreading disinformation, artists of all mediums are already losing out on a huge amount of work. Art theft has been a big issue online for a long time, however when your image is stolen, there is still the chance someone will be able to reverse image search it, and still find you and your work. However people are now treating GenAI as a free loophole for stealing imagery due to the ‘transformative’ nature of the program. This is in part due to a long used copyright loophole that allows artists to take inspiration or even direct parts of each others’ work without being legally implicated for plagiarism.
Last year, photographer and digital artist Jingna Zhang (@zemotion), won a landmark case against a painter in Luxembourg who had painted one of her images, entered it into a competition, and won a cash prize. She argued that the painting was so exact to the image it was created from, it could not be considered transformative, and despite the fact he’d painted it, it was a copy of her work. GenAI models have been seen regularly including watermarks and signatures of artists, only obscured by its inability to form legible text. How transformative can the images truly be, if it’s even replicating a version of someone’s watermark, put there to protect the work from theft? Does the fact it’s technically creating new pixels count as transformative? I personally don’t think so, especially where the style recognisably belongs to a living human artist. If tracing an image and claiming it as your own is classed as plagiarism, then how is GenAI any different?
This case was a hugely significant win, and Zhang is now focusing her efforts on research into the ethics and applications of GenAI, I’d hugely recommend following her for updates on this research. Recently, she spoke about a report that has been released showing that some AI models had started to purposefully deceive the user, and overwrote instructions with its own. Given these programs are as I mentioned, not sentient, this information is visible to the researchers and able to be stopped, but it still shows that the technology is not ready to be implemented. Despite this, GenAI truly is everywhere you look now, even typing this article in Google Docs, it’s asked me several times if I’d like AI assistance, now wouldn’t that be ironic. Finding reliable information online is becoming more and more difficult. People are treating Google’s AI snippets and ChatGPT answers as infallible truth, but this just isn’t correct. These text based models are essentially very clever search engines, as mentioned earlier, one of the ideas behind LLMs is being able to talk to machines in our language, rather than the computer’s language.
You can get accurate and detailed search results in most engines if you know the right shortcuts and keywords, much like you can in ChatGPT. However, when you search for this information yourself, it’s on you to verify the information you find and check your sources. As stated earlier, AI models lack this human discernment, it can’t tell the difference between fact, fiction, and satire. For a completely hypothetical example, I might ask ChatGPT to tell me all about apples, and it might tell me that apples are a superfood that can heal all ailments because “an apple a day keeps the doctor away.” That phrase is featured so heavily across English speaking sources, that unless the model was told that it’s just a saying, it might not understand, and may present the information as fact. People are then taking these (often needlessly lengthy,) explanations from ChatGPT, and offering them up on posts, in comments, and in forums, seemingly believing they’re being helpful, whilst not realising they’re proliferating misinformation.
A similar issue is happening with AI image generation. People are so excited by this technology, and the speed at which they can create bizarre dreamlike versions of images they have in their heads. Simply by using either a short prompt, or a long and complex series of prompts. But people are creating and sharing these images at such alarming speed, real artists and real images are being bumped down the search results in favour of AI. Pinterest is a great example of this, my feed is now about 80% AI and is now unusable as a tool to find inspiration from other photographers and artists, when I select an image, I am usually given images that are visually similar. However now, it’s a seemingly never ending string of AI replicas with only a slight difference to each of them, as if the same prompt has been run over and over again changing only one or two words. There are history pages that once stole content now using AI writing and images to make up things that never happened or to alter historical events. There are science pages sharing misinformation, and a seemingly endless amount of scams. These images have absolutely no inherent value. They can’t be used as inspiration or reference to any success because the lines and the proportions don’t make enough sense to replicate, I can’t find what clothes or makeup they’re wearing as a consumer because they don’t exist, and I can’t find the actual inspiration behind the visuals because the AI generator makes that impossible on purpose. They are, essentially, digital scrap, and yet people are wildly protective over them.
Situations are already occurring where searching for the work of a current living artist, instead brings up pages and pages of AI reproductions. This is potentially costing artists opportunities to be found and hired. This training data was largely acquired illegally, the team uploaded any large database of imagery and the associated artists’ information they could find, into their models’ training data. Supposedly, beginning with the massive Magic: The Gathering archive. I noticed that many of the patterns seen in early AI, particularly the swirls and fractals it often made, are similar to repeating motifs found in the architecture and machinery in the world of MTG. This is largely the work of concept artists, whose entire job is to create things no one has seen before based on a description, this is what people think GenAI is doing. However the idea is not the art, the amorphous blob of an idea I see in my head is quite similar to the look of AI, but if I could take that blob and put it on a page it wouldn’t make it a finished product. The artwork still needs to be made. And whilst an artist can use all they’ve seen and learned to create something brand new, a computer can only approximate based on what it's already seen.
Stable Diffusion started its life as a denoising tool, used to take pesky noise (a digital artifact caused by camera sensors not being able to keep up with low light) out of images. It uses this process to build up images from prompts pixel by pixel using its large database of images that have been encoded into datasets comprised of their attached keywords.
The best way to imagine this is to picture a photograph of an apple on a table. This image has been uploaded online and contains metadata that tells people who the photographer is. Let’s say it’s me, so it has a tag to attribute the copyright to ‘Isobel Greenhalgh’. The image also contains alt text, this describes what is in the image for people with sight impairment, their screen reader reads the alt tag as an audio description of what they’re seeing. The alt text reads “a red apple sits on a light brown wooden table in a white room”. So now let’s feed this image into the AI’s training data. It’s already seen lots of images of apples and tables, so because both have been mentioned, it knows to file those parts into “apple” and “table”. However the image doesn’t stay an ‘image’ in order to do this, it’s broken up into raw data that’s smaller and easier to store. This data is attached to those keywords, it would also sort the table under “wooden” and “light brown”, and the apple under “red” etc. Now when someone asks the model to create an image of a red apple on a light brown table in the style of Isobel Greenhalgh, the image may automatically be in a white room, despite this not being asked for. It will use everything it’s seen about apples and tables to stitch together an image that looks roughly like what you’ve asked for, using the data it already has. However since it cannot create what it hasn’t seen, what you’re looking at is a Frankenstein's monster of an image comprised of an unknown number of existing images.
A GenAI image generator doesn’t actually know what an apple is, it’s simply programmed to recognise the patterns that make pixels look like an apple based on what it’s seen. GenAI cannot create anything that is not already included in its training data. The best example of this was early attempts to get AI to render an image of a centaur, except it had never seen a centaur before, so it kept spitting out strange horse/man combinations that were never quite right. There was no way to get the AI to create something that did not exist within its training data. This also means that if you were to remove the training data, it would lose the ability to create images entirely, the dataset is what it creates the images from.
This, combined with the number of artists’ watermarks that have ended up minced into images proves to me without a doubt that GenAI is not creating anything new, and cannot be treated as such. However we should also be treating what it is creating with deep suspicion. If you see something made with AI that depicts inappropriate material, it’s because it’s included in the training data. AI companies must be held accountable for this and be forced to be discerning and transparent with their training data. Right now the UK government is preparing to remove copyright protections from artists and creators alike in favour of tech giants that want the right to include any publicly accessible material in their training data. This is a massive breach of our rights as creators and sets a dangerous precedent to the viability of all future creative careers. We cannot allow ourselves to be replaced by machines, but we also can’t allow these gigantic companies to get away with using our labour, and our life’s work, to make themselves richer. This is being done at our expense to benefit a tiny proportion of the population, and it’s not just costing our rights, but our environment too.
The environmental impact of AI is one point that is talked about regularly so I only wanted to touch on it. All computations require computers, the bigger and more complex the computation, the bigger the computer and the more resources it needs. Computations also require data, and whilst digital, it does need somewhere physical to store it. The cloud isn’t somewhere that only exists online, it just means it's not being stored locally to you. Cloud data is stored on servers, imagine a harddrive but the size of a computer tower, and when you access your files, you are accessing them from that server. The more data, the more servers you need to store that data as each one can only hold so much. This leads to rooms filled with servers all stacked on top of each other, and these are housed in data centres that are filled with these rooms that are filled with servers. As these servers are on all the time, they use a massive amount of electricity. However they also generate heat, I’m sure you’ve experienced a hot laptop on your lap with the fans whirring whilst you try to play minecraft with 15 mods. But this amount of heat also requires an enormous amount of cooling, which is done using cold water. So now you have all these data centres using huge amounts of electricity and water, and they’re still building more. This is quickly becoming unsustainable for the communities in which they’re housed, such as Ireland which is home to many tech company offices and data centres due to the low corporate tax rates.
A lot of the arguments for and against GenAI are becoming heated and emotional on both sides. Whilst I can see the potential for application of the actual technology being used in art where the entire training set can be controlled by the artist. And I also recognise there are many out there trying to make it more efficient and environmentally friendly. I cannot support it in its current form and I intend to continue rallying MPs for tighter legislation to protect the livelihoods of artists, and to stop corporations using it to devalue their workforce. I understand new technology is exciting and feels like progression, but if it's only progressing the wealth of a handful of people, how can any of us actually benefit? Please reader, don’t stop making art, don’t stop creating, the act of creation is one of the beautiful things that makes us human, and allows us to connect with each others’ experiences. They can never take that from you, though they may try. If they take everything, I’ll still be drawing pictures in the mud with sticks.