In the late 18th and early 19th Centuries, as the First Industrial Revolution began dramatically increasing production of items such as cotton clothing, a movement of textile workers and weavers arose which protested against new technology and machinery that they believed was threatening their jobs, lowering wages, increasing unemployment, and reducing the quality of products.

The Luddites, as they came to be known, were so named after Ned Ludd, a mythical figure who was said to have destroyed two stocking frames in 1779 in protest against the mechanical knitting machines that were being introduced at the time.

The emergent Luddite movement responded to rapid changes that technology brought to the economy and society through staged protests and attacks on factories and machinery, sometimes destroying the equipment that was seen as a threat to livelihoods.

The Luddite movement eventually declined in the mid-1820s as it was impossible to reverse that course of innovation and the economic imperative to adopt these technologies drastically outweighed the concerns of resistant workers when compared to the progress of society as a whole.

Imagine if we had stalled the development of steam power to protect those textile workers, the consequential impact on modern society, trade, standards of living and science is unfathomable.

The term “Luddite” has continued to be used to refer to people who are opposed to technological innovation or who are resistant to the systemic change caused thereby. In the years since, we have seen many similar movements arise whenever innovation makes prior systems of labour or revenue generation obsolete. Contemporary examples include the opposition by metered taxi operators to Uber or coal truck drivers to renewable power generation.

Enter the rise of artificial intelligence (AI). In the last few months, the opposition to AI on the basis of it making existing means of work redundant has entered mainstream discourse in part due to the meteoric rise in popularity of text-to-image AI. These AI can turn words into drawings, or create fantastical images of one in exotic locales, all within seconds.

Very quickly, an outcry from digital artists arose, centred on two points: that the AI plagiarises the work of human artists and that people should rather commission actual human artists than pay for an AI app.

The arguments surrounding plagiarism focus on how the AI is trained and generates output images. So it is useful to first establish how this occurs. This is an incredibly complex and advanced area of computer programming, but to try and explain it as simply as possible, the current text-to-image AIs are based on generative adversarial networks (GANs).

Text-to-image GANs are a type of machine learning model that can generate images based on a given text description. They are made up of two neural networks — simply think of them as two separate computer programs which interact with each other: a generator that creates the images and a discriminator that tries to distinguish real images from the generated ones.

To train a text-to-image GAN, you need a dataset of images and corresponding text descriptions. The training process involves using the generator to create images based on the text descriptions and then presenting both the generated and real images to the discriminator. The discriminator tries to identify which images are real and which are fake, while the generator tries to create images that can fool the discriminator.

As the training process continues, the generator gets better at creating realistic images and the discriminator gets better at distinguishing real from fake images. This process is repeated until the generator is able to create images that are almost indistinguishable from real ones.

An argument which has been advanced against these AIs is that they are simply splicing together existing images using the text descriptions, like some sort of advanced Photoshop. This is untrue.

The starting point for the generator with every image is random noise, basically an image which would look like static from an untuned CRT TV. From there it attempts to find patterns in the noise which resemble the patterns it has associated with a given text description. But how does it determine what those patterns are?

To provide a very simplified example, let us say we ask it to draw a cat. To do so we need to take an image of a cat and provide it to the discriminator model and say to it, this is a cat, now try and determine from this image and a bunch more real images of cats, whether the images from the generator are a cat or not.

The generator then submits its random noise which the discriminator easily identifies. The generator then changes a few parts of the image and submits again. We then repeat this thousands of times over until the generator has, through trial and error, refined how it creates cats from the random noise, to fool the discriminator into believing its images are actually of a cat.

This, I would argue, is no different to how a human artist learns to create art — over years in childhood, learning what a cat looks like by being shown images (or real life) cats and told these are cats, and from that attempting to create depictions of the cats, and over time refining their ability to create, from nothing, a depiction of a cat in a chosen medium.

Further, this often involves referring to reference works, depictions by others of cats, seeing how they drew/painted/sculpted etc, the cat. Artistic creation is fundamentally an iterative process, iterating both on one’s own work and on the work of other artists.

The argument for plagiarism in AI art which has gained the most traction comes from the provision of existing reference works to the discriminator. As described above, we need to tell the discriminator what a cat is by providing existing images of a cat taken/created by a human.

Visit Daily Maverick’s home page for more news, analysis and investigations

Many of these artists, so the argument goes, did not consent to their art being used to help train these AI, to which it must be asked, is their consent required?

Remember, the existing works are not being used in any way except to help the AI understand what the text appears like in an image form. It is not being spliced together nor reproduced, but instead is being used to create the conceptual framework upon which the generator interprets patterns from random noise to align to the text prompt.

This can range from the vagueness of knowing that many artistic images have a signature in bottom left-hand corner (which AIs often try to replicate, adding their own signature of sorts to the artwork as that is what convention dictates for those styles of artwork), to the specific such as the artistic style of a single prolific artist.

This, I would argue, is no different from an art student studying the works of Rembrandt to gain an understanding of his use of light and shadow, or Picasso for his use of colour and form to convey meaning and emotion, and then replicating those features in their own artwork or using it to better inform their own work. Neither of these artists gave consent for their artwork to be used in this way, nor was it needed.

The reality is that once an artist makes and displays their artwork, it is out there for the world to study and critique. Whether this is done by a human artist or a computer one, is in my mind, irrelevant. To snip the commercialisation argument, human artists are equally empowered to go on and sell their works.

Fundamentally what informs the output of the AI is not really the images which were fed to it, but the text prompt used by the generator to guide the patterns into which it should curate the random noise.

Hence, if I ask it to create The Scream by Edvard Munch, it would give me an image which looks strikingly similar, but not identical, to the original artwork because it is pulling from those patterns. If I asked an art student to do the same, I would also get an image (depending on the skill level of the student) which would also be strikingly similar, but not exactly the same as the original.

Here the potential plagiarism is not from the ability to create the artwork, but from the prompt given and the intention of the use thereof.

The important distinction which needs to be made is that the tool itself does not plagiarise, but it can be used for plagiarism. The personification of AI means that the above distinction is often missed.

The broader conversation to be had here then is not around the tool plagiarising, but rather about users using it to plagiarise and how we can gain the benefits of the technology while limiting nefarious uses thereof. This is a conversation relevant to all technology, but very rarely does it result in the resolution to not utilise the technology at all.

The second critique of rather commissioning a human artist is perhaps more genuine to the concerns at hand. The concerns of these artists are rooted in fear that their livelihoods will no longer be viable, their skills no longer of value or more profoundly that art itself will lose its meaning or value due to no longer being human-created.

I can sympathise with the fear. Being a lawyer, I am already seeing tools coming which are dramatically altering our workflow and can easily envisage a near future in which clients inquire from an AI how to address their legal issues instead of me. As with the Luddites, this fear itself is not justification for the abandonment of promising technologies.

It is also important to note that new technologies often create new job opportunities and industries that didn’t exist before. Text-to-image AI tools will be able to augment rather than replace human art practices. They may allow artists to work more efficiently, freeing up time for them to focus on more creative and complex tasks.

We are already seeing this manifest through an explosion within the VFX industry of better quality VFX shots created with fewer resources. This has lowered the barrier to entry for independent artists to produce amazing work which has in turn exponentially increased high-quality visual media being created and shared.  

As for the concern that art itself will lose its value or meaning if it is no longer solely created by humans, one must recognise that art has always evolved alongside technological advances. From the invention of the printing press to the rise of photography, art has always found new ways to incorporate and respond to new technologies. AI text-to-image tools are simply the latest example of this.

While they may challenge our traditional notions of what constitutes “authentic” art, they also offer new opportunities for artists to push the boundaries of creative expression. Ultimately, the value of art lies in its ability to evoke emotion and inspire thought, regardless of the tools used to create it.

In conclusion, the concerns raised by artists about the use of text-to-image AIs are reminiscent of the fears of the Luddites, who opposed the use of mechanical looms out of fear for their livelihoods.

Just as with the mechanical looms, text-to-image AIs are here to stay and will continue to progress and improve in ability. Like Pandora’s box, the technology has been opened and cannot be shut again.

Text-to-image AIs do not plagiarise existing works and have the potential to augment rather than replace human artists. While it’s true that new technologies can sometimes lead to economic disruption, they also create new opportunities and industries that didn’t exist before.

Rather than lamenting their existence or trying to stop their progression, it’s crucial that we embrace these tools and have a productive discourse about how they can be used to enhance the art world and expand the boundaries of creative expression. DM