In barely half a year, Artificial Intelligence has gone from drawing its first decent piece of art to multiple competing and open-source pieces of software that challenge the best of artists. But how did we get here?
AI art could be said to have started with the perceptron, invented in 1943 and first implemented in 1958. The perceptron is a portmanteau of perception and neuron, fittingly, as it is an artificial neuron – with artificial intelligence, a neuron is just something that computes outputs from an array of inputs – that has the ability to ‘see.’ It does this by analysing a grid of colours in a manner too complicated to explain here. Over time, machines learned to differentiate, say, cats from dogs (Lefkowitz).
Fast-forward a couple of decades, and AI can classify hundreds of thousands of different things, from breeds of dogs to traffic signs (Wei). However, going the other way – from describing images to generating them based on description – turned out to be a lot harder.
The best tool AI artist programmers had in creating realistic art were GANs – Generative Adversarial Networks – built in the early 2010s and popularised around 2017. These consist of a generator, which tries to make art that looks human, and a discriminator, which tries to decide whether given art looks human. As the generator produces better and better art, the discriminator gets better and better at critiquing it (Deacon).
In 2020, however, the world was shown the power of Diffusion Models. This seminal generative technique, inspired by thermodynamics, consists of continually adding noise to images and learning how to reverse this diffusion (Weng). This is the preferred technique by multiple fantastical AI models developed in the 20s, including Google Imagen, MidJourney, OpenAI’s Dall E 2, and Dream Studio. The latter is based on the Open-Source software Stable Diffusion. Open-Source means I can download the software and, on my PC, generate professional-grade images.
As expected, professional, especially digital, artists were in an uproar. If an amateur AI artist can write, ‘In the style of …’ plus any famous artist’s name and create artwork comparable to the original, then what foothold will professionals who have worked hard to express themselves have in society?
Furthermore, these new pieces of software can be used unethically. For example, the development of related deepfakes has been used to nonconsensually humiliate and harass women, and even more disgustingly, children, with indecent material. In response to this, the release of new software is often in stages. Blocks against words such as “preteen” and “teenager” have been put in place, as well as ToS’s generally prohibiting images of sex, violence, and public figures. The use of photorealistic faces is also frequently debated (“AI Can Now Create”).
We have developed an arsenal of AI weapons designed to fake content: Art generation, deepfakes in posts, and even live superimposition (Corridor Crew). AI music, voice imitation, and text generation are capable of tricking Google researchers. And what’s more, it’s getting better, and better exponentially.
However, there is still one part of the human mind we can still reassure ourselves won’t be conquered by Artificial Intelligence. Or at least not yet. As fascinating and as frightening as AI art can be, there is often something that the machine fails to produce: it lacks originality. The google text generator that convinced a google researcher it was sentient? It’s called LaMDA, and all it did was pass the Turing test by imitating through pattern recognition (“The Google Engineer”). The same goes for AI art; all it does is try to match existing works. This isn’t to say that there will never be an original machine mind, but to get to that point there have to come at least a few seminal papers in AI research. By the current rate of improvement, that could come in only a couple of years.
Right now there is an arms race in AI research, and just like the nuclear one decades ago, it’s leading us to a revolutionary territory, drastically altering the way the world works. Yes, it’s dangerous, unknown, and urges ethical premises to be rewritten. But a step into the unknown is a step into the future, and this is a rather exciting prospect, isn’t it?
By Eskil