What is AI art?
Before we break down how art comes into play with artificial intelligence (A.I), let’s start by defining A.I. John McCarthy, a renowned computer scientist defines A.I as “the science and engineering of making intelligent machines, especially intelligent computer programs” (2004). He adds on by saying it’s similar to using computers to understand human intelligence, but A.I doesn’t have to limit itself to methods that are possible by human standards. How this is possible, is through a process called machine learning (Smith, 2019, p. 47). Put simply, the A.I is made up of various layers of processors called neural networks, that work to receive an input, break down the information in its hidden layers and produce an output.
People such as computer scientists and statisticians often feed massive amounts of sample data (otherwise known as “trained data”) so it can learn and “recognize” mappings between certain variables (Smith, p. 50). Without being fed data produced by us humans, there is no way for machine learning and A.I to properly function. Machine learning in A.I is not omnipotent - it can make classifications, calculations and predictions based on statistics and patterns (Coleman & Allin 2021). It cannot predict the future, produce unheard knowledge or give us the next winning lottery numbers, but give us interpretations based on data that reflect views that already exist in the world (Coleman & Allin 2021).
By now you might’ve gotten a sense of where A.I has its weaknesses. But to outline them, A.I has three major limitations: algorithmic bias, the quality of the data and its black-boxed nature (M.P et al). Algorithmic bias is an unavoidable part of the recognition process of A.I. Algorithms, which falls into machine learning, is a set of instructions that machines follow to complete a task. If the algorithms are considered faulty or biased in a way that the results in the task are “unfair”, then depending on the context in which the results belong, it can be considered unreliable (M.P et al, Allen 2020). More on algorithmic bias is explained in the section Ethical Dilemma.
As mentioned previously, A.I are fed the data that we provide it. Though if the data is incomplete or inaccurate for the algorithm’s task, then it impacts the outcome - “AI can only be as smart or effective as the quality of data you provide it” (M.P et al). Perhaps before viewing this article, you may have thought that A.I was like a “black box”, a system where a person puts something in, the machine does something super high-tech in secret, and it spits something out. Because the box is “black” and not transparent, no one can tell what’s happening inside. For many of the general public, this is true.
​
​
What you’ve read so far is only an introduction to the complex world of A.I and machine learning. How it occurs in the industry is limited by the accessibility of code and functions which is often the private property of companies (Delfanti 2022). That’s why for many A.I systems, it cannot easily tell how it arrived to its conclusion. Despite all these weaknesses, it can still be a powerful tool depending on its user.
​
So, how do you create A.I generate art? Otherwise known as algorithmic art, it begins with training the algorithm with data, in this case, it’s a collection of images (often thousands) (Hencz). The user, or artist looking to create art, would have to give the algorithm instructions on the desired art style and aesthetic. This is often in the form of typed prompts, detailed code or selected themes (Elgammal). The A.I then analyzes from its database of images and based on the given information, it generates images that mimic the forms, shapes, figures, colour, text and patterns that best suit the prompts it’s recognized (Elgammal). This, in essence, is the machine learning process of creating A.I generated art/algorithmic art.
However, there are currently many types of A.I that produces art with slight variations of this process. Firstly, there’s the General Adversarial Network (GAN) which is a system that operates with two main components: one that tries to produce images and another that tries to filter out images that are not original and that best align with the prompts (Dupelessis 2022). An example of this is the VQGAN+CLIP system. There is also the Convolutional neural networks (CNN) that were made popular by Google’s A.I program, DeepDream (Dupelessis). It was invented to help scientists and engineers to visualize patterns learned by neural networks, but when turned towards artwork and images, it enhances the patterns to the point that it creates an abstract and psychedelic effect (Dupelessis).
​
​
​
​
​
Lastly, there’s Neural style transfer (NST) which takes an inputted image and reproduces it in the style of another - like the anime filter trend in Tiktok (Dupelessis).
A.I art is a field that is always growing and changing to meet the demands of artists, computer scientists and the everyday people that help to define what is art. Its rise in popularity has also led to many issues of misuse and misinformation. Arti wants to provide all those who visit the opportunity to learn and have fun, free of charge and free of bias - all for a more ethical future for A.I in Art.
​
