The development of GAN (Generative Adversarial Network) model, DALL-E, Contrastive Language-Image Pre-training (CLIP), and Diffusion Model, plays a huge role in developing the text-to-image art generators. OpenAI is one of many companies that are currently developing these futuristic features, including their newest DALL-E 3 text-to-image generator that can translate nuanced requests with prompt texts into extremely detailed and accurate images. There are a lot of different types of AI generators (e.g. ChatGPT as a text generator) but now let’s focus on the AI text-to-image generator because a lot of people don’t really know the existence of these features, but it’s affecting a lot of creative workers (illustrators, graphic designers, etc).
As aforementioned, text-to-image art generator has a lot of different models and it’s developing as time goes by. One of the most developed and commonly used by OpenAI, MidJourney, and Imagen is the Diffusion Model. In understanding the IP protection aspect of the AI-generated image arts, let’s dive into how the Diffusion Model works generally.
After the user inputs the prompt text, the first step is the training data, which is simply a dataset of collecting all images based on the alt texts of these images scattered throughout the internet
When all word descriptions have been collected in the dataset, the next stage is through deep learning. AI tries to classify the prompt texts according to the visual concept that best matches the prompt text, this process is simply like teaching AI to mix and match millions of inputs in the form of algorithm code and this process is carried out by AI in the form of deep learning.
The next stage is latent space. All the variables that have been collected in the deep learning process then begin to be formed in the latent space.
The simple illustration of the process in the latent space/Dokpri
Keep in mind that the variables collected by deep learning obtained from the dataset are in the form of pixels, so in the latent space, there is a mathematical space that can determine whether the text description inputted by the user has a round/square shape, green/orange color, or has a shadow/not, and other indications that indicate 1D/2D/3D or even more than three dimensions (multi-dimensional latent space). The illustration of the latent space above is the process where each pixel-shaped variable is measured for its corresponding position.
After the AI forms pixels with the formula according to the prompt text instructed by the user in the latent space, the pixel in the form of a mathematical matrix is converted into an image, this process is called diffusion.
It’s all very technical and seems to be ‘confusing’ for people who are not familiar with AI processing. However, what we can conclude is that the user (as the human) is not involved directly in the process of the AI-generated image art. Users input the prompt text and wait for AI to process it into an image or an illustration as the user requested. The question is: regardless of the prompt text by the user, can we say that the user is the creator/originator of the art?
The question regarding the authorship of the artwork determines the possibility for such artwork to be protected by copyright law or not. In Regulation Number 28 of 2014 on Copyright (“Indonesian Copyright Law”), the answer is very simple and straightforward: the image art created by AI cannot be protected by the Indonesian Copyright Law, because in order for the works to be protectable, it has to be created by the “natural” author.