Introduction: The use of AI in creative works

In recent years, Artificial Intelligence (AI) has rapidly evolved from a subject of science fiction films to an integral part of everyday life. It has transformed various industries, including healthcare, finance, entertainment, and transportation. A fascinating and possibly understudied area of AI is its collaboration with human creativity. AI tools are being used more and more by artists, musicians, and writers to enhance their work, streamline processes, and explore new artistic frontiers.

The emergence of AI in the creative realm is not without controversy. While many are excited to explore AI’s creative potential, which allows everyone to produce various forms of creative works, many have raised ethical concerns over its ability to mimic, or worse, replace human creativity. Sceptics argue that AI's forays into the arts could signal the end of human creativity and lead to a time when machines take the place of artists. Optimists, on the other hand, see AI as a powerful tool that can complement and amplify human creativity. The truth, as it often does, is likely to be somewhere in between.

This article will first explain the fundamentals of Generative AI which uses AI in creative works. It will then discuss the debate over creative AI.

The building blocks of Creative AI

There are three basic building blocks to generative AI that are used in creative works: “Text2Text”, “Image2Image” and “Text2Image”.

“Text2Text AI” models like ChatGPT, QuillBot, and WriteSonic, hinge on ‘tokens’ – units that can be words, partial words, or single characters. Text is first converted into tokens, and are subsequently processed and analysed. Let’s consider how it works in the next paragraph.

After tokenisation comes ‘embedding’, where each token becomes a high-dimensional vector that represents the token's meaning. Then, contextual analysis happens as the AI model processes tokens sequentially, much like how a human read word by word. The model employs artificial neurons and interconnected layers to understand relationships, identify patterns, and predict token sequences. Next comes the fun part, namely ‘generation’, where the model predicts the next token based on context and selects the most likely one, continuing this process until a full sentence or paragraph is formed. Finally, ‘detokenisation’ brings the tokens back to human-readable text.

So, what is the purported “magic” behind AI models like ChatGPT? It is their ability to swiftly process massive volumes of text data. They learn to recognise the patterns and nuances that make up human language, allowing them to generate coherent, contextually relevant, and surprisingly human-like texts – a testament to the power and potential of machine learning.

"Image2Image", as the name suggests, involves transforming one image into another. It is not just about applying filters or adjusting brightness and contrast; it is about fundamentally altering the contents of an image, creating or modifying its elements, or even generating entirely new images from scratch. Among the plethora of techniques available for image transformation, Stable Diffusion stands out for its effectiveness and versatility.

Drawing inspiration from the phenomenon of diffusion observed in physics and chemistry, it injects controlled noise into the original image, much like adding unexpected colour to a canvas, leading to creative possibilities. Next, the model undergoes ‘iterative diffusion’, which is comparable to ink dispersing in water. It redistributes pixel values, reshaping the image in a high-dimensional latent space. In this high-dimensional space, the model can manipulate the image's essence like a sculptor. ”Reconstruction”, translates the transformed latent space into a unique image. It is the culmination of the creative process, akin to the final brushstrokes of a painter, revealing a unique and transformed image

The final building block, "Text2Image", is a method for creating visual representations from textual descriptions that borrows ideas from "Text2Text" and "Image2Image" models. Tokenisation and embedding capture semantic meaning of the words and provide a basis for the AI model to understand the relationships between them. Here, the AI model makes use of multiple layers of artificial neurons that resembles human brains to provide contextual analysis that takes into account the relationship between the text and visual elements, recognising patterns and forecasting visual elements. In the next step, generation, iterative redistribution of pixel values, namely the diffusion process is guided by the contextual analysis to create a visual representation in the latent space after a controlled noise is introduced into a canvas. Finally, reconstruction transforms the latent space back into a tangible image, bringing the textual description to life.

What is creativity?

Before we delve into the debate around creative AI, let us pause and consider what exactly is creativity? While it is an elusive concept, we will discuss it using the debate around creativity and originality. Cognitive scientist Margaret Boden defines creativity as “the ability to come up with ideas or artefacts that are new, surprising, and valuable” in her insightful book on human creativity and AI. In the same writing, she suggests three main forms of creativity, namely:

i) ‘Combinational creativity’: the process of combining existing ideas to come up with something new.

ii) ‘Exploratory creativity’: the process of creating new ideas within a predefined conceptual space.

iii) ‘Transformational creativity’: the act of ignoring fundamental conventions to conceive potentially impossible but highly creative ideas.