Does AI spell the end for human creativity?

If you ask someone today what Artificial Intelligence (AI) is, most people now will think of a tool such as ChatGPT. Based on a particular form of machine learning called neural networks, the essence of this Generative AI is a system that is trained on very large amounts of data (for example, all the fairytales ever written) and is able to produce its own outputs (for example, a new fairytale) in response to a prompt from a user (eg “write me a fairytale about a boy who finds a magic stone in his back yard”).

Since ChatGPT and similar Generative AI – think Claude or Bard – became widely available, beginning around June 2021, there has also been an explosion of hype about this technology. This ranges from predictions of a technological Singularity in which AI escapes the bounds of human control and runs amok, to lesser predictions of the widespread replacement of humans by AI in jobs such as data entry, customer service and graphic design.

However, one of the more popular topics prone to AI hype is human creativity.

Almost as soon as Generative AI models became widely available, we saw an explosion of “creative” outputs. A popular predecessor to ChatGPT was DALL-E 2, an image generation model. Instead of generating text in response to a user prompt, this model produced images. Disparate elements could be combined (eg a dog riding a horse) in different media (eg oil painting) and even in a particular style (eg cubism) to produce what seemed to be genuinely creative works.

People started submitting AI-generated works to art competitions, and a panicked debate grew. What does this mean for human artists?!

Almost as quickly, however, some people began to question the apparent creativity of these artificial works. Not only did they often contain many errors (eg too many fingers on human hands), but questions also began to be asked about the material used to train such a model: whose art was used, and was permission obtained to use it? If an AI-generated image is an amalgam of existing images, media and unique artistic styles, then is it really creative, or is it just a sophisticated reworking of existing art? Even more importantly, if a creative image results from a user prompt (eg “an oil painting of a young woman, scrolling through her smartphone, looking bored, all in the style of Da Vinci’s Mona Lisa”) then is any resulting creativity due to the AI, or due to the human prompt? In other words, is the value of DALL-E 2 simply that it saves the user the trouble of learning how to paint?

In fact, the creativity, or lack thereof, of Generative AI can be understood most readily by considering how ChatGPT writes a story.

If you prompt Generative AI to write a story, what happens is the following. Your prompt is first broken down into individual words (called tokens). The model then converts those tokens into numbers (called vectors). These numbers attach meaning to each token, based on the model’s training. The model then interprets the relationships between all of the words (tokens) of your prompt (in a transformer layer), in simple terms understanding the context of your prompt. Finally, an output layer produces a list of probabilities for each possible output word in the model’s vocabulary, giving the highest probability to the word that best fits your prompt. For multiple words (ie a story), this process continues, always seeking the best (ie highest probability) next word given your prompt.

There are a couple of problems with this if you are looking for creativity in the output. First, a pre-requisite for creativity is novelty. Novelty means newness, or originality, or statistical uncommonness. If Generative AI is designed to choose the words in a story based on high statistical probability, then it is designed, in effect, to avoid novelty. Second, another pre-requisite for creativity is relevance and effectiveness. Effectiveness means appropriateness. Generative AI is designed to choose the words in a story that best fit the prompt – that is, words that are most appropriate – biasing the model in favour of effectiveness.

Therefore, mathematically, Generative AI favours effectiveness at the expense of novelty. Creativity, on the other hand, seeks to maximise both effectiveness and novelty. This places a fundamental, technological limit on the creative capacity of Generative AI.

So here is the issue. Generative AI is limited in how much creativity it can produce, by design. If it tries to increase novelty, it can only do so at the expense of effectiveness, and vice versa. Humans don’t have this same constraint. There is nothing stopping us, in theory, from maximising both effectiveness and novelty simultaneously. Not everybody does this on a daily basis, but many can, and many do. Mozart, Dickens and Rembrandt prove this.

What does the technological limit on Generative AI creativity look like in practice? Well, we have some good evidence. A recent study by Chakrabarty et al (2024) compared stories written by professional human authors, with similar stories written by several different AIs. Expert human judges then rated the creativity of these stories, without knowing who had written which. These judges had no trouble working out which stories were written by humans, and which were written by AI. Not only that, but they consistently found that the stories written by professional human writers were the most creative. Finally, most telling of all, they determined that the AI-written stories lacked many of the features of professional human writing, and resembled stories that they said were typical of amateur human authors. Generative AI, in other words, and thanks to the technological barrier described above, is limited to amateur creativity.

Does AI spell the end of human creativity? Only if we let it. If we stop demanding professional human creative writing, professional human music, professional human art, and simply surrender ourselves to bland, mediocre, amateur artificial creativity, then yes. But if we keep reading, listening to, appreciating, and demanding, the best of literature, music, art and other creative outputs, then human creativity will always be in demand.