AI is doomed to an "average" level of creativity

Generative neural networks like ChatGPT will not be a full-fledged replacement for talented writers, screenwriters or artists: their creative potential is inherently limited by the inherent architecture of the models. That's the conclusion of creativity expert David Crowley, a professor of engineering innovation at the University of South Australia.

In a new study published in The Journal of Creative Behavior (article ""The Cat Sat on the ... ?" Why Generative AI Has Limited Creativity"), Crowley attempted to formalise how "creative" large language models (LLMs) can be. Using standard mathematical principles, he estimated the upper limit of their creativity to be about 0.25 on a scale from 0 to 1, where 1 is the maximum possible creative expression.

According to the scientist, this corresponds to the level of the average person, but falls short of talented creators.

"AI can mimic creative behaviour - sometimes very convincingly - but its actual creativity is limited. Under current design principles, it cannot reach the level of truly talented humans," Crowley explains.

He emphasises that many people overestimate the creativity of AI, confusing content generation with creating new and original content:
the models learn from vast amounts of pre-existing text and images and produce expected, statistically predictable variations.

Crowley reminds us that in the science of creativity it is not the fact of producing a text or picture that matters, but its novelty, originality and efficiency. At the same time, people's creativity is unevenly distributed: about 60 per cent of the population is below average. Therefore, some people, when seeing ChatGPT results, do consider them "creative", while more creative people are better at noticing the limitations of such systems.

This is the first study to attempt to assess "AI creativity" based on the internal workings of language models, rather than subjective user impressions.

According to Crowley, AI is good as a support tool - for drafts of texts or ideas, for example - but is not capable of replacing a truly gifted author, artist or designer:

"A human can occasionally create something truly original and effective. An LLM won't do that: by definition it will produce an average result. If industries start relying on it too much, we'll get formulaic, repetitive work."

For artificial intelligence to truly approach an "expert" level of creativity will require fundamentally new architectures capable of generating ideas that are not tied to statistical patterns of past data, says the scientist.

"These results show: the world still needs creative people - perhaps even more than before," Crowley summarises.