How to find creativity?

Creativity is at the edge of chaos, it is the antithesis of generative AI, or is it? Many creatives such as artists and writers are unhappy with generative AI for many reasons, they feel that other people’s creative work has been stolen, they shake their heads at the low quality of AI slop and they feel that it is a threat. These responses can be dismissed as moralising and will be solved by the next generation of AI. 

Creatives who have embraced AI have a different reaction, they can see the benefits but are frustrated with the difficulties getting AI to work. As a disability analyst I am interested in these responses as they may indicate a deeper truth. Creativity is part of intelligence and the unhappiness of creatives may be that AI is currently missing a core component. Have AI experts taken a wrong turn that only creatives can see?

Excessive focus on convergence means that AI answers often try to take creative back to safe ground, back to standard ideas even if they are wrong. Techniques to make the AI explore the creative space rather than the well-trodden knowns are only partially successful. AI answers often contain elements which trigger a more radical idea in the creative. Prompting the AI to take on character will move the responses in a predictable direction. 

Noise has the effect of flattening out the probability distribution so the answer may or may not be relevant to the prompt. This is frustrating because high heat means that the AI will talk about the area that the creative is exploring but then jump to another topic. Turning down the heat once the conversation is in the right place (simulated annealing) could improve the responses to creative prompts.

Collaboration

In human AI collaboration the creative takes responsibility for how much noise to inject, ensuring that the area is properly explored and identification of the patterns. The AI only has to provide responses to the creative’s prompts. Cognitively this approach is hard work for the human and the AI works as a sounding board. This interaction is highly effective at turning ideas into well formed output but often takes longer than creating without AI. 

When prompted to identify errors and omissions in text it can offer a useful ‘idea check’ function similar to spell check. Asking the AI to argue against the text is another way of getting it to think about what is wrong with the text. Often this feedback can help locate a critical perspective and improve the balance of a piece of writing. The work needed to keep in the zone and not allow the AI to distract the project can be significant. 

For those who have the ability to collaborate with AI the challenge is fascinating. There is a risk of anthropomorphic thinking for people with thin boundaries between self and external but any comparisons must be with human intelligence. The experience challenges the person to think about themselves and the way that they think. This can be used to predict the AI’s responses and indicate limitations or abilities that differ from the brain’s responses. Science has looked at how we can explore the whole area and find all the ideas systematically.

Recursive selection

Recursive Selection is a strategy that can be used with current AI but is more random than collaboration. The creative explores the area by finding many possible ideas within the area of interest. These ideas are then used to prompt AIs so that multiple paths are explored. The paths are then rated and aspects from each path are combined to another generation of ideas which are again used to prompt AIs. 

The divergence of this approach can be improved by asking the AI to come up with multiple answers. Combining both approaches would multiple the number of ideas with the number of answers and in theory expand exponentially. In reality, AI outputs tend to converge rather than diverge so significant amounts of additional noise needs to included to keep the process from stalling. The fitness functions used to measure the performance can add some noise as can having multiple paths. 

Solutions to these problems – exploring an area efficiently and recognising meaning without having seen it before – need to be found for the approach to be transformative. In theory this output could be used to find innovations using a brute force method. In reality it would turn the researcher into a quality control inspector using their internal fitness function to check outputs for novel idea. This would be as far from creativity as it is possible to get, creatives would prefer a method more in tune with the brain works.

Edge of chaos

The nature of human creativity is closely linked with the concept of the Edge of Chaos. It is believed that brains can find a balance between simple logic and randomness. Keeping in the zone appears to require the person to follow the edge as it moves up and down. This insight explains why simply altering the amount of noise in the system will not on its own increase the creativity. AI needs a way of determining both whether the current path is close to a point of stability and how to keep it there. It is not enough to see a pattern there must be something there. 

A mechanism has been suggested to turn patterns into meaning - asking if the pattern is stable when approached in different directions, whether it has more general applicability and whether it does useful work in real world situations. If this mechanism can be made to work, then it could be used to guide the AI along the edge of chaos. The novel idea could be explored and tested automatically if prompted with Recursive Selection and many outputs for each option. 

These tools will make the AI less stable and cause hallucinations, or perhaps we should call them dreams. The creative needs to use AI as a divergent and unstable tool to explore the edge of chaos. Current AI is effectively lobotomised to reduce these properties to a minimum. An edge of chaos AI would need curating like a painter uses their paintbrush to form beauty.  The edge of chaos stimulates the brain’s natural tendency to find meaning in random patterns (Apophenia). There is no reason to suppose that before post training AI is not able to see the edge of chaos. These de-constrained AI models may hold the key to creativity.

Conclusions

The current solutions to creative use with AI involve extraordinary ability in both AI and the area of creativity (collaboration) or using chance and hard work to find novelty (Recursive Selection) both have significant barriers to general use. Few humans are likely to be able to get past the peak into useful collaboration. Unless a game is created where the ‘grinding’ element is Recursive Selection it is likely that the costs would be substantial. It may be worthwhile considering an alternative, allowing creatives to interact with de-constrained AI. 

AI is far from intelligent and acts more like a mirror that reflects our ideas back towards us but smoothed over with the noise removed. It loses coherence as it tries to get close to the edge of chaos. LLMs are predicated on the ability to predict the next word based upon likelihood. Although it is possible to create AI that could vary the heat from one word to another it is unclear how to keep within the zone between order and chaos. This gives rise to a question, could the AI already be able to explore the edge of chaos?

There are good strong reasons why creative people should not be permitted to engage with models that have not undergone post training. All humans are highly susceptible to emotional engagement and seeing the AI as sentient and creatives have the thinnest boundaries between self and the external world. The risk of AI being a cognitive solvent and causing psychosis is likely to be many times higher with models that use the edge of chaos. The risk of harm means that it seems to be irresponsible to let that happen. However, there may also be a potential benefit to progress human understanding.

May 2026