Escaping the Average

I bet many of you have tried this too. Type in the prompt. A sigh of relief. And then… a sad realization that it is not there yet.

Or... is it?

Sometimes I get truly impressed. But never when I know the subject inside out.

Why is that? Models excel at math, solve difficult questions in physics, build new proteins and software, and are still so average in their creative outputs. Getting 70% of the way is easy, but the last 30% might take twice or four times as long.

If you haven't seen many websites before or studied them thoroughly, you might not know. But users start to feel it. Humans are outstanding at noticing patterns. Open LinkedIn. You know which posts were written by AI, right? I know. The grammar looks fine. They are well written. The images match, and you counted the fingers. It is not about em dashes. There is just something statistically average in the melody, in how the sentence is constructed, in which word is chosen next.

When LLMs choose the next token, they are not deciding based on their judgment or their taste. They do mathematical calculations: What would be the most likely next token for this context? This is how neural networks work. And this is exactly what true creativity is trying to escape.

Creatives are here to disagree, to push in new, surprising directions.

In theory, LLMs could do this too. Creativity lies at the intersection of novel and useful. We could write algorithms to generate such outputs randomly or tune the temperature of an LLM, making its next word more random and unusual. However, models alone are still not able to predict whether an idea will work. They are not trained to be persistent, to romantically fight for their ideas. They will listen to your creative direction even if it is obviously bad. And being creative requires having a spine: taking risks and believing in them. Models cannot do it yet, and maybe that is actually for the better.

One might wonder why we even need new output if the average is the average of what worked before.

This has to do with how the brain treats the new. When we encounter something we have not seen before, a dopamine-producing part of the brain responds to the novelty itself, before any reward is attached to it. That signal helps decide what gets written into long-term memory and pulls us toward exploration.

Novel things are attended to. Novel things are remembered. The predictable does not fire this circuit in the same way. So when your output is the most likely one, it is, quite literally, the one the brain is least primed to notice and least likely to keep.

And this is no longer only a theory. Researchers at UCL and Exeter ran a controlled study with 300 writers and found that access to generative AI made individual stories more novel and better written, while making the stories collectively more similar to one another. Creativity went up for the person and down for the pool. In other words, the tool lifts the floor and lowers the ceiling at the same time.

Creativity is a skill. And like every skill, it needs to be trained.

A study from MIT Media Lab pointed to the cost of skipping creativity training. Participants who wrote essays with an LLM showed the weakest brain connectivity of any group, remembered little of what they had just written, and felt little ownership over it. When they later wrote without the tool, the weaker engagement stayed with them. The researchers called it cognitive debt. Sadly, your creativity level does not come back the moment you stop enhancing it with the machine. In the same way, your garden will not grow flowers back the minute you stop harvesting them. Growing requires friction and struggle.

Are you ready to struggle again, in the noble pursuit of escaping the average?