The Problem With AI: Averages All the Way Down

After some time interacting with Large Language Models (LLMs) like ChatGPT, Anthropic Claude, and DeepSeek, you notice a a key limitation in their tendency toward a samey, average-of-all-responses output.

If you ask for a story, you receive the average of all stories in the training data — the most generic protagonist, the most generic plotline, the most generic setting and so on. In a parallel with human creativity, you get better results by imposing restrictions than by granting freedom. You can ask, instead, for a story in a specific genre, with a specific setting, featuring a specific sort of protagonist.

The results will be much more interesting, but on examination one finds that the story is the most generic example possible within the provided constraints. In order to make the story interesting at a finer grain, more details have to be provided by the user, which constrain the story to the most average version within those constraints. Inevitably one finds that any deviations from the expected must be compelled by the user. It’s like trying to approximate a curve with line segments. It’s averages all the way down.

Of course this is expected. A deep learning model works by minimizing loss, and any deviation from the average is a source of loss. Average, to the model, is the same as correct.

For many uses, maybe even most, average output ranges from acceptable to ideal. For a code function or a cookie recipe, the average output is probably fine. In some cases it might be perfect, while in others it might be stuck in a decent local maximum that needs some prodding to get into the right shape. This is possible because those subject domains are well-trodden, open meadows of knowledge.

But what about when we want the unexpected? Can a temperature parameter give us that, or do we just get a fuzzier average?

What about the frontiers of knowledge? Can we really hack away at the thicket of scientific ignorance from the sunny meadows of the averages of existing research? Can reinforcement schemas get us to the frontier, or just another local maximum?

In my next article I will introduce a simple test I have designed for judging the creative ability of AIs and what it’s shown me about the creative potential of current LLM models.

Go on to next article — A Creativity Test for AI: Part 1.