The bell curve theory of AI

Everyone has seen a bell curve. Most values cluster in the middle, a few outliers sit at the edges. It describes height, test scores, marathon times - and, it turns out, how AI thinks.

AI learned from billions of texts written by millions of people. Some brilliant, most average, some terrible. It internalised the entire bell curve of human expression, and it gravitates toward the middle.

That single insight explains almost everything people find confusing: why AI is impressively competent yet rarely brilliant, why it is confidently wrong, and why talking to it the right way changes everything.

01Training is learning the distribution

When a model trains, it reads the internet: medical papers, forum threads, textbooks, blog posts, answers both correct and incorrect. It learns what the typical response to any question looks like.

Think of every possible human response to a question as a point on a curve. The peak is the most common answer. The tails hold the rare ones: breakthrough insight on one side, nonsense on the other.

AI learned the whole curve. By default it hands you the peak - the most statistically likely answer.

02Smart, but not a genius

A doctor with twenty years of niche experience lives in the right tail. AI learned from that doctor - and also from thousands of medical students, health blogs and symptom-checker pages. The result is solid median competence across everything and mastery of nothing.

This is why AI nails the common case. The last 20% - the part that requires real expertise - lives in the tails, where the curve gets thin.

03Temperature controls the width

Models have a setting called temperature. It literally controls how wide the curve is when the model picks its next word.

Same model, same training, different width of the same curve. That is all temperature is.

04Why how you ask matters

A vague prompt samples from a wide distribution. The model could go anywhere, so it goes to the peak. Average question, average answer.

A specific prompt does something more powerful than narrowing the curve: it moves the center toward the tail you actually want.

This is why prompting works at all. You are not tricking the model; you are telling it which part of the curve to sample from. The more specific your context, the further from the generic peak you pull it.

Good prompts do not make AI smarter. They move the center of the curve.

05Why it struggles with the genuinely new

The curve only exists where there is data. Truly original ideas - never written, never thought, never combined - have no distribution to sample from.

Ask it to combine two well-known concepts in a standard way and it excels; it is sampling from a rich, well-populated curve. Ask it to invent something genuinely new and it falters, because the curve is empty. There is no peak to aim at.

That is not a flaw, it is a property of the system. AI recombines existing human knowledge in useful ways. It does not generate knowledge that never existed - the curve cannot reach past its training data.

Original thought is still a human monopoly. For now.

06The world's best-read average person

If you keep one thing from this piece, keep this.

AI is the world's best-read average person. It has read everything, remembered the patterns, and gives you the most statistically likely answer. That is enormously useful - and it is the middle of the curve, not the edge.

Once you see that, you stop being disappointed by the limits and start being strategic about the strengths.

The bell curve is not a limitation to fight. It is a tool to wield. Whoever understands it gets far more from AI than whoever treats it as magic.