Essay · AI & designFebruary 2026 · 4 min read

The outlier problem in AI prediction.

Language models predict the most likely next word, so they drift toward the average. This essay is about why the moments that change the world are the ones that probably would not come next.

most likely ≠ most interesting

How ai prediction works

Large language models work by predicting the most statistically likely next word (or token) based on patterns learned from massive amounts of text. At its core, the model asks: "Given everything before this point, what word or response would most probably come next?"

This means the model naturally gravitates toward the center of the distribution—the average, the conventional, the consensus answer.

What this means in practice

The AI is structurally trained to favor the fat middle—the safe, common, and well-represented response.

Why outliers matter

The most meaningful breakthroughs in history were never the median prediction:

Scientific breakthroughs were once outlier ideas (heliocentrism, germ theory, plate tectonics).
Creative genius often means seeing connections nobody else sees.
Interesting conversation often comes from unexpected angles.
Innovation is almost by definition an outlier—something that departs from the norm.
Humor fundamentally depends on subverting expectations.

Scenario comparisons

ScenarioMost likelyThe outlier

What's a chair for?SittingA chair is a social contract about where power sits in a room.

Medical diagnosisCommon illnessThe rare underlying condition that actually fits the symptoms.

Business strategyFollow industry best practicesThe counterintuitive move that disrupts the entire category.

Art & expressionSomething that resembles existing artSomething that redefines what art can be.

The deeper concerns

1. Creativity vs. Probability — Creativity is, almost by definition, improbable. If an idea is highly predictable, we don't call it creative. By optimizing for likelihood, AI systematically filters out novelty.
2. The Homogenization Problem — If everyone uses AI to write, think, and create—and AI always gravitates toward the statistical mean—culture risks becoming blander and more uniform. A kind of intellectual regression to the mean.
3. Confirmation of Consensus ≠ Truth — The most likely answer reflects what most people have said, not necessarily what's correct. History is full of cases where the consensus was wrong and the outlier was right.
4. The 'Interesting' Lives at the Edges — The most interesting people aren't average. The most interesting ideas aren't conventional. The most interesting art isn't predictable. The most interesting scientific findings are surprises.

A nuance worth adding

AI models can produce outlier responses—temperature settings, sampling methods, and prompting techniques can push toward less probable outputs.

But even then, the model is still drawing from patterns it has seen. It can recombine in novel ways, but it's not generating ideas from genuinely outside its training distribution.

The real outliers—the things nobody has thought of yet—are structurally harder for a prediction engine to produce.

The bottom line

“A machine that tells you what probably comes next is enormously useful. But the moments that change the world, spark genuine insight, or make you laugh unexpectedly—those are the moments that probably wouldn't come next.”

Thanks for reading.

Back to the index