Every general-purpose AI is, structurally, an averaging machine. The model doesn't know what's true. It doesn't know what's interesting. It doesn't know what's right. It knows what's most likely to come next given what came before. That probability is calculated across a training corpus of hundreds of billions of words, where every word is essentially a vote about what should come next.
The output, then, is the consensus. The middle of the distribution. The thing that, statistically, fits best. For most kinds of work, that's fine. For some kinds of work, it's the entire problem.
THE WORK WHERE AVERAGES KILL YOU
If you're writing a how-to article about email marketing, the average answer is the right answer. Most of what's true about email marketing is well-documented and the consensus is largely correct. The model gives you a competent summary. You publish it. Nobody is harmed.
If you're writing the positioning statement for your brand, the average answer is the wrong answer. The whole point of positioning is to be specific, distinctive, exclusionary. To occupy a piece of mental real estate that nobody else can claim. The average gives you the position that everyone else also could have arrived at. By definition, that position is not yours. It's the category's.
This is what marketers and brand builders run into when they use AI for strategic work. The output looks competent. It reads professionally. It's also indistinguishable from what a competitor would produce if they asked the same question. The differentiation collapses. The work flattens.
WHY PROMPTING DOESN'T FIX THIS
You can prompt the model to be contrarian. You can give it a persona. You can paste in 5,000 words of your brand voice. The output gets more polished. The underlying mechanism doesn't change. You're still pulling from the average of a contrarian-shaped distribution. The contrarian average is still an average. It's just shifted to a slightly different point on the same curve.
This is why "AI prompt engineering" tutorials all have the same basic shape. They're moving you around inside the distribution. They're not getting you outside of it. The thing outside the distribution (an actual original perspective) requires a different generation mechanism, not a better prompt.
WHAT TASTE IS, MECHANICALLY
Taste is the ability to prefer one thing over another in a way that isn't reducible to popularity. It's saying "this word, not that word" because the rhythm is right, even if a survey of writers would have voted for the other one. It's reaching for the strange image because the obvious image is dead. It's the willingness to lose the casual reader in service of the serious one.
Models have no taste because they have no preference. They have probabilities. Probability and preference look similar from the outside. They're not the same thing. A probability is a calculation. A preference is a choice that overrides the calculation. Models don't override anything. They calculate, and then they emit.
This is why even excellent AI writing has a particular texture. It's the texture of probability. The smoothness of the average. The way a fast-food cheeseburger has a particular flavor that's recognizable even when the recipe varies. It's not bad. It's just the same taste, optimized.
THE ALTERNATIVE
An AI that doesn't think in averages has to be built on something other than the average. Specifically, it has to be shaped by a curated philosophy. A defined point of view. A corpus that's selected for its perspective rather than its volume.
This is what the AI thinking partner category exists to do. The premise: if you train a model on the open internet, you get an AI that produces the average of the open internet. If you train it on the work of a small set of minds with strong points of view, you get an AI that produces something closer to those minds. Not perfectly. Not always. But meaningfully closer than the consensus you'd get otherwise.
The category is small. It's hard to build because curating the right corpus requires actual editorial judgment. It's hard to market because "more thoughtful" is a worse pitch than "faster." But it solves the problem most strategists are actually having, which is that their tools are giving them the same answer they'd get if they hadn't asked.
WHEN TO CARE
For 80% of what you do in a day, the average is fine. Drafting an email. Summarizing a meeting. Translating a document. Quick research. The model's averaging tendency is invisible because the average is also the right answer.
For the other 20%, the average is the enemy. The strategic work. The positioning. The naming. The headline that has to break the pattern. The decision that defines the next year of your work. For those, you need a tool that doesn't average. Or you need to do enough of the work yourself that the model never gets to finish the sentence.