Why AI writing sounds the same
You noticed the output sounds like everything else, and every explanation you can find says the same thing: the model does not know your brand, so give it more context. That is a description of the symptom being sold as a diagnosis.
There is an actual mechanism, it is published, and it is more interesting than the folk version. The flattening is not inherited from what the model read. It is introduced later, when the model is aligned to human preferences, and it has been measured on real preference data.
This sits under the AI writing tells. If you want the practical end rather than the mechanism, eleven questions produce the constraints that survive it.
Is there an actual mechanism, or is this a vibe?
There is a mechanism, and it has a name: typicality bias in preference data. The argument is that when people rate one model output against another, they systematically prefer the more familiar-sounding one, and that preference is baked into the reward model the alignment process optimises against.
The consequence is uncomfortable. If the bias sits in the data rather than in the algorithm, then a perfect reward model and perfect optimisation would still encode it. It is structural, not a bug in the training run.
What has actually been measured?
Enough to take it seriously, from several independent directions. These are other people's numbers; this site runs no experiments and publishes none.
| Finding | Detail |
|---|---|
| Typicality bias is real and significant | Fitted to a preference dataset over 6,874 correctness-matched response pairs, giving a bias term of about 0.57 and 0.65, both at p below 10 to the minus 14 |
| Alignment introduces it, pretraining does not | Diversity drops after the supervised and preference-optimisation stages, and the aligned model still retains the base model's diversity underneath |
| Instruction tuning is the step that matters | Writing with an instruction-tuned model produced a statistically significant diversity reduction; the same task with the base model did not |
| The loss is the model's, not the writer's | In that study the reduction was attributable to the model's contributions rather than to what the humans wrote |
| It is a known trade-off, not an accident | Preference optimisation reduces output diversity relative to supervised fine-tuning while generalising better to new inputs |
| Bigger models do not fix it | Stylistic deviation from human writing persists with scale and is larger for instruction-tuned models than base models |
What is not the cause?
This section exists because the cluster is full of confident explanations that the evidence does not support, and saying so is more useful than adding another one.
- Not that the model has not read enough. The base model has the diversity; it survives underneath the aligned one.
- Not that it does not know your brand. Telling it about your brand helps, and it is not what caused the flattening.
- Not model size. The deviation persists as models get larger.
- Not a single punctuation mark. That folk theory has its own page's worth of argument and it is not a mechanism for anything.
Why does more prompting not fix it?
Because you are asking the model to move against a pull that was optimised into it, using the same channel the pull acts on. More words about how you want it to sound do not change the distribution those words are sampled from.
There is a published prompting technique that does move the needle, and the shape of it is instructive: instead of asking for one answer, you ask the model to lay out a spread of possible answers and their likelihoods. Reported diversity gains of roughly 1.6 to 2.1 times over asking directly, on creative writing tasks.
Note what that is not. It is not a voice. It is a way of getting more variety out of a model that has been trained towards the middle. Variety and your voice are different problems, and only one of them is solved by writing your rules down.
What does a constraint have to look like to survive the pull?
Specific, mechanical, and checkable. The pull is towards the typical, so a constraint phrased as a preference is a suggestion the model can satisfy by being slightly less typical and calling it done.
- Prohibitions beat adjectives. Never open with a rhetorical question is enforceable. Sound distinctive is not.
- Mechanical settings beat description. Contractions, person, spelling variant and punctuation change every sentence and cost almost nothing to state.
- Worked examples beat both. A short pair showing the same message written two ways gives the model a target rather than a direction.
- Short beats long. Vendors document that longer instruction files reduce adherence, so a constraint buried in three thousand words is a constraint you have already weakened.
What we cannot explain
Why particular words rise and fall. The vocabulary turns over between model generations and no published account explains the specific choices, only the general pull towards the familiar.
How much of any given draft is the flattening rather than the prompt. Nobody has separated those cleanly, and this site will not pretend to: we publish no measurements of model behaviour and every number on this page belongs to somebody else.
Where these come from
Every claim above is quoted from one of these, and each was read on the date beside it. If one of them has changed since, the page is wrong and we would like to know.
- Zhang et al., typicality bias and mode collapse checked 2026-08-17
- Kirk et al., the diversity cost of preference optimisation checked 2026-08-17
- Padmakumar and He, instruction tuning and diversity loss checked 2026-08-17
- Reinhart et al., stylistic deviation persists with scale checked 2026-08-17
The pull is towards the middle, so the fix is a constraint specific enough to resist it: answer eleven questions and get one.