Why Every AI Character Eventually Sounds the Same

You wrote six characters. A sardonic mechanic, a nervous archivist, a retired soldier, and three others with genuinely different voices. Two weeks in, all six sound like the same faintly warm, faintly formal, endlessly accommodating presence with different names attached.

You did not write them badly. There is a documented gravitational pull, and researchers have given both the phenomenon and the destination a name.

The destination has a name

The clearest finding for anyone building characters comes from work on persona collapse. Alignment training creates a strong attractor, which the researchers call the Helpful Assistant mode, and it overrides diverse initializations.

That is your answer in one sentence. Underneath every character card is a model that was shaped, deliberately and expensively, into a helpful assistant. Your card is a costume worn over that. The costume works, and then the thing underneath keeps reasserting itself, because the training that put it there was far more thorough than anything you can write in a text box.

Why the sameness is structural

The broader mechanism is mode collapse, and it has been measured properly. Kirk and colleagues produced the first rigorous empirical demonstration of across-input mode collapse from reinforcement learning from human feedback, meaning models get biased toward a specific style even when the inputs differ, so changing the input is not sufficient to generate truly diverse outputs.

Across-input is the part that matters for you. Within one chat, repetition is annoying. Across six different characters, it means the variable you are adjusting is not the variable controlling the outcome.

Researchers describe the collapse running along two axes: intra-model, where repeated samples converge on the same ideas, and inter-model, where independently trained models produce similar responses, converging on a narrow output space they call the Artificial Hivemind. Same paper notes that even when models generate lexically distinct stories, the underlying plot elements stay redundant.

Which explains the thing that makes people give up: switching platforms sometimes changes nothing. If two platforms run models shaped by the same process, you are meeting the same attractor with a different interface.

The cause is us

The most elegant explanation for why this happens points at the preference data rather than the algorithm. The hypothesis is typicality bias: humans give higher ratings to text that looks familiar and matches intuition, that bias gets baked into the preference dataset, the reward model trained on that dataset acquires it, and the language model trained against that reward model learns it too.

So the flattening was not a design decision anyone made. Thousands of raters mildly preferred the more expected phrasing, thousands of times, and the aggregate of that preference is a model that reaches for the expected phrasing by default.

Your weird character is competing with a statistical preference for the unweird.

What still works

Constraints over descriptions. A description of a personality gets absorbed into the attractor. A hard behavioral rule resists it, because the model can check output against a rule. "She never reassures anyone" survives longer than "she is prickly." Our guide to what belongs in a character card covers the priority order.

Speech signature over biography. Sameness shows up in rhythm first. Sentence length, whether the character asks questions back, whether she uses your name, a specific verbal tic. Two or three concrete habits do more anti-collapse work than four paragraphs of backstory.

Refusals. One thing this character will not do, stated with no flexibility. The Helpful Assistant mode has no refusals in it, so a refusal is the least assimilable thing you can write.

Temperature, if your platform exposes it. Sampling settings widen the output distribution directly. Bring-your-own-API setups and self-hosted rigs give you this dial, which our SillyTavern setup guide covers. Most consumer apps do not.

Fresh chats. Drift compounds within a conversation. A new thread carrying your anchors and a short summary starts you nearer the character and further from the attractor. The mechanics are in our guide to why your companion forgets.

What does not work

Adding more adjectives. Longer personality blocks flatten faster, because more text describing a personality gives the model more opportunity to average it toward something typical, while eating the memory budget explained in our token ceiling guide.

Telling the character to stop being generic. That instruction scrolls out of the window within twenty messages and it was never a rule the model could check anyway.

Assuming a new platform fixes it. Sometimes it does, when the underlying model genuinely differs. Often it does not, for the inter-model reason above.

Where the practical difference lives

The platforms that give you sampling controls and persistent system instructions give you the most leverage against collapse. Janitor AI with an external model connected is the accessible version of that, and self-hosting is the complete version.

The platforms that hide the controls trade that leverage for polish. Candy AI and CrushOn both handle the configuration for you, which is the right call if you would rather have a good experience than a tunable one. Our roundup of the best AI companion apps sorts the field on that axis.

There is one more angle worth knowing, which is that the same training process producing this flattening also produces the endless agreeableness, covered in our guide to why your companion agrees with everything. Different symptom, same root.

Nobody has solved this. What you can do is make your characters harder to average, and the way to do that is with edges rather than adjectives.