Why AI calls everyone Elara

Ask any model for a fantasy name. Go on — open a new chat and ask for ten names for a young woman in a cold northern port.

You will get Elara. You will probably also get Lyra, Seraphina, Isolde and Maeve, and somewhere in the list a surname that is Thorne, Vex, Vance or Voss. Ask it for a city and you will get Eldoria, Silvermere or Stormhaven. Ask what the magic is called and it will be Aetherium.

You will get these from GPT, from Claude, from Gemini, from Llama. You will get them tomorrow. You will get them if you ask nicely and you will get them if you beg. In December 2025, a naming blog that has run a Name of the Year award since 2004 gave the title to Elara, on the grounds that it is the name AI likes.

This is not a small aesthetic complaint. Readers of serialised fiction have learned these names. Seeing Elara on page one is now, for a lot of people, the moment they decide what they are reading was not written by anyone. The names are a tell, and the tell arrives before your prose gets a chance.

So: why does this happen, why does telling the model not to do it not work, and what does?

It is not laziness. It is what the model was rewarded for.

The intuitive explanation is that the model is being lazy, or that "fantasy names" is a shallow part of its training data. Both are wrong, and the real answer is more interesting.

Modern models are tuned with human preference data: people are shown two responses and asked which is better. The problem is that human raters have a well-documented cognitive bias toward the familiar, the fluent and the prototypical. Ask a hundred people which of two invented names is "better" and they will systematically prefer the one that sounds like a name they already know.

A 2025 paper calls this typicality bias and shows it is the root of the problem: "Verbalized Sampling: How to Mitigate Mode Collapse and Unlock LLM Diversity". The reward model learns that typical means good. The policy trained against that reward model sharpens onto the single most typical answer. And once a distribution has collapsed onto a peak, asking it politely for something off the peak does not move it — you are asking the model to be something it was specifically trained not to be.

The trade is measurable and it was measured. RLHF-tuned models generalise better than supervised-only ones, but with substantially lower output diversity, both per prompt and across prompts: "Understanding the Effects of RLHF on LLM Generalisation and Diversity". A companion result, bluntly titled "Creativity Has Left the Chat: The Price of Debiasing Language Models", found aligned models show reduced syntactic and semantic diversity and cluster into what the authors call attractor states. Elara is an attractor state.

And the bigger model does not save you. NoveltyBench evaluated twenty leading models on their ability to produce genuinely distinct outputs and found all of them far below human writers — and, awkwardly, that larger models within a family are often less diverse than their smaller siblings. Upgrading is not a fix. It may be the opposite of a fix.

Two more things go wrong that nobody warns you about

It gets worse inside a single conversation. Ask for twenty names and the first five will be the best you get. A 2026 paper on barriers to diversity in model-generated ideas (arXiv:2602.20408) separates two distinct mechanisms: fixation, where a model's early outputs constrain its later ones within one session, and knowledge aggregation, where a single model is one averaged distribution while a room of human writers is a partitioned one — each person occupying a different corner of the space. This is why your cast drifts. You name one character Kael and by chapter twelve you have a Kaelin, a Kalen and a Kelric, all distinct strings, all the same name to a reader.

And the model literally cannot count syllables reliably. This one surprises people. Models read text as byte-pair tokens, not letters, so instructions like "two syllables, hard consonants, no sibilants" are not merely ignored — they are close to unfollowable. See "Tokenization Constraints in LLMs" and work on how tokenization limits phonological representation. Every "give me names with this exact sound shape" prompt is asking for something the architecture does not support.

The fixes people reach for, and why they mostly do not work

Turning up the temperature. The most common advice, and it is close to useless here. A study asking directly whether temperature is the creativity parameter (arXiv:2405.00492) found it only weakly correlated with novelty and moderately correlated with incoherence. You do not get more imaginative names. You get the same names plus some garbage.

Fancier sampling. min-p sampling was proposed as a way to get diversity and quality together (arXiv:2407.01082). A 2025 reproduction (arXiv:2506.13681) found the human evaluation had dropped a third of its data and that the benchmark gains vanished once hyperparameter search was controlled for. It is a reasonable safety valve at high temperature. It is not a diversity solution.

Random seeds in the prompt. "Here is a random string: 7f3a9c. Use it to vary your answer." This one is the most interesting failure, because there is now a framework that measures exactly how much of an injected random choice reaches the output. "Where You Inject Diversity Matters" ranks interventions by where the randomness enters:

Where the randomness goes in	How much of it reaches the output
Nowhere (direct prompting, temperature)	baseline
Surface — a random string, a random unrelated word	0.003
Specification — the draw decides the generative configuration	0.46 – 0.55

That is not a marginal difference. Sprinkling randomness on top of a request is, to two decimal places, doing nothing. Randomly choosing what kind of thing to make before making it is doing almost everything — and, the paper notes, at comparable quality rather than trading quality away.

Telling it what not to write. This is the one that stings, because it is what everybody tries first, and in our own testing it actively backfired. We once replaced a terse instruction in our chapter prompt with a careful, exhaustive one listing every unwanted opening. Measured on the same fixtures, the unwanted behaviour went from appearing zero times to appearing in almost every chapter. Naming a failure mode in a prompt appears to prime it. "Never write Elara" puts Elara in the context window, and the next thing the model writes is Elara-adjacent.

So what actually works

Put together, the research points somewhere specific and slightly humbling: stop asking the model for the name.

The evidence says diversity has to be decided before generation, at the level of the specification rather than sprinkled on top. It says phoneme-level structure cannot be enforced through a prompt and has to be enforced somewhere that can actually count. And it says the list of names to avoid must never be shown to the thing you are trying to steer away from them.

None of that is a prompt-engineering problem. All of it is an ordinary-software problem.

That is the approach we took in webnovellm. The sound of your world is settled once, up front. Names are then built rather than recalled — the shape is drawn first and the letters are filled into it — and anything that comes out looking like the handful of names every model already reaches for is thrown away before you ever see it, along with anything too close to a name your novel has already spent. A model is still involved, but its job is reading your world and choosing between finished candidates, which is a thing models are genuinely good at. It is not asked to invent, because inventing is the part it cannot do.

The practical upshot is that it costs nothing and takes no time, so it can sit inside the writing itself: when the chapter reaches a person or a place that needs a name, the name is already waiting.

Here is a small version of the same idea, running entirely in this page.