on fantasy names

Why does every fantasy protagonist seem to be named Kaelen?

¹

Recently, while experimenting with LLMs for fantasy and RPG prompts, I started noticing the same kinds of protagonists appearing over and over again in replies. Not literally the same character, but close enough: similar names, similar archetypes, similar story styles. One recurring example was names like Kaelen, Elara, Thorne, and Rowan.

At first, I dismissed it as a coincidence and maybe the tool was picking up from my earlier conversations in memory, but after seeing the pattern play out across different models, I wondered why.

My initial question was simple: Why does every fantasy protagonist seem to be named Kaelen? ² The more I thought about it, the more that felt like the wrong question.

The interesting part wasn’t the name itself. Kaelen was just an example of a broader pattern. The real question was why different models, given so many possible options, kept arriving in roughly the same region of the creative space³.

That led me to the idea of mode collapse and the concept of statistical attractors⁴. Certain names occupy a sweet spot where they satisfy multiple constraints at once: they’re fantasy-sounding, easy to pronounce, familiar but not tied to a specific famous character, and consistent with genre expectations. So when given a vague prompt like ‘Give me a name for the hero of a fantasy story’, LLMs are naturally pulled toward those valley regions.

What I took away is this: Generative systems tend to converge on culturally reinforced defaults. Culture develops conventions, training data captures those conventions, models learn them, and generic prompts activate them.

Once I understood the mechanism, I wondered: How do I escape those attractors during prompting?

The answer was to create friction in my prompt:

• Adding unusual constraints (e.g., “58-year-old irrigation engineer from a Bronze Age river civilization” instead of “fantasy hero”)⁵
• Ask for the cliché answer first, then avoid it. The meta-prompt that worked best: «Before answering, identify the most likely cliché response. Do not produce that response. Generate something meaningfully different while still satisfying the request.»

Looking back, the thing I found most interesting wasn’t the explanation itself but the path that led there. A small observation about recurring fantasy names gradually unfolded into a more general mental model about how generative systems behave:

graph LR
    A["<strong>Observation</strong><br/>Recurring fantasy protagonists"]
    B["<strong>Pattern</strong><br/>Outputs look similar"]
    C["<strong>Mechanism</strong><br/>Statistical attractors"]
    D["<strong>Principle</strong><br/>Convergence on reinforced defaults"]
    E["<strong>Application</strong><br/>Design prompts for exploration"]
    A --> B --> C --> D --> E

The original question turned out to be less interesting than the one hiding underneath it. That’s often a good sign that you’re learning something worth keeping.

Have any of you noticed this when using AI tools?

Any examples where you kept seeing the same kinds of answers, ideas, or recommendations come up?

Footnotes

1. Kaelen ‘Wraith’ Vashti is the main character in my Active Play Fiction story, The Shattered Obelisk. ↩

2. I use Kaelen or Kael in a couple of my stories, which made this pattern personal 🧐 ↩

3. Of all the gin joints in all the towns in all the world, she walks into mine. - Rick Blaine in Casablanca (1942) ↩

4. Wikipedia entry for attractors ↩

5. I have a story in mind for such a character. ↩