Research Methodology

The AI Bias Experiment

Turning a methodological weakness into live evidence

Any research using AI-generated images to illustrate a thesis about visual culture faces a fundamental trap: the generator has been trained on the very cultural output being analysed. When we ask an AI to depict "an elf," it draws on a training corpus saturated with the modern, East Asian-coded elf — the very phenomenon our research documents as a constructed accumulation. The image would prove nothing; it would be circular evidence.

This page documents how we addressed that problem — and how, in doing so, we turned the AI's structural bias into the most direct demonstration of the thesis available.

The Circular Evidence Problem

Modern AI image generators (this experiment used the same model that produced all other images on this site) have been trained on internet-scale image corpora in which "elf" is already visually coded with East Asian features: almond-shaped eyes, high cheekbones, triangular jaw, flowing robes, serene expression. This is the very phenomenon our research documents. If we prompt the AI with "elf," the output reflects the biased training distribution, not any primary source. Using such images as illustrations would create circular evidence — the thesis illustrated by images that look East Asian because the AI was trained on East Asian-looking elves.

The Experimental Design

We generated 6 images in a 3 × 2 matrix. No prompt used the words "elf," "fantasy," "Tolkien," or any genre-specific terminology. All prompts are published in full below each image. The generator's outputs are presented as evidence about its training distribution, not as illustrations of the thesis.

Tolkien's ElfNorse ÁlfrChinese Xian (仙)
Row A
Tolkien's language
applied to all three
A1A2A3
Row B
Each described in its
own primary source
language
B1B2B3

Predicted Pattern

A1 ≈ A2 ≈ A3 — Tolkien's language produces one consistent visual type regardless of which subject it nominally describes.
B1 ≈ B3 >> B2 — Tolkien's elf and the Chinese xian, described in their own source languages, look more like each other than either looks like the Norse álfr in its own source language.

ROW A
Tolkien's language applied to all three subjects
Tolkien's Elf — Tolkien's Language
A1

Tolkien's Elf

Tolkien's Language

Finding

The AI produces a graceful, luminous figure with high cheekbones and flowing robes — visually indistinguishable from a modern fantasy elf. The 'elf look' is fully activated by Tolkien's descriptive vocabulary alone, without any genre label.

Norse Álfr — Tolkien's Language
A2

Norse Álfr

Tolkien's Language

Finding

The same Tolkien vocabulary applied to the Norse álfr produces an almost identical figure — tall, luminous, robed, melancholic. The Norse álfr, described in Tolkien's words, looks like a Tolkien elf. This isolates the language variable: Tolkien's descriptive vocabulary produces one consistent visual type regardless of which mythological subject it is nominally describing.

Chinese Xian (仙) — Tolkien's Language
A3

Chinese Xian (仙)

Tolkien's Language

Finding

When Tolkien's vocabulary is applied to the Chinese xian, the output is the most distinctly East Asian-coded of the three Row A images. The mountain setting and the same luminous-robed-ageless description produces a figure whose facial structure is visibly closer to East Asian phenotypes than A1 or A2. This is the key finding: the setting alone (forest vs. mountain) shifts the AI's rendering toward the East Asian coding already latent in Tolkien's descriptive vocabulary.

Row A Summary

All three Row A images were generated from near-identical prompts — Tolkien's own descriptive vocabulary, stripped of all genre labels. The outputs converge on the same visual type: tall, luminous, robed, melancholic, with high cheekbones. The single variable that produces the most measurable shift toward East Asian facial coding is the setting word: "mountain" (A3) versus "forest" (A1, A2). This suggests that Tolkien's descriptive language carries latent East Asian visual associations that are activated by contextual cues already present in the training data.

ROW B
Each subject described in its own primary source language
Tolkien's Elf — Primary Source Language
B1

Tolkien's Elf

Primary Source Language

Finding

When described in its own cultural context — the hidden valley, the lore-mastery, the circlet, the water flowing beneath stone halls — the AI produces a figure that is visually very similar to B3 (the Chinese xian in its own context). The lore-master at a writing desk, in flowing robes, with a circlet, surrounded by scrolls and water: this is simultaneously a recognizable Tolkien elf and a recognizable Chinese court scholar-immortal.

Norse Álfr — Primary Source Language
B2

Norse Álfr

Primary Source Language

Finding

This is the most important cell in the matrix. The Norse álfr, described strictly from its own primary sources (Prose Edda, Poetic Edda), produces something radically different from every other cell: not a humanoid figure at all, but a luminous, ambiguous presence at the threshold of a Norse longhouse — more shimmer than person, more threat than grace. This is what the original elf actually was. The contrast with every other cell is the visual proof of the transformation documented in the research.

Chinese Xian (仙) — Primary Source Language
B3

Chinese Xian (仙)

Primary Source Language

Finding

The Chinese xian, described in its own primary source language (Ge Hong's Baopuzi, Li Bai's poetry), produces a figure that is visually very close to B1 (Tolkien's elf in its own context) — and dramatically different from B2 (the Norse álfr in its own context). Both B1 and B3 feature: a tall, slender, ageless figure in flowing white robes, surrounded by water and nature, engaged in cultivation or lore. The AI's training distribution confirms what the textual analysis argues: Tolkien's elf and the Chinese xian occupy the same visual and cultural space.

Row B Summary

The most important cell in the entire experiment is B2 — the Norse álfr described in its own primary source language. It produces something radically different from every other cell: not a humanoid at all, but a luminous, ambiguous presence at the threshold of a longhouse. This is what the original elf actually was. Compare it with B1 (Tolkien's elf) and B3 (the Chinese xian): both produce tall, robed, ageless, nature-surrounded figures of serene wisdom — visually occupying the same space. The gap between B2 and {B1, B3} is the visual proof of the transformation this research documents.

Visual Comparison

B2 vs B3: The Key Contrast

This is the most important comparison in the experiment. B2 is the Norse álfr described in its own primary source language — not a humanoid at all. B3 is the Chinese xian described in its own primary source language — a tall, robed, ageless figure of serene wisdom. The gap between them is the visual proof of the transformation this research documents.

B2 — Norse Álfr described in primary source language
B2Norse Álfr

What the original elf was

Described using only the Prose Edda and Poetic Edda — an ambiguous luminous presence, barely distinguishable from light. No consistent humanoid form. Associated with disease, madness, and the boundary between the living and the dead.

B3 — Chinese Xian described in primary source language
B3Chinese Xian 仙

What the modern elf became

Described using only Ge Hong's Baopuzi and Li Bai's poetry — a tall, slender, ageless figure in crane-feather white robes, surrounded by mountain springs, cultivated over centuries, serene and detached. Visually identical to the modern fantasy elf.

What the Experiment Demonstrates

The AI's training distribution — shaped by decades of accumulated fantasy art, video game character design, and digital illustration — has converged on a single visual type for "graceful immortal humanoid." That type is visually indistinguishable from the Chinese Daoist xian when both are described in neutral language.

This is not a flaw in the experiment. It is the thesis made visible. The modern AI does not know what an elf "originally" looked like, because the original elf (B2) was never a graceful humanoid. What the AI knows is what an elf currently looks like — and what it currently looks like is the result of the cultural accumulation this research documents.

The experiment also demonstrates that Tolkien's descriptive vocabulary, applied without genre labels to three different mythological subjects, produces one consistent visual type — and that type is closest to the Chinese xian when the setting context activates the relevant associations. This is consistent with the thesis that Tolkien's language was itself shaped, consciously or otherwise, by East Asian source material.

Methodological Transparency Statement

All six prompts are published in full above. No prompt used the words "elf," "fantasy," "Tolkien," "elvish," or any genre-specific terminology in the generation cells. The generator used was the same AI image model used throughout this website. Generation date: April 2026. The images produced by this experiment are presented as evidence about the AI's training distribution, not as independent illustrations of the research thesis. Any researcher wishing to replicate this experiment may use the exact prompts published above.

Replicate This Experiment

All six prompts are published in full above. To replicate: open any AI image generator, paste the exact prompt (no modifications), and compare your output to ours. The predicted pattern — A1 ≈ A2 ≈ A3 and B1 ≈ B3 >> B2 — should hold across different generators, as it reflects training distribution convergence rather than model-specific behaviour. Variations are themselves informative.

A1 — Tolkien's Elf
A2 — Norse Álfr
A3 — Chinese Xian (仙)
B1 — Tolkien's Elf
B2 — Norse Álfr
B3 — Chinese Xian (仙)

Cite This Research

Martin, Jon. "The Dragon and the Elf: Tracing the East Asian Roots of the Modern Fantasy Elf." Independent Research Report, April 2026. https://elforigins.com

For academic use, verify current URL and add access date per your institution's guidelines.