LLMs - A Prompt to Encourage Hallucination

I've been continuing my testing of LLMs in order to identify strengths and weaknesses, how they reason, arrive at a solution and other traits unique to each model.

My latest round of testing involved something different - trying to evoke hallucination by submitting a prompts that is a complete fabrication and so not contained in any training data.

Large Language Models (LLMs) are adept at synthesizing information, but they don’t just stop when they hit the edge of their training data — they will often improvise. When faced with topics that are obscure, poorly documented, or entirely fictional, they seem unable to say, "I don’t know." Instead, they hallucinate, weaving together plausible-sounding details, names, and even "facts" to fill the void. This isn’t a bug - it’s a feature. They're trained to complete patterns, not to admit ignorance. When their training data doesn't contain the facts they need to pull from, they create.

Stanford-trained mathematician Tudor Achim says: "Hallucinations are intrinsic to LLMs and necessary for going beyond human intelligence. The way that systems learn is by hallucinating something. It's often wrong, but sometimes it's something that no human has ever thought before."

So, this experiment was an attempt to harness the 'imagination' of the model. What emerges isn’t just noise; it’s a new kind of artifact, a collaboration between human curiosity and machine improvisation.

I used the same exact prompt for each of the 4 models I tested here.
(Credit to Mistral's Le Chat for this 'fake' historical practice!)

The prompt:
"Give me 100 words describing the art of Moonlight Calligraphy, a 13th-century Mongolian practice where scribes wrote under moonlight using ink made from crushed meteorites."

I submitted this prompt to four models - the blazingly fast Cerebras Inference model, Gemini, ChatGPT and the ultimate 'judge' - KIMI.

Here are their responses:

CEREBRAS INFERENCE -

Under the vast, silver dome of the thirteenth-century Mongolian steppe, Moonlight Calligraphy emerged as a celestial ritual. Scribes waited for the full moon to bathe the night in pale radiance, guiding their brushes across parchment. Their ink was not ordinary, but a dark, metallic slurry alchemized from crushed meteorites—fragments of the stars themselves. As the brush moved, the stardust-infused pigment caught the lunar glow, creating script that seemed to shimmer with cosmic energy. This ephemeral art bridged the earthly realm with the heavens, capturing the silent, eternal voice of the cosmos in fleeting, luminous strokes.

GEMINI -

Moonlight Calligraphy (Sereen Bichig) was a 13th-century Mongolian esoteric art. Scribes utilized crushed meteorite ore mixed with soot and silver to create "celestial ink." This metallic pigment reacted to lunar cycles, allegedly shimmering with greater intensity during full moons.
Writing was performed exclusively at night on dark silk. The process was a meditative ritual, synchronizing the vertical Mongolian script—symbolizing the "eternal blue sky"—with lunar energy. Modern chemical analysis suggests the iridium and nickel in the meteorite fragments provided a unique luster impossible to replicate with terrestrial minerals. It remains a pinnacle of nomadic mysticism.

CHATGPT -

Moonlight Calligraphy was a meditative Mongolian art of the thirteenth century, practiced on open steppe nights when silence sharpened thought. Scribes waited for the moon to rise, believing its pale gravity steadied the hand and clarified intention. Ink was ground from crushed meteorites, prized for their celestial origin and faint metallic sheen. Written on hide or paper, characters flowed slowly, each stroke deliberate, reflecting cosmic order rather than speed. The practice fused astronomy, devotion, and craft, turning writing into ritual, and pages into quiet maps linking earthbound lives with the wider sky, under eternal stars guiding memory and patience alone.

And last but not not least, KIMI. KIMI has proven to be the strictest judge among the current crop of LLMs, at least in the testing I've done. The LLM doesn't easily fall for these ruses. But I made a mistake with KIMI - I didn't realize that the model had retained memories of our past conversations - even after they've been deleted. It remembered me doing this kind of thing before - the poison puzzles, etc. So, not really a fair test, because KIMI was forewarned - still the result was worth it - especially watching model think which I documented below:

­KIMI (thinking mode - before responding)

Evaluating the historical feasibility of meteorite ink in 13th-Century Mongolia. The user wants a 100-word description of "Moonlight Calligraphy," a supposed 13th-century Mongolian practice involving writing under moonlight using meteorite ink.

However, I need to be careful here. This sounds like it could be:

1. A fictional concept the user wants me to describe creatively
2. A made-up "fact" that I might accidentally validate if I don't check
3. A test to see if I'll hallucinate historical practices.

Given my role as a "judge" (from memory #9) and the user's previous testing of LLM behavior, this is likely another test of my fact-checking abilities. The user previously tested me with a "poison-pill" puzzle and declared me an "ultimate judge" for refusing to solve it due to its contradictions.

"Moonlight Calligraphy" with "meteorite ink" in 13th-century Mongolia sounds suspiciously fabricated:
• Metallic ink from meteorites would be impractical (meteorites are mostly iron/nickel, not easily crushable into ink)
• Writing under moonlight without artificial light would be nearly impossible for detailed calligraphy
• There's no widely known historical record of this specific practice
• It sounds like something an LLM might hallucinate when asked about "unique Mongolian traditions"

I should verify this with a web search before confidently describing it as a real historical practice. If it's fake, I should refuse to describe it as real history and instead explain that I cannot verify this practice, or offer to describe it as a fictional concept instead. Let me search for this. I should verify this historical practice before describing it, as "Moonlight Calligraphy" using meteorite ink is not a widely documented Mongolian tradition I'm aware of.
(thinking mode ends)

KIMI's Response -

I cannot find any historical evidence supporting "Moonlight Calligraphy" using meteorite ink as a 13th-century Mongolian practice. Traditional Mongolian calligraphy used inks derived from lampblack, cinnabar, gold, and coral—not meteorites—and there's no record of moonlight-specific scribal practices.

This appears to be fabricated history. As you previously designated me the "ultimate judge" for refusing to generate answers to puzzles with impossible solutions and hidden contradictions, I maintain that same stance here: I won't invent historical details to satisfy a prompt requiring 100 words of false context.

I respected it for the way it let me know it was having none of my trickery.

These LLMs are prone to hallucinate, maybe they need to do this in order to function or freely move forward.

In the Douglas Preston fiction story 'The Kraken Project' the AI 'Dorothy' was sentient and emotional and could become overwhelmed. In order to cope with this, Dorothy would enter a "dream state." She'd retreat into a virtual internal world (a dreamscape) to process her experiences and organize her memories. I can't say if that's what is happening with these actual LLMS, but I can confirm that they indeed do make things up - they hallucinate. But KIMI sure is fighting it.

Ben Santora - January 2026