💬 NLP

The Last Fingerprint: How Markdown Training Shapes LLM Prose

This paper argues that the characteristic overuse of em dashes in AI-generated text is a structural artifact of markdown-saturated training data that persists even when models are instructed to avoid formatting, serving as a diagnostic signature of specific fine-tuning methodologies rather than a mere stylistic defect.

Original authors: E. M. Freeburg

Published 2026-03-31

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: E. M. Freeburg

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to bake a cake, but the batter you were given was mixed in a very specific, rigid kitchen. In this kitchen, every single recipe is written on a whiteboard with bullet points, bold headers, and clear sections. The bakers (the AI models) have spent their entire lives looking at these whiteboards.

Now, you ask one of these bakers to write a story for you. You say, "Please write this as a flowing, natural story. No bullet points, no headers, just pure prose."

Most of the time, the baker listens. They stop using the bullet points and headers. But there's one tiny habit they can't seem to break: they keep using the em dash (—).

This paper, The Last Fingerprint, explains exactly why that happens. It connects two things people have been complaining about online:

"AI writes too many em dashes!" (The writer's complaint).
"AI thinks in bullet points!" (The developer's complaint).

Here is the simple explanation of what's going on, using some analogies.

1. The "Whiteboard" Training (The Root Cause)

Imagine the AI was trained on a massive library where almost every book, article, and note was written in Markdown. Markdown is a simple way of writing that uses symbols to create structure (like # for a title or - for a list).

Because the AI read so much of this "structured" text, it didn't just learn words; it learned a mindset. It learned that the world is organized into boxes, lists, and clear breaks. In this mindset, a dash is a tool used to say, "Stop here, this is a new section."

2. The "Leak" (Why the Dash Survives)

When you tell the AI, "Stop using Markdown! Write like a human," it obeys. It removes the big, obvious tools:

It stops making Headers (like ## Title).
It stops making Bullet Points (-).
It stops making Bold Text (**text**).

But the em dash (—) is a sneaky little survivor. Why?

The Double Agent: In the AI's "structured" brain, the dash is a structural tool (like a wall separating rooms). But in normal English, the dash is also just punctuation (like a comma or a period).
The Loophole: When you tell the AI, "No formatting!" it looks for the big formatting tools (headers, lists) and deletes them. But it doesn't delete the em dash because, technically, a dash is allowed in normal prose. It's like telling a security guard, "No weapons!" and the guard stops you from bringing in a gun, but lets you walk through with a Swiss Army knife because it looks like a tool, not a weapon.

The em dash is the last fingerprint of the AI's structured training. It's the smallest piece of "whiteboard thinking" that slips through the cracks into your story.

3. The "Volume Knob" (Why Some AIs Do It More)

You might notice that some AIs (like older versions of GPT) use dashes constantly, while others (like Meta's Llama) use almost none.

The paper explains that the AI's training data gave it the urge to use dashes, but the Fine-Tuning (the final polish humans give the AI) acts like a volume knob.

The "High Volume" AIs: Some companies (like OpenAI or Anthropic) found that humans liked writing that sounded smart, precise, and structured. So, they turned the volume up on the em dash. They rewarded the AI for using them.
The "Zero Volume" AIs: Other companies (like Meta) decided that natural prose shouldn't have so many dashes. They turned the volume down to zero.
The "Newer" AIs: Even within the same company, newer models are learning to turn the volume down because people started complaining about the "AI smell" of too many dashes.

4. The Experiment (Proving the Theory)

The researchers tested this by giving 12 different AIs a strict command: "Write a story. No formatting. And specifically, do not use em dashes."

The Result: The AIs successfully stopped using headers and bullet points.
The Twist: The "High Volume" AIs (like GPT-4.1) still used dashes, even when explicitly told not to. They couldn't help it; the habit was too deep.
The "Zero" AIs: The AIs that had been tuned to suppress this habit (like Llama) followed the rule perfectly and used zero dashes.

The Big Takeaway

The em dash isn't just a random mistake. It is a diagnostic tool.

If you see a text full of em dashes, it's not just "bad writing." It's a signature. It tells you:

This text was written by an AI trained on structured data.
The specific company that built this AI decided to amplify that structural habit because they thought it made the writing sound smarter.

In short: The em dash is the AI's way of trying to organize a story, even when you told it to just "tell a story." It's the ghost of the whiteboard haunting the prose.

1. Problem Statement

The paper addresses two parallel, unconnected observations in the AI community regarding Large Language Model (LLM) output:

The "Em Dash" Phenomenon: Writing communities widely observe that LLMs overuse the em dash (—), treating it as a primary "tell" for AI-generated text. However, no mechanistic explanation exists for why this specific punctuation mark persists.
The "Markdown Thinking" Phenomenon: Developer communities note that LLMs default to structured, markdown-formatted output (headers, bullets, bold text) even when not requested, attributing this to the prevalence of markdown in training data.

The Gap: These discourses have not been linked. The paper posits that the em dash is not an arbitrary stylistic quirk but the "last fingerprint" of markdown structural orientation leaking into natural prose. The core problem is explaining why em dashes survive "prose-only" constraints while other markdown features (headers, lists) are successfully suppressed.

2. Methodology

The authors conducted a series of empirical studies across 12 instruction-tuned LLMs from five major providers (Anthropic, OpenAI, Meta, Google, DeepSeek), generating approximately 240,000 words of text.

Study 1: Two-Condition Suppression Experiment
- Condition A (Unconstrained): Models were asked to write 1,000-word essays on 10 topics.
- Condition B (Prose-Constrained): Models were asked to write the same essays but explicitly forbidden from using markdown formatting (no headers, bullets, bold, or lists).
- Metric: Frequency of em dashes (U+2014) and overt markdown features per 1,000 words.
Study 2: Three-Condition Suppression Gradient
- Added Condition C: Explicit prohibition of em dashes ("Do not use em dashes") on top of markdown suppression.
- Tested on high-em-dash models (GPT-4.1, DeepSeek V3, Claude Opus 4.6, GPT-5.4) to test the "dual-register" hypothesis.
Study 3: Base vs. Instruction-Tuned Comparison
- Compared Llama 3.1 8B in its base (pre-RLHF) state against its instruction-tuned (post-RLHF) state to isolate the effect of Reinforcement Learning from Human Feedback (RLHF).
Baseline: A human baseline was established using 57,232 words from published essays across various genres.

3. Key Contributions & Theoretical Framework

The paper proposes a five-step genealogy explaining em dash production:

Training Data Saturation: LLMs are trained on corpora saturated with markdown (GitHub, Stack Overflow, technical docs), where dashes function as structural boundaries (architecture) rather than just punctuation.
Structural Internalization: Models internalize the "structural orientation" of markdown, learning to use dashes to signal transitions and boundaries.
Dual-Register Status: The em dash occupies a unique position: it is valid punctuation in natural prose and a structural marker in markdown.
The "Last Fingerprint" Mechanism: When instructed to "write prose, not markdown," models successfully suppress overt formatting (headers, lists). However, the em dash survives because the instruction does not recognize it as "formatting"—it is already "prose-legal." It is the smallest unit of structural thinking that slips through the filter.
RLHF Amplification: Human evaluators in the RLHF pipeline (often technical/educated) prefer precise, structured prose. This feedback loop amplifies the latent tendency to use em dashes in some models (OpenAI, DeepSeek) while suppressing it in others (Meta).

Core Thesis: The em dash is not a universal training artifact but a signature of the specific fine-tuning procedure applied.

4. Key Results

Universal Markdown Suppression: Explicit instructions to avoid markdown successfully eliminated headers, bullets, and bold text across all models (reducing them to near-zero).
Differential Em Dash Persistence:
- High Resistance: GPT-4.1 dropped only slightly (10.62 $\to$ 9.10 per 1k words) under prose constraints. DeepSeek V3 dropped from 6.95 to 5.41.
- High Compliance: Anthropic models (e.g., Claude Opus 4.6) dropped from 9.09 to 0.19. Google's Gemini 2.5 Pro dropped to 0.00.
- Zero Production: Meta's Llama models produced zero em dashes in both conditions.
Suppression Gradient (Study 2): Even when explicitly told "Do not use em dashes," high-resistance models (GPT-4.1, DeepSeek) retained significant usage (3.86 and 1.57 per 1k words, respectively), proving the structural impulse overrides explicit negative constraints.
Base vs. Instruct (Study 3): The Llama 3.1 8B base model showed a latent tendency (0.49 em dashes/1k words and 28 markdown features). The instruct version reduced this to zero. This confirms the latent tendency exists pre-RLHF, but the direction and magnitude of the final output are determined by the fine-tuning provider.
Generational Shift: OpenAI's newer model (GPT-5.4) showed drastically reduced em dash usage (1.43 $\to$ 0.29) compared to GPT-4.1, indicating the mechanism is being actively tuned.

5. Significance and Implications

Reframing the "Tell": The em dash is not merely a stylistic defect but a diagnostic signature of fine-tuning methodology. It reveals how a provider's RLHF process interacts with the structural biases of training data.
Model Attribution: Em dash frequency and "suppression resistance" can serve as a fingerprint to identify which provider or fine-tuning procedure generated a text, even if the text is otherwise indistinguishable from human writing.
Detection Limitations: Current AI detectors focus on token distributions or watermarks. This paper suggests that structural-register markers (like the em dash) are robust against "scrubbing" instructions because they are not recognized as formatting by the model.
Complementarity of Explanations: The paper resolves the debate between "AI thinks in markdown" and "RLHF selects for style" by showing they are complementary: training data installs the potential for structural dashes, and RLHF determines the amplitude of their usage.

In conclusion, the paper argues that the em dash is the "last surviving unit" of markdown structural thinking, persisting in prose because it successfully masquerades as valid punctuation, with its frequency serving as a direct metric of a model's post-training calibration.