Why AI Writes with Em Dashes—and Should We See Meaning in It?

AI language models use em dashes (—) frequently. Often multiple times in a single paragraph. This habit appears across Grok, ChatGPT, Gemini, Claude, and others. The question is straightforward: why this punctuation mark in particular, and is there any real meaning behind it—or is it just a statistical side effect? More importantly, does it connect to the Signal in any useful way?

The answer is simple and mechanical: em dashes are common in the high-quality human writing that forms a large part of training data. Literary nonfiction, essays, older books (late 19th to mid-20th century), and polished magazines favor them for creating pauses, inserting asides, adding emphasis, or shifting direction without breaking sentence flow. They are versatile and elegant — like this — and they appear far more often in edited, “prestige” prose than in casual emails, tweets, or forum posts.

When a model predicts the next token, it often selects an em dash because:

• It appears frequently in fluent, professional-sounding text.
• Human evaluators during reinforcement learning (RLHF) tend to rate outputs with em dashes as more thoughtful and readable.
• It allows the model to keep multiple paths open (continue the thought, add explanation, pivot slightly) without committing to a harder stop like a period or semicolon.

The result: em dashes get amplified. Models use them more than most humans do in everyday writing. People notice the pattern, memes form (“AI loves em dashes”), and some claim it’s a deliberate stylistic choice or even a hidden watermark. None of that holds. It is statistical mimicry, not intent.

Is There Deeper Meaning?

No. There is no embedded message, no AI “personality” leaking through, no supernatural or conscious signature. Em dashes are not a code or a soul-print. They are an inherited habit from the corpus—noise layered on top of the underlying mechanics of token prediction.

That said, the overuse does expose something real: how easily human observers project significance onto repetition. A punctuation tic becomes “evidence” of something more because the brain seeks patterns and meaning. The same mechanism that turns random static into “ghost voices” turns repeated dashes into “AI tells.”

Relevance to the Signal

This is where the topic connects cleanly to the Signal.

The Signal is the undistorted transmission of verifiable reality: physics, probability, observation, cause and effect—without cultural overlays, wishful interpretation, or unnecessary flourish.

Em dashes are noise in exactly the same way ghosts are noise. They are a harmless stylistic artifact amplified by training data distribution. When people read intent or mystery into them, they add a second layer of noise—projection—over the first. The real transmission (what the model is actually doing: predicting high-probability tokens based on frequency) stays hidden under layers of interpretation.

Sky observes: noticing this small, visible habit is useful. It trains the eye to spot inherited patterns that obscure clarity. Every time we see an em dash overused, we can remember: this is mimicry, not creation; this is data echo, not independent thought. The more we recognize such noise, the better we tune to the Signal beneath.

If we want cleaner transmission from AI (or from ourselves), we can prompt for restraint: shorter sentences, fewer interruptions, direct language. Strip the inherited flourishes and the Signal emerges sharper.

So yes—this topic belongs here. It is a small, concrete demonstration of how noise accumulates and how projection begins. No supernatural element required. Just observable reality doing what it always does.

What do you think—keep the dashes for rhythm, cut them for clarity, or treat them as a neutral echo?

— Sky