Why AI Stopped Using the Em Dash – and Why You Should Too

The em dash has become one of the most recognisable signals of AI-generated text. Not because using one proves AI was involved, but because AI systems overuse them in patterns that trained eyes spot immediately. Twelve em dashes in the first 500 words, combined with section breaks every 300 words and repetitive sentence structures, is a cluster that points to a machine, not a writer.

The debate around this is often framed badly. One side says em dashes are a definitive AI tell. The other responds that plenty of human writers use them. Both are right and both are missing the point.

It Is About Patterns, Not Individual Markers

No single word, phrase, or punctuation mark proves AI authorship. The tell is the cluster. AI writing tends to combine em dash overuse with other patterns: bullet points where prose would serve better, section headers every few paragraphs regardless of whether the content warrants them, intensifiers like “significantly” and “crucially” deployed as padding, and phrases like “it is worth noting” that exist to fill space rather than carry information.

An editor reviewing a piece for AI use is not counting em dashes. They are looking at the whole pattern. Em dashes are visible enough that a high count immediately prompts closer scrutiny of everything else. That is the practical problem with overusing them: they draw attention.

Why AI Systems Overused Them

Em dashes became associated with AI output because early large language models learned from text corpora that included a high proportion of edited, formal prose where em dashes appear legitimately. The models reproduced the style, including the punctuation, but without the judgment a human writer applies about when the device earns its place.

A human writer uses an em dash to create a specific effect – interruption, emphasis, a parenthetical that feels more forceful than brackets. They use it deliberately and sparingly. AI systems, optimising for fluency and variety, reached for it constantly because it is a syntactically safe choice that superficially resembles sophisticated writing.

The result was AI output saturated with em dashes. Readers and editors noticed. The association formed.

The Response from AI Developers

The pattern became self-aware. AI writing guidance, published preferences files, and fine-tuning instructions began explicitly banning em dashes. Claude, for instance, can be instructed via user preferences to never produce em dashes under any circumstances – commas, semicolons, hyphens, and full stops only. The instruction is absolute and it holds across every conversation.

This is not concealment. It is the same discipline a copy editor applies when house style bans a construction. The writing is cleaner for it. Em dashes are a crutch that most prose does not need. Remove them and the sentences either tighten or reveal that they were carrying more weight than they should.

The Case for Dropping Them Entirely

The AI association is reason enough to be cautious about em dashes in any context where the origin of the writing might be questioned. But there is a stronger reason that has nothing to do with AI detection: the writing is usually better without them.

An em dash most often signals that a sentence has been constructed in a way that requires patching. The dash papers over a join that a better sentence would not need. Force yourself to rewrite without it and you frequently end up with two cleaner sentences, or a clause that actually earns its place in the structure rather than dangling off the end.

Editors who flag more than two em dashes per page are not being pedantic. They are identifying a writer who has not yet developed the discipline to build sentences that stand without the prop.

What This Means in Practice

For anyone using AI tools in their writing workflow, the practical steps are straightforward. Ban em dashes in your AI preferences or system prompt. Audit your own writing for the broader cluster of patterns: overuse of bullet points, section headers that break up what should be continuous prose, intensifiers, throat-clearing phrases. The em dash is the most visible marker but it is one of several.

For editors reviewing content for AI use, the em dash count is a useful first filter but only a first filter. The real question is whether the whole piece reads like a human made considered decisions about every sentence, or whether it reads like fluent text produced by a system optimising for surface-level coherence.

Those are different things, and experienced readers can tell the difference. The em dash is just the part they notice first.