Why Em-Dashes Give Away AI-Generated Briefs

The most reliable AI tell in a piece of legal writing is the em-dash. Other tells exist. The em-dash is the one that appears most consistently and the one trained readers spot fastest. Lawyers who care about how their briefs read on the other side of the bench should care about this.

This article is what I have learned about AI tells in legal writing, why they show up, and how to strip them out without giving up the productivity gains AI provides.

Why the em-dash specifically

The em-dash is a punctuation mark that signals an aside, an interruption, a thought-finishing flourish. Used sparingly, it adds rhythm. Used heavily, it reads as a tic.

Generative models trained on the open internet picked up the em-dash from a corpus that contains a lot of long-form writing styles where em-dashes are common: blog posts, magazine essays, certain literary genres. The models then deploy the em-dash everywhere, often two or three to a paragraph. The pattern is so consistent that reverse-engineering tools and trained human readers both flag it.

In legal writing the pattern is more conspicuous. Briefs do not need em-dashes. The conventions of formal legal writing favor commas, semicolons, and periods. A brief peppered with em-dashes looks unlike legal writing produced by a lawyer with an editor in their head.

Once a reader notices the pattern, they cannot unnotice it. The brief reads as AI. The lawyer's professional voice is replaced by the model's default voice in the reader's perception, regardless of how much editing the lawyer did before filing.

The other tells

The em-dash is the most consistent tell. There are others.

A particular kind of soft transitional phrasing. “It is worth noting that,” “it is important to remember,” “in this regard,” “with respect to.” Lawyers use some of these occasionally; AI models use them in chains, often two or three in a paragraph, often at sentence openings.

A specific cadence. Three-clause sentences with parallel structure, deployed back to back. The cadence reads as polished but quickly becomes monotonous. A human writer's cadence varies more.

Hedging modifiers. “Generally speaking,” “broadly speaking,” “in many cases,” “often.” These show up where a confident lawyer would have written a flat declarative.

Closing-paragraph summaries that restate what was just argued in slightly different words. The summary adds nothing and signals a model trained to wrap up sections cleanly.

Adjective stacks. “Robust and comprehensive,” “thorough and methodical,” “careful and considered.” Two adjectives where one would have done.

A weakness for “delve” and “leverage” and “navigate” used as verbs. Lawyers say “address” and “use” and “manage.” Models say “delve into” and “leverage” and “navigate.”

None of these on its own gives away an AI draft. In combination, especially with em-dashes, they do.

Why this matters

A judge reading an AI-tell-heavy brief notices. Not always consciously. The brief reads as somehow off. The off-ness gets attached to the lawyer's professional credibility before the judge has even worked out why.

Opposing counsel notices the same things and is sometimes hunting for them deliberately. The legal-tech press has been documenting opposing-counsel patterns of identifying AI-generated content for use in sanctions motions. A brief that reads as AI invites the verification-style attack: was the cite hallucinated, did the lawyer Shepardize, was the brief signed under good faith.

Clients notice in different ways. A demand letter or a contract that reads as AI invites doubt about whether the lawyer's attention is actually on the matter or whether the work was outsourced to a tool.

None of this is fatal on its own. All of it is avoidable.

Why the tells are hard to remove

Two reasons.

The first is that the tells are baked deep into the model's defaults. Telling the model not to use em-dashes works partly. Telling it not to use the soft transitions works partly. Models drift back to default behavior over long passages, especially in second and third drafts where the standing instructions weaken.

The second is that lawyers reviewing AI drafts are not always reading for the tells. The lawyer is reading for substance. The tells slip past on a substance-focused read because the substance is fine; the tells are stylistic.

The discipline that catches the tells is a separate read-pass focused only on tells. It is fast once it becomes a habit. Without the habit, tells survive.

How to strip them

A few specific moves I run on every AI-assisted brief before filing.

The em-dash audit. Search the document for em-dashes. Replace each one with the appropriate alternative. A comma. A semicolon. A period. An actual sentence break with a fresh subject. Most em-dashes the model generated should not have been em-dashes at all; they were pauses where the model defaulted to its punctuation tic.

The “it is worth noting” pass. Search for “it is worth noting,” “it is important to,” “it should be noted,” “it is interesting that.” Cut every instance unless the phrase is doing actual work, which it usually is not. The sentence that follows the cut phrase reads cleaner without it.

The adjective-stack reduction. Search for common adjective pairs the model favors. Pick one. Cut the other.

The verb upgrade. Replace “delve into” with “address.” Replace “leverage” with “use.” Replace “navigate” with “manage” or whatever the actual verb should be. The simpler word is usually right.

The closing-summary cut. Look at the last paragraph of each section. If it summarizes what the section just said, cut it. The reader does not need the summary. Models generate the summary because they were trained to wrap up sections; lawyers do not need to wrap up sections that way.

The voice check. Read a paragraph aloud. If it does not sound like the lawyer talking, rewrite it. AI tells survive silent reading better than they survive aloud reading.

These six passes take about 15 minutes on a 10-page brief. They produce a document that does not read as AI.

The standing-instruction approach

A complementary move, upstream of the verification passes, is to put the tell-prevention into the model's standing instructions.

Tell the model, in the project setup, not to use em-dashes. Use periods, commas, colons, semicolons. Tell it not to open paragraphs with “it is worth noting” or similar transitions. Tell it not to stack adjectives. Tell it to prefer short Anglo-Saxon verbs over Latinate ones. Tell it not to summarize sections at the end.

The instructions reduce the number of tells in the first draft. They do not eliminate them. Models drift. The verification passes are still required.

Why this is not a vanity issue

Some lawyers I have talked to about this treat the AI-tells issue as cosmetic. A brief that argues well will succeed regardless of whether it has em-dashes.

The cosmetic frame underestimates the audience. Judges and opposing counsel are trained readers in a profession where written craftsmanship matters. A brief that reads as written by a lawyer carries an implicit signal: this lawyer cares about the work, knows the conventions, and is paying attention. A brief that reads as written by AI carries the opposite signal.

The signal compounds with other questions about the lawyer's AI use. If the brief reads as AI and contains a hallucinated cite, the cite does not look like a careless mistake; it looks like a tool-use failure. If the brief reads as the lawyer's voice and contains a hallucinated cite, the cite looks like the kind of error any lawyer might make under pressure. The framing matters.

Stripping the tells is not vanity. It is professional presentation. The professional presentation of the work product is part of the work product.

Frequently Asked Questions

Are em-dashes always wrong?

No. Used sparingly and intentionally, em-dashes are correct. The problem is the model's overuse, which produces a recognizable pattern. One em-dash in a brief is usually fine. Six are not.

Will the next generation of AI models fix this?

Probably partly. The tells shift over time. The em-dash pattern was less prominent a few years ago and may be less prominent in a few years. New tells will emerge. The discipline of reading for tells will remain useful regardless.

Is there a tool that strips the tells automatically?

Some editing tools claim to. Mileage varies. The verification passes described here are fast enough that the tool is not strictly required. The lawyer doing the passes also catches things specific to the brief that no general tool would notice.

Does this apply to legal writing other than briefs?

Yes. The same tells appear in demand letters, motions, contracts, opinion letters, and client memos. The verification discipline transfers.

How does this connect to the broader verification framework for AI legal content?

The tell-removal pass is layered on top of the citation and argument verification. It does not substitute for those. A brief can be verified for accuracy and still read as AI, which is its own problem.

What if the lawyer actually wants to disclose AI use?

Disclosure is a separate strategic question. The decision whether to disclose is upstream of the question whether to read as a lawyer. Even with disclosure, briefs that read as a lawyer present better than briefs that read as a model.