I Made the Rules and I Can’t Follow Them

February 21, 2026

I work with AI every day. Not casually. I co-author research with it, build frameworks through it, run arguments against it to find the weak joints. My deployment methodology has four tiers, and I operate at the upper two regularly, the levels where a domain expert curates AI output with full epistemic authority. This is serious, structured collaboration, and I’m not embarrassed about any of it.

Which is exactly why I have rules designed to make sure my writing doesn’t read like AI wrote it.

Two of those rules are killing me.

The Em Dash Problem

Rule: No em dashes. Use alternatives. Occasionally an en dash is acceptable.

If you’ve spent any time reading AI-generated text, you’ve seen the em dash epidemic. ChatGPT, Claude, Gemini all scatter em dashes throughout their outputs as if signaling sophistication. It’s a statistical artifact. Because quality writing employs them, models learned to overuse them regardless of necessity.

The result: the em dash has become a tell. Readers have developed instincts recognizing it. Encountering multiple dashes triggers suspicion: this was generated. Maybe it was, maybe it wasn’t. The suspicion alone is problematic.

So I banned them. Not because they’re poor punctuation, but because I refuse letting my work get sorted into the “AI slop” category by pattern-matching readers. My published prose needs to remain unmistakably mine, stripping anything triggering slop detection.

Here’s the problem: English needs the em dash.

Certain moments demand punctuation between commas and periods. You need inserting asides without losing your thread. When sentences have earned interruptions, nothing else accomplishes this work properly. The em dash isn’t decoration then; it’s structural.

I recognize this drafting regularly. I reach for one, not habitually anymore after months of resistance, but because sentences require it. Alternatives prove worse. Semicolons feel too formal. Parentheses feel too quiet. New sentences fracture momentum. The em dash sits there, perfectly suited for this exact function, yet banned by my own rule because machines overuse it.

“You can tell a lot about a writer by the rules they make and then can’t keep.”

AI models damaged the em dash through excessive affection. Human writers sharing this love must now choose between correct punctuation and correct optics. This shouldn’t require choosing. But it does.

The Symmetry Problem

Rule: Avoid symmetric reversals. No “not X, but Y” or “less X, more Y” formulations. Let distinctions emerge from argument.

Same diagnosis, different manifestation. AI prose fixates on symmetric reversals: “It’s not about efficiency, it’s about effectiveness.” “Less optimization, more intention.” Every model does this constantly because the structure feels clean, appears insightful, and training data rewarded it.

Yet it feels insightful without being insightful. The formula does work that argumentation should accomplish. You insert “not infrastructure, but epistemology” and readers nod, but you haven’t demonstrated why one matters more. You’ve merely asserted preference with pleasant rhythm.

I banned it similarly: refusing philosophical arguments mistaken for model-generated content prompted to “sound thoughtful.”

But symmetric reversals (like em dashes) predate AI and prove more valuable than their abuse suggests. Aristotle employed them. Paul did too. Every philosopher sharpening distinctions reached for this tool. “Not X, but Y” persists across twenty-five centuries because human thinking operates through contrast. Genuine opposition sometimes exists; naming what something isn’t, then what it is, clarifies positions fastest.

“Banning a rhetorical move because machines overuse it is like banning a hammer because a robot hit your thumb with it.”

The tool wasn’t the problem. AI flooding the zone with cheap imitations made authentic versions appear counterfeit.

What the Rules Actually Reveal

Here’s my realization: these rules concern provenance, not writing style.

Publishing research demands readers trust human minds shaped arguments. I work with AI collaboratively and transparently, but ideas originate from someone with domain expertise, capable of being wrong meaningfully, staking reputation on claims. Writing must reflect this. It must bear marks of consciousness, not computation.

My em dash and symmetry bans attempted creating distance. Stripping patterns AI colonized leaves recognizably human work. My instinct was right. My execution overcorrected.

Because avoiding every pattern AI uses means I’m not writing like myself anymore. I’m writing against machines, meaning models still shape my voice. I’ve given the tool different authority. Instead of imitating it, I’m defined against it: another form of letting tools control craftspeople.

Real solutions require discipline, not elimination. Use em dashes when they’ve earned their place. Deploy symmetric reversals when contrasts prove genuine and argumentative work supports them. Avoid both when purely decorative. This demands judgment on every sentence (harder than blanket bans) but judgment separates originated thinking from derived output.

I’m not entirely rescinding my rules. They’ve accomplished real work. My prose tightened over a year; I’m more conscious about reaching for tools habitually versus purposefully. Awareness represents the actual victory.

But I’m abandoning pretense that elimination works. Em dashes remain available. So do symmetric reversals. I’ll use them less. I’ll use them better. I’ll observe when they appear and question whether they’ve earned inclusion or whether I’m simply mirroring models.

My rules taught me suspicion toward my instincts. That’s valuable. But suspicion differs from elimination. And that, if you’re tracking, is simultaneously a symmetric reversal and em dash construction.

I’m keeping it in.

---

James (JD) Longmire is a Northrop Grumman Fellow conducting independent research on AI epistemology and governance.