How to make AI writing sound human — 3 big signs to avoid

AI-written content has a smell. Readers can't always name it, but they can feel it within two sentences and they bounce. The smell is made up of three specific patterns that virtually every default-tuned model emits constantly. Strip those three and your output reads like a person wrote it.

Here's how to fix it — and the exact custom prompt you can paste into Claude's settings today.

Why AI slop kills your views

The three patterns aren't subjective taste preferences. They're tells. Once a reader spots a pattern, every following sentence is suspect. Trust collapses, the audience scrolls. Killing the patterns isn't about style — it's about retention.

Mistake 1 — robotic emojis and signs

Default ChatGPT loves to drop a sparkle emoji or a checkmark in front of every list item. Nobody texts like that. The fix: ban emojis and decorative symbols outright in your system prompt. They almost never improve information density.

The secret — switch from ChatGPT to Claude

ChatGPT has a stubborn baseline tone that's hard to break out of. Claude is dramatically more steerable — give it a system prompt and it actually holds the constraint across long outputs. If you've been fighting ChatGPT to sound human, switching to Claude is the single biggest improvement you can make.

Mistake 2 — eliminate em dashes

The em dash. Every AI loves it. Real writers use them sparingly. The fix: ban them in the system prompt. Force the model to use periods, commas, or sentence breaks instead.

Mistake 3 (the big one) — forced negation

The "not only X but Y" pattern. The "not just a tool, but a revolution" cadence. Repetitive parallel phrasing. Staccato three-word sentences. These are the deepest tells in the AI-output fingerprint, and they're everywhere.

The secret prompt — copy this into Claude

Drop this directly into your Claude general settings:

"Never use em dashes in any response. Very important. Eliminate all emojis and signs. Avoid AI-sounding writing patterns: no forced negation ('not just a tool, but a revolution'), no staccato sentences, no repetitive parallel phrasing. Write naturally like a person would talk or type. If something can be said in one sentence, say it in one sentence."

Save it once. Every Claude conversation from then on inherits the constraint.

The result — writing that gets views

After switching to Claude with this system prompt, the difference is immediate. Outputs read like a person typed them. No tells, no slop, no smell.

Apply this to video ads too

Scripts have the same patterns. Use the same constraints when you ship AI-written ad copy through the AI Media Machine. The platform finds the best-performing ads in your niche, clones their winning visual patterns, and inserts your product into the video — and now the script underneath sounds human, not robotic.