How to use AI for content without sounding like AI

Every creator using ChatGPT, Claude, or Gemini for captions hits the same wall: the output is grammatically perfect, technically on-topic, and somehow completely lifeless. Audiences smell it instantly. Engagement craters.

The "AI smell" isn't mysterious. It's four specific mechanical signals that LLMs default to when nobody tells them otherwise. Fix those four signals and the same model produces captions indistinguishable from human writing.

# Signal 1 — the em-dash addiction

GPT-4 and Claude default to em-dashes (—) about 5x more often than human writers. It's the single fastest tell. If a caption contains 2+ em-dashes in 200 words, readers' AI-detector instincts fire.

Fix: explicitly forbid em-dashes in your prompt, or use a tool (like GrowhtOS) that strips them in post-processing.

# Signal 2 — the "perfect-grammar" trap

Humans write with sentence fragments. They use commas where periods should go. They start with "And" or "But." AI by default writes like a high-school English teacher.

The fix is counter-intuitive: tell the model to write less correctly. Real instruction:

Write the way a person texts a friend. Fragments OK. Start sentences with And or But when it sounds right. Don't fix run-ons that have rhythm.

# Signal 3 — the four-bullet structure

LLMs love structure. Heading → intro → 4 bullets → conclusion. Humans don't write this way in captions. Audiences notice immediately.

For social captions specifically: ban bullets entirely in the prompt. Write in flowing paragraphs or single-sentence punches with line breaks.

# Signal 4 — the absence of specifics

The biggest tell of all: AI captions are about things in general. "Building a brand requires consistency." Human captions are about one specific moment. "Last Tuesday, I almost deleted my account at 11:47pm."

Specifics — names, dates, dollar amounts, weather, the exact thing someone said — are the signal humans use to mean "this is real." Force AI to include them.

# The prompt that actually works

Here's the skeleton of a prompt that gets human-sounding output from any modern LLM:

Voice: I write the way I text. Fragments OK. Start sentences with And or But.
       No em-dashes ever. No bullets in captions.

Rules:
- Include at least one specific detail: a name, a date, a dollar amount,
  or a quote of what someone actually said.
- Open with a sentence under 8 words.
- End on a question or a punch line, never a CTA.

Topic: {topic}
Examples of my voice: {3 of your past best captions}

The "examples of my voice" part is the magic. Few-shot examples drag the model's output toward your actual rhythm. Without them you get default-AI-voice no matter what you specify in rules.

# The two-pass critique

The most powerful technique we've found — and the one GrowhtOS uses on every generation — is to have the model critique its own draft, then rewrite. First pass produces caption. Second pass scores it 0-100 on "does this sound like the user's actual past captions?" If under 70, rewrite with specific feedback. This catches 80% of AI-smell issues.

This is purely a prompt trick — no special API. You can do it in raw ChatGPT by pasting your draft back in and asking "what makes this not sound like me? Rewrite it."

# What never works

A few approaches that look smart but don't move the needle:

• Asking for "casual tone" — every model has a default "casual" that all read identically. Be specific about your casual.
• Increasing temperature — produces weirder output, not more human output.
• Switching models — Claude doesn't sound human in default mode either. Same four signals.

The fix is always the same: feed examples of your actual voice + explicitly ban the four AI smells. Do that and you'll get drafts you only need to lightly edit, not rewrite from scratch.