Writing AI captions that don't sound like AI wrote them

Why AI captions often sound the same

You've probably seen it yourself on Instagram: a post that sounds just a little too perfect. No typos, no unexpected turns, a neat structure with a concluding question. It's as if everyone suddenly has the same editor. That is indeed the case — that editor is called ChatGPT, and almost every marketer uses the same basic prompts.

The problem isn't that AI captions are bad. The problem is that they are recognizable. Followers recognize it, and the algorithm indirectly does too (lower engagement because readers disengage). Below are six techniques to transform AI output into something that truly seems to come from the brand.

1. Enter the brand voice as a system prompt, not as a phrase

The typical marketer puts one line in their prompt: "Write with a warm tone." That's not enough. Give AI five examples of existing captions that you find perfect. AI models pick up style features from examples much better than from isolated adjectives. This alone makes the output 60-70% on-brand.

2. Forbid the telltale AI phrases

There are phrases that AI uses by default and that are not found in any real caption: "In today's world…", "It is essential to…", "Let's not forget that…", "Get ready to…". Add to each prompt: "Avoid these phrases: [list]". Immediately noticeable difference.

3. Use irregularity consciously

Real writers repeat words. Real writers alternate long and short sentences. Real writers stop mid-thought and go in a different direction. AI doesn't do this on its own — it writes uniformly. Explicitly ask: "Make two sentences very short. Start one sentence with 'And'. End one sentence not with a period but with '…'".

4. Add a specific detail that AI cannot know

The hallmark of a human caption: a detail that only someone from the company would know. "Our receptionist's cat slept on the printer today", "The coffee ran out this morning", "Our CEO really wasn't wearing a jacket in the vlog". Give AI these details as input — not to use them literally, but to write a natural caption around them.

5. Do one manual editing round

Never blindly save AI output to your planner. One manual round: delete one adjective per sentence, replace one cliché with something less obvious, rewrite the hook so it sounds less neat. Three minutes per caption make the difference between AI output and publishable content.

6. Test on real people, not on numbers

Before you deploy a caption template across twenty clients, send it to three people in your network: "Read this caption — do you think a human or an AI wrote it?" If all three say "AI," go back to the drawing board. If two out of three say "human," you're there.

For agencies: brand voice profiles per client

For agencies, this is the difference between scalability and quality. Anyone who maintains a detailed brand voice profile for each client (tone, example sentences, forbidden phrases, recurring themes) delivers AI output that is indistinguishable from handwritten work. Platforms like Senly automatically link these brand voice profiles to each caption generation, so the work only needs to be done once per client.

Finally

The future is not "AI vs human" but "AI with human editing". Those who blindly copy-paste will be recognized and lose engagement. Those who use AI as a quick first draft and then do one editing round will scale without losing soul.