From 'obviously AI' to 0% on AI detectors: human-sounding content for a client

Full walkthrough

At this point, most people can start reading a blog post and immediately tell it was written by AI. In this case study I'll show how I solved that problem for my client, and the key insight that led to the solution.

The client and the problem

I was working with a Denmark-based health and well-being company. Like many companies, they used AI to write blog posts for their website, and they weren't happy with the quality - they ended up rewriting most of it. The client wondered if there was a way to keep using AI but get genuinely higher quality out of it.

One quick note: the demos here are in English so they're easy to follow, but the client's content was all in Danish, and that was its own problem. Most of the content on the internet is English - roughly 640 million websites, versus about 5 million in Danish. Since AI is trained on that internet content, it's much worse out of the box at anything with small representation online, Danish included. I'll demonstrate in English to keep it clear.

Showing the problem

First, the starting point. I generated a blog post with a reasonable prompt - not just "write me a blog post," but one that explains the audience, the voice, the language, and the topic. The result reads like AI. To make that objective, I paste it into an AI content detector. This one is good, and it flags the text as 100% AI.

Just to show the detector is fair, let me test it on human writing. Here's a trick to find guaranteed-human text: Google a topic and add "before:2007" (OpenAI was founded in 2015, so 2007 gives a comfortable buffer). Paste that older, human-written text in, and the detector says 0% AI.

The key insight

Now watch this. I wrote a short piece myself, by hand, right now - and the detector says it was likely written by AI. Why? Because I deliberately wrote it in the same tone and style that AI uses. That's the whole insight: if I, as a human, can imitate how AI writes, then AI should be able to imitate how a specific human writes - not 50%, not "a bit," but a full imitation of one person's persona, style, and voice.

The solution

Here's the application I built for the client - a small blog manager where you can create, edit, and publish a post. The important part is the writing. I take the same topic and, this is the key step, I give the AI a reference example of content written by one real human. For the client, I asked for at least 14 pages of Danish content written entirely by a person - whoever wrote their blog posts before, or a copywriter. The AI picks up on that person's style and imitates it.

I drag and drop the human-written reference, generate, and paste the result into the detector: 0% AI, and on topic. I tried it on another topic - a piece about well-being as a business investment - using the same reference, and it came out at 0% again. It even reuses the kind of dry humor the reference author used, because it's imitating that specific person rather than inventing "AI jokes." Side by side, the AI-flavored version scores 100% and the imitation scores 0%, and you can feel the difference in style and formatting. It reads like a real person wrote it. From there it just needs a light edit, a few words, and it's ready to publish.