I went into this thinking that an AI frontier model, facilitating trillions of calculations per second, would easily write like a human. This is the fourth article I’ve written on this. My first article was generated from a single prompt with Claude, followed by passing it through an AI editor, a few tweaks, a couple of deletions, and the article was done. I posted it the draft and ran it through Pangram. It came back 100% AI. I didn’t publish it.
Find 100% Human and Copy
What did I do next? I decided to use Claude to assess and review 100% human-written articles, to create an actual formula for what makes writing read as human. I scraped a total of 56 articles, around 185,000 words. And from there, it developed a set of criteria to establish a formula that mimics human writing.
The areas it looked at: entity density, which is proper nouns, acronyms, and alphanumeric product names per hundred words. The next thing it looked at was dash rate, em dashes, trying to have a balance of those throughout. Transition density, which is formal connectives: furthermore, moreover, in conclusion, so those within the three-hundred-word frameset. A vocabulary rate to make sure it matched against a list of overused machine writing words, to minimize that.
The next part it looked was the structural measures, paragraph irregularity, so the size and framing of them. And then other things it looked at were the cadence measures. The mean sentence length, standard deviation to that sentence length, and then the key one is burstiness, which is borrowed from physics, where it describes where the events cluster arrive at regular intervals, applied within a sentence.
This is what it pulled together to shift it writing output to being human. The outcome of this, from my perspective read exactly like my voice, exactly what I consider to be human. I uploaded the article and once again, ran it through Pangram. The result, 100% AI. I didn’t publish it.
Averages of All vs One
This time, my assumption was that it was because I was using the average of fifty plus articles and because it was the average this is why it was detected as AI. The reason, there wasn’t enough variation with what had been generated. Therefore, I opted to focus on a single writer and used his content, his frequency, his sentence length, his lexical diversity, his pronoun use, his use of em dashes, his vocabulary, as a means of structuring the article. This time I had AI produce a shorter piece, a two-hundred-word count article mirroring this writer’s style. Once the article was generated, I uploaded its and ran it through Pangram, the result, 100% AI. I didn’t publish it.
Defeat
For the final phase of this battle, I asked Claude to audit its output, and I asked specific questions around why its content still came up AI.
In response, it confirmed it was good at mimicking the style of a human. You have conventional AI detectors, they use key phrases, dash rates, sentence length, variance, these are all surface stats. It the humanizer, write like a human skill we’ve all come across and Claude can do this.
The key difference with Pangram is that it’s a trained classifier. It doesn’t focus on surface features such as burstiness, sentence length, or dash rates, it looks for deeper structural, meaning, and patterns learned from both human and AI writing. In other words, LLMs generate language in a way that has a structural fingerprint or pattern. It not checking whether a piece merely “looks” human, it is seeing LLM identifies we can’t perceive.
My conclusion: at present I can’t defeat Pangram based on my current use of AI but I thinks it’s doable. I’m also in two worlds about content creation and writing. On one hand, it’s faster to produce content with AI. On the other hand, I prefer the 100% human moniker.
With this article. I’ve had AI draft it based on my thoughts and intent. I’ve then dictated it back to Claude in my own words. I’ve had Claude correct spelling mistakes, remove ums and recurrences with no structural changes. I’ve then edited it twice across two different days. The result 100% human but it was touch and go there with Claude editing.
Thanks for reading! I build things with AI and share what I learn as I go. Subscribe. It’s Free. Subscribe on Substack.