How to Distinguish if a Text was Written by a Human or an AI?

Well, let’s be honest, almost all of us have moved to AI. When I say “all,” I mean people like me who create content online… I even see some who criticize generative AI all day long while using it. You can feel it in their articles. Or maybe it’s because they’re lifting content created by AI? I don’t know.

But we must admit, it’s chaotic!

Personally, I conduct a lot of tests here on AI-assisted writing. I say “assisted” because I refuse to go into “automatically generated content mode” like some media outlets do. However, it’s true that I use it, and I hope in the smartest way possible. So when I need to analyze a topic for an article, I first ask the AI for a comprehensive analysis.

I then have a working document that helps me understand the concepts related to the topic, and especially to have analytical angles I might not have thought of. I also get a few examples of introductions that I can draw inspiration from if I’m stuck with writer’s block, and a draft outline that I’m free to follow or not (generally, I don’t because I’m lazy…).

From there, I can start writing the old-fashioned way as usual. And when I’m done, I submit the article to my favorite LLM (namely Claude), which critiques my article and gives me tips to improve it. If I ask, it even makes slight modifications to integrate the information I might have forgotten and its feedback.

So, I’m an AI-assisted blogger, and I’m fine with that because I use it as a tool to support me, just like I use Wikipedia, a spell checker, or sometimes an SEO tool. And of course, I read and validate everything. I am the sole master on board, and I am 100% responsible for what is published under my name. By the way, I noticed that by 2026, websites that use AI will need to clearly disclose it. I looked into it a bit, and apparently, the way I use it isn’t subject to this warning because there is human oversight and validation of what is published.

Now when it comes to images, that’s another issue. In the past, I used royalty-free images (and my own images), but that caused me problems because there are unscrupulous companies that put royalty-free images online for a few years, just long enough for someone like me to use them and then remove their CC0 status (or similar) to send lawyers demanding money for photo theft. So today, I generate them using AI, and that works perfectly for me.

As a reader, I’m accustomed to and can immediately notice when an LLM has been used in generating an article (hence my intro), and at a time when Gemini, ChatGPT, and similar tools are flooding the web with automatically generated texts, it’s not easy to separate the wheat from the chaff. Fortunately, Australian researchers have discovered the invisible signature that reveals artificial writing.

With 98.3% accuracy, their algorithm can distinguish between a text written by a professional journalist and one produced by AI, even when your eyes see no difference. And it’s not just a subjective impression like mine. These researchers from Charles Darwin University conducted a rigorous scientific study comparing 150 articles written by award-winning journalists with 150 articles generated by Gemini (Google’s new toy that has replaced Bard) on the same topics.

It turns out that humans and AIs have different “stylistic fingerprints” that can be measured and quantified. It’s a bit like recognizing a robot by its too-regular gait… Yep, AI tends to write with a metronomic cadence that betrays it in the eyes of algorithms.

The five most revealing characteristics are:

1. Sentence Length Range: We humans sometimes write very short sentences and sometimes lengthy ones that wander off before returning to the initial subject after a detour through our disordered thoughts. In contrast, AI is more consistent.
2. Paragraph Variation Coefficient: Our paragraphs vary greatly in length, while those of AI are more uniform.
3. Verb Ratio: Humans use about 16% verbs compared to 13% for AIs, which isn’t a huge difference, but it matters.
4. Grammatical Complexity: Our sentence structures are less predictable (well, that depends on who…lol).
5. Paragraph Variability: Again, AI loves order and consistency too much, which gives it away…

The algorithm that detects all this is called Random Forest, which counts each comma and analyzes every sentence structure to see what repeats the most. The most amazing thing is that this method achieves 98.3% accuracy!

Another interesting discovery is that AI prefers nouns over verbs. Indeed, it describes the world rather than putting it into motion. It’s a bit like a beginner sports commentator saying “hey, a pass from Messi” instead of “Wow, Messi dribbles and passes the ball!!”. It’s less dynamic and much more static.

Now, if AI produces content that is more standardized and thus “boring” but equally readable and informative, is this really a problem for readers? I don’t think so.

Anyway, it doesn’t bother me as long as it’s cool and enjoyable to read. Especially since some types of basic journalistic content that lack “artistic” value, like financial reports, sports summaries, or coverage of predictable events, I believe can be fully delegated to AIs. This frees up time for humans to focus on their added value: investigation, deep analysis, emotion, and creativity.

And especially, the way I explained my “collaborative” usage allows time savings while adding that touch of creativity that machines cannot (yet) reproduce.

In short, one thing is certain: detection tools like the one developed by Charles Darwin University will be essential to maintain transparency in this evolving ecosystem. But we also must be careful not to fall into a ridiculous witch hunt because everyone remains free to read what they like, regardless of who writes it (AI or human), as long as the information is reliable and published responsibly.