The Art of Making AI Sound Human
Veterans of the AI world—those of us who were involved just over two years ago—may remember extensive discussions about how search engines would recognize AI-generated content.
The first detection tools quickly emerged, typically looking for a single factor in the text that would reveal AI authorship. Many assumed Google Search had even more powerful tools and might "ban" AI content from search results.
The Evolving Landscape of AI Content Detection
Since then, Google has repeatedly affirmed that they don't specifically look for AI-generated content. Now, as before, it's the quality, usefulness, and relevance of the content that matters. That's what Google says, and as always, there's no reason to doubt what tech giants tell us...
By the way, that last quip about tech giants' credibility is clear evidence of human authorship. An AI tool would never write something like that—they avoid controversial statements, and perhaps that's actually a good thing? Again, this pattern of stopping abruptly and posing an unanswered rhetorical question is something AI struggles to master. You can request it, and AI will attempt it, but it often produces slightly stiff outputs, even in the best language models.
It's perfectly fine to use AI as a co-editor and "typewriter." But how do you do it?
The Human Touch in Writing
Everything you've read until now has been written by a human, word for word. The somewhat choppy style, the attempts at humor: It's not necessarily good writing, but it's definitely not AI. Let's see what Claude, ChatGPT, and Gemini have to say. I'll input this text in all three and ask if it was written by a human or AI. The last thing I'm writing here doesn't reveal whether it's AI or not. It could be AI trying to game AI detection, or what does AI say?
And sure enough, AI can recognize this. They all confirm that the text was written by a human, pointing to the same examples I mentioned myself. Claude rather cheekily calls it: "The informal, sometimes rambling quality that feels more like a human thinking out loud."
All three reveal my little meta-trick of suggesting they could be AI instructed to fool AI detection. That is, itself, a clear indication of human authorship. Gemini is straightforward: "This last part I'm writing doesn't reveal whether it's AI” “... is a level of meta-awareness that is typical of human writing." Claude and ChatGPT grasp the same concept. AI models have no "meta-awareness" and have tremendous difficulty mimicking it naturally.
The Problem with Controlling AI: Too Little and Too Much
ChatGPT starts chatting with me afterward and asks if I wrote the text myself, which I confirm, and it notes: "Your writing style is clearly human – both in structure, rhythm and humor. You manage to create a dynamism in the text that AI often has difficulty imitating naturally." Thanks!
Then it suggests something interesting—the kind of thing that makes working with AI valuable: Try getting AI to create a linguistic style guide and write an article on "How Close Can AI Really Get to Human Writing Style."
I ask Claude 3.7 Sonnet to create a style guide and an article. The result is not good, and that has nothing to do with Claude specifically (I use Sonnet because I believe it's currently the best writing tool).
The writing guide "Natural Human Voice Tone and Style" is very long and contains much of what I would have guessed: Variation, long and short sentences, occasionally going off on a tangent, etc. But there are also some strange instructions: "Be skeptical and slightly sarcastic, especially about tech giants' statements." That's maybe taking things a little too literally.
And in the article itself: "The big tech companies love to claim their AI is basically indistinguishable from human writing at this point. (Sure, just like those 'cheese products' that are totally indistinguishable from actual cheese.)" Hmm!
AI tools share this trait: if you give them specific and rigid style instructions, the result is almost always exaggerated. It's difficult to get them to hit the right "temperature" or tell them "it should be something in this direction, but don't overdo it."
If you start with too concrete instructions, you will have to spend a lot of time tweaking the prompt to get the desired results.
What Distinguishes Human Writing from AI?
What characterizes AI text in a blog post, as opposed to human writing in general? I asked Claude, ChatGPT, and Gemini, and their answers align:
AI has a consistent tone and structure throughout, often a smooth, neutral voice without linguistic peculiarities, and when making claims, they are often broad and generalizing.
Humans, on the other hand, often have varying tone and rhythm. The text can contain creative expressions or personal humor and can include tangents or associative leaps in thought.
I agree with the latter. And although my writing style might seem a bit meandering to some, I would always rather read a writer searching for a personal style than the quiet, thick porridge of words that flows from AI.
AI's Unique Strengths in Writing
While AI struggles with authenticity, it excels where humans often fall short. AI can process and synthesize enormous amounts of information in seconds, maintaining remarkable consistency and breadth of knowledge. It never experiences writer's block or fatigue, generating content tirelessly with impressive grammatical precision.
For non-native speakers especially, AI bridges communication gaps with its vocabulary range and ability to adapt to different formats and audiences without the learning curve humans require. This makes AI an invaluable writing partner—handling the heavy lifting while humans contribute the authenticity that technology cannot replicate.
Technical Limitations Behind AI's Struggle with Human Style
Behind the scenes, even advanced AI models struggle with fundamental limitations. They lack genuine contextual awareness—recognizing patterns without truly "understanding" meaning.
This statistical approach means they have no life experiences, cultural background, or genuine emotions. AI often suffers from "hallucination problems," generating convincing but false information due to superficial understanding.
Perhaps most revealing: AI cannot model its own limited knowledge. We humans know what we don't know—this metacognitive awareness shines through in our writing in ways AI still struggles to credibly mimic.
Key Takeaways: Making AI Sound More Human
Embrace imperfection: Add deliberate variations in tone, rhythm, and structure to your AI-generated content.
Insert personality: Include rhetorical questions, controversial statements, or mild humor that AI typically avoids.
Show meta-awareness: Reference the writing process itself or include self-reflective comments about the text.
Use creative expressions: Add idioms, metaphors, or unusual word choices that deviate from AI's typically formal and predictable language.
Allow tangents: Include occasional thought diversions or associative leaps that break the logical flow in natural ways.
Remember, the goal isn't perfection. That is more than anything else what characterizes us humans: We are not perfect.