Introduction

A year and a half ago, we published "Does it really matter if a blog is written by AI or a human?" We were curious, a bit skeptical, and just trying to figure out what all the fuss was about.

So we reread it this week to see what held up.

Some things we got right, some things changed completely, and one thing became crystal clear: the battle with AI writing isn't about teaching it to write better, but about knowing what to delete. Back then, we were asking whether people could tell the difference between AI and human writing. Now? The difference is pretty obvious. AI writing all sounds the same.

I believe it is, because it got everywhere. And when everyone's using the same tools, everything starts to blend together. 

Why AI saves time then costs it?

Yes. You'll probably spend 50% of your time deleting what it gives you. And that time is not in editing, improving … but deleting. 

AI generates fast, which is the whole pitch, right? But then you sit there reading through what it gave you, and you start cutting. First you remove the buzzwords, and all those words nobody says that much. Then you're fixing those weird em dashes it loves for some reason, and stripping out the three-part lists it can't seem to stop making. You're deleting sentences that sound like they came from a corporate memo nobody wanted to read in the first place.

And somewhere around the twentieth deletion, you realize something funny: this is taking longer than just writing it yourself would have taken.

The irony is kind of brutal when you think about it. AI saves time generating the initial draft, but then costs you more time fixing everything. You trade writing time for editing time, and somehow the editing feels harder because you're trying to make someone else's voice sound like yours instead of just using your own voice from the start.

I've done this hundreds of times now, and the pattern is always the same. Write a detailed prompt, get back 1200 words, delete about 500 of them, rewrite new ones, keep them. And every time I think: why didn't I just start with a blank page?

The answer, obviously, is that we all want the shortcut. AI promises one, and we want to believe it works. But the shortcut has a hidden cost, and that cost gets paid in delete-key presses and rewriting time.

Why does everything sound the same

So why does AI writing sound so similar, no matter which tool you use? Turns out there's actual research on this, and the answer is kind of fascinating in a slightly depressing way.

Researchers at University College Cork used something called literary stylometry, basically computational methods for analysing writing style, to compare AI writing to human writing. What they found was that AI models produce what they call "compact, predictable styles" while humans show "greater stylistic range shaped by personal voice." Even when ChatGPT is specifically trying to sound human, it still carries a detectable fingerprint, as the researchers call it.

GPT-4 is actually more consistent than GPT-3.5, which means the better the model gets technically, the more uniform it becomes. It's like the models are getting really, really good at being average. And there's a reason for that. AI works by predicting the most statistically likely word that should come next in a sentence. It looks at millions of examples of writing and figures out what word usually comes next in this kind of situation. The result is safe, familiar, and completely average. It's optimised for the most common pattern, not the most interesting one.

That's the whole problem, really. AI converges toward the average of everything it's seen, which would be okay if it were only trained on human writing. But, by April 2025, over 74% of newly created webpages contained AI-generated text. AI is now being trained on content written by other AI. Researchers actually have a term for this, they call it "model collapse." It's a feedback loop in which each generation of AI gets a bit worse because it's learning from corrupted data rather than from actual human creativity and experience.

Meanwhile, 86% of marketers say they edit AI content before publishing it, which means most of what you're reading online was AI-generated and then human-fixed. But it still sounds AI-generated because most people don't really know what to fix. They leave in the buzzwords, the em dashes, the three-part structures, the "in today's world" openings, all the exact phrases that make you immediately think "oh, AI wrote this."

Can AI ever write like a human?

Probably not. And the reasons why are pretty interesting.

The UCC study showed that human authors naturally display "far greater stylistic range, shaped by personal voice, creative intent, and individual experience." AI writing is polished and clean, sure, but it's also really uniform. It doesn't vary the way human writing does because it has no real reason to vary, it's just optimizing for the most likely output, not trying to be interesting or different or weird in the ways that human writing can be.

Think about what AI fundamentally can't do. It can't take a stance that costs it something. It can't have a bad day and write differently because of that bad mood. It can't change its mind about something it believed yesterday. It can't contradict itself in ways that come from genuine uncertainty. It can't sound tired, or genuinely angry, or bored, or excited in ways that aren't just performed for effect.

And here's the big one: AI has no stake in anything. It doesn't care if it's wrong about something because there's no consequence to being wrong. It has nothing to lose, no reputation on the line, no relationships that could be damaged by saying the wrong thing.

This is what people miss when they say AI writing is getting better. The advantage humans still have isn't that we're better at grammar or sentence structure, AI is often better at those technical things. The advantage is clear thinking. It's knowing what matters and what doesn't. It's being able to read something that sounds perfectly fine and still think "no, this is bullshit" and delete it, even though it's grammatically correct and flows well.

AI can polish sentences all day long, but it can't know which sentences shouldn't exist in the first place. That requires judgment, and judgment requires having something at stake.

What works

Okay, so if AI writing is this problematic, what's worth trying?
Forget most of those "humanizer prompts" people sell.
They don't work the way everyone thinks they do.

But here's what may help atm: 

The Claude Humanizer Skill

There's a skill for Claude that's based on Wikipedia's "Signs of AI Writing" guide, and it actually does something useful. It strips out the most common AI patterns, things like inflated symbolism, promotional language, those vague attributions AI loves, the em dashes, the rule of three structure, AI vocabulary words, and all those negative parallelisms that make text sound robotic.

Does it work perfectly? No. It catches patterns and helps you see what needs fixing, but you still have to do the rewriting yourself. 

Importing your own voice

If you're using ChatGPT, you can export your memory from there and import it into Claude. Suddenly Claude has a much better sense of how you write, your sentence rhythm, your vocabulary, the kinds of phrases you use naturally.

But even better than that? Feed Claude actual examples of your writing. Not prompts that describe how you write, but real things you've written. Let it learn from what you've done, not from what you tell it your style is supposed to be. The difference is pretty noticeable.

The editing approach

I saw a prompt on LinkedIn a while back that completely changed how I think about editing AI content. It's pretty harsh, but it works better than anything else I've tried.

The basic idea is to be ruthless about certain things:
Here's that section rewritten with bullet points:

The basic idea is to be ruthless about a few specific things:

• Delete every sentence about importance, impact, legacy, or broader trends. AI loves writing these, but they almost never mean anything concrete. If you can't point to a specific example of the impact, cut it.
• Kill the AI vocabulary. Words like pivotal, crucial, enhance, underscore, landscape, vibrant, testament, foster, delve, showcase, intricate, and evolving. These words show up in almost every AI-generated piece. Once you notice them, you can't unsee them.
• Get rid of contrast structures. You know the ones: "It's not X, it's Y" or "Not A, not B, but C." They sound smart and clever, but they're usually just filling space without saying anything real.
• Cut out vague authority. If it says "experts say" or "studies show," either name the specific expert or study, or delete it. Vague authority is just a way to make something sound credible without actually being credible.
• Remove those neat summary endings,  "In conclusion," "Overall," "In summary", they're all just filler. If your piece actually makes a point, you don't need to announce that you're wrapping up. Just wrap up.
• The big test: if a sentence works in any article, delete it. Copy a sentence and imagine pasting it into a blog about gardening, then software, then marketing. If it works in all three, it's saying nothing specific enough to matter.

If a sentence could apply to a thousand different topics, delete it. That's the real test. If you can copy-paste a sentence into an article about gardening, an article about software development, and an article about marketing, and it works in all three, it's saying nothing specific enough to keep.

This isn't about making AI sound human, but removing everything that sounds like AI trying to sound important. Most AI writing fails because it's trying too hard to impress you instead of just trying to tell you something useful.

Starting with your own mess

The best way I've found to use AI isn't to let it write from scratch. It's to write your own messy first draft, all your ideas, your perspective, your voice, even if it's disorganised and rough… and then let AI help clean it up, but also keep what makes it yours.

You do the thinking and get your ideas down. Then AI helps with the technical stuff: fixing grammar, tightening sentences, catching repetition you didn't notice, suggesting better structure. But the substance stays yours. The voice stays yours.

Because here's what I've learned, after working as a content creator for 8 years…  AI is actually pretty good at structure and mechanics. It's terrible at substance and originality. If you give it real substance to work with, it can help clarify and clean up that substance. If you ask it to generate substance from nothing, you just get generic content that sounds like everything else. And we don’t love that :) 

Why clear thinking still wins and always will

Simple. Because AI can't originate ideas. It can only remix and recombine things it's already seen. It's like a really sophisticated pattern-matching machine, not a thinking machine.

You can teach AI all the patterns in the world. You can train it on millions of examples. But you can't teach it judgment, because judgment requires having stakes in the outcome. And judgment is what separates content people remember and care about from content they scroll past without thinking.

When I edit AI writing now, I'm not really fixing sentences or improving word choice. I'm asking different questions: Does this mean something, or is it just words? Would I say this to someone if we were talking? Does this have a real point, or is it just taking up space because it sounds like the kind of thing you're supposed to say?

Most AI content is just filling space. It looks like writing (proper sentences, good grammar, logical structure), but it's not really saying anything you couldn't have guessed. It's optimised to sound correct without committing to anything specific or taking any real position.

Real writing commits to something. It takes a position. It says "I think this is true" or "this approach is wrong" and then defends that position. It has a point of view.

AI hedges constantly. It qualifies everything. It presents both sides of every question without ever really landing anywhere. And there's a reason for that: it has no skin in the game. It's not trying to convince you because it has no stake in whether you're convinced or not. It's just trying to sound convincing, which is a completely different thing when you think about it.

Conclusion

When we wrote that original blog year and a half ago, we thought the important question was: can you tell if AI wrote it? Can you spot the difference between human and AI writing?

Turns out that was kind of the wrong question to focus on.

The real question should have been: does it matter who wrote it if there's actual thinking behind it? And the answer to that is yes, it matters a lot.

Not because human writing is automatically better than AI writing in some inherent way. But because writing without real thinking is just pattern-matching, and we're already drowning in pattern-matching. The internet is essentially training itself at this point. Models are learning from other models instead of from human creativity and experience. And that feedback loop means everything is slowly converging toward this same safe, polished, ultimately meaningless middle ground where nothing sounds wrong, but nothing sounds particularly right either.

The only way to break out of that pattern is to think. To have a real perspective on what you're writing about. To be willing to delete the 50% that sounds like everything else and keep only the the other 50% that sounds like you, or at least sounds like something specific instead of something generic.

We were right about one thing in that original blog though: it really doesn't matter if something is written by AI or by a human, technically speaking.

What matters is whether there's actual thinking behind it. 

And AI doesn't think. It predicts what comes next. It generates plausible sequences of words. It fills space efficiently.

But you think. You decide what matters. You figure out what to cut and what to keep.

That is the work now. Not writing faster or generating more content. It's knowing what to delete, what to change, and what's worth keeping. It's being the editor, not just the writer.

Interested in reading more insights on AI? Check our previous blogs! 

‍