AI vs Human Writing: Can You Tell the Difference?

I got kind of obsessed with this question a few months ago. There's so much confident claiming in both directions, people saying AI writing is obviously detectable, other people saying it's completely indistinguishable, and I wasn't sure either crowd actually had good evidence. So I did my own informal thing.

I'm not a researcher. This wasn't a controlled study. But I collected a bunch of writing samples, some from AI, some from humans, and asked people in my life to guess which was which. And the pattern of what confused people and what didn't confused me too, in the best way.

What I actually did

I pulled twelve writing samples. Six from AI (I generated them myself using a few different models, with varying amounts of effort put into making them sound natural) and six from humans (a mix of blog posts, newsletter excerpts, and a couple pieces I wrote myself). I stripped out any identifying information and any obvious giveaways like author names or bylines.

I sent them to about twenty people, mix of ages, mix of how much they use AI in their own lives. I asked: which ones do you think are AI-written? Rank your confidence. And optionally: what made you think that?

The results were not what I expected.

What people actually got right

The samples that were correctly identified as AI most often were the ones where I'd put the least effort in. Asked the model to write something with minimal instruction, didn't do any editing, just published what came out. People flagged those pretty consistently. The reasons they gave were interesting: "too balanced", "doesn't commit to any opinion", "the structure feels formulaic", "there's no personality behind it".

That last one is the one I keep thinking about. No personality behind it. Because that's not really about grammar or vocabulary. It's about something harder to define, the sense that someone specific wrote this, from their specific experience and perspective.

The heavily edited AI samples and the human samples got mixed up constantly. People were wrong about those at basically coin-flip rates. Someone who was convinced one of my own blog excerpts was AI-written because "it's too clean" made me feel some kind of way.

What people thought they were detecting (and were often wrong about)

The stuff people mentioned as giveaways was really telling. A lot of people said things like "short punchy sentences feel human" and "long academic sentences feel AI". But I had human samples with long, dense paragraphs and AI samples that were deliberately punchy, and both categories fooled people who were using those signals.

Some people said they look for "hedging", phrases like "it's worth considering" or "this may vary" as AI tells. That did track for some samples. But then there were human writers who naturally hedge a lot and AI samples where I'd specifically asked the model not to hedge, and those threw people off.

Spelling and grammar mistakes? Several people said they look for those as human signals. I'd argue that's backwards in 2026, most humans have spell check and most AI makes subtle errors humans don't. But the assumption persists.

Okay but here's what I actually think is detectable

After running this experiment and then spending a lot of time thinking about why certain things tripped people's detectors and others didn't. I think the real signal isn't any specific word or structure. It's specificity of experience.

Human writing at its most human is full of weird specific stuff. A specific memory. A specific example that's almost too niche to be useful but that's exactly what the person thought of. An opinion that's kind of strong and a little idiosyncratic in a way that a balanced overview wouldn't include. An aside that's only there because it's funny to the writer.

AI writing, even when it's technically polished, tends to produce the general version of things. The example it gives is the most obvious example. The opinion it takes is the most defensible opinion. The structure it uses is the most logical structure. It's optimizing for reasonable and clear, which means it loses the texture of someone actually having lived through something and then written about it.

That's what people were detecting without being able to name it. Not the sentences. The experience.

What this means for AI detection tools

There are a bunch of tools that claim to detect AI writing. I've tested a few of them. Honestly? I'm not impressed. They tend to flag things as AI that aren't and miss things that are. They seem to be picking up on some of the surface signals I mentioned, sentence structure patterns, certain vocabulary frequencies, without getting at the actual thing that makes writing feel human.

And as AI writing gets more sophisticated and more people are actively trying to make it sound natural, those surface signals are going to become even less reliable. I don't think detection tools are going to "win" this arms race. The only reliable detection I've seen is exactly what the people in my experiment were doing unconsciously, looking for whether a specific human seems to be behind the words.

My actual opinion on all of this

I'm not going to pretend I have a clean take here, because honestly I find this kind of complicated.

I don't think AI writing is automatically bad or fake. I use it all the time. Sometimes I use it to draft something I then rewrite heavily in my own voice. Sometimes I use it for stuff where voice genuinely doesn't matter, documentation, templated emails, structured summaries. That seems fine to me.

What bothers me is AI writing that's pretending to be a specific human's personal experience and isn't. That's where the detection question actually matters, I think. Not "is this technically human-written" but "is this actually someone's real perspective and experience, or is it a plausible simulation of that?"

Because readers are extending a certain kind of trust when they read personal writing. They think they're getting someone's actual take on something. If that's a generated impression of a take, the reader is being misled in a real way.

Can you tell the difference? Sometimes yes, sometimes no. But I think you can often tell when something is missing from the writing, that specific, weird, particular thing that only comes from a person who was actually there and actually thought and felt something about it. That absence is the tell. And no prompt engineering has fully closed that gap yet.