I’ve started noticing something when I read.
Sometimes it’s obvious, like a phrase I’ve seen too many times, or a perfectly balanced sentence. Other times I couldn’t tell you exactly what gave it away. The writing is good. There are no glaring mistakes. The ideas may even be interesting.
But something about it makes me think: AI wrote this.
I’m also aware that I could be completely wrong.
That’s part of what makes this moment so strange. We’re becoming suspicious of writing that feels too polished, too structured or too familiar, while AI is getting better at producing the quirks and imperfections we associate with human writing.
I’ve noticed something else too. It isn’t only that some AI-generated writing doesn’t quite sound human. It’s that so much of it sounds like the same human.
Human writers have always had fingerprints. We recognize them in ways that go well beyond vocabulary or grammar. There is rhythm and peculiarity. What someone notices. Where they linger. The connections they make. What they leave unresolved. Sometimes you can recognize a writer you know before you ever see the name.
In my last article, Made by a Human,[i] I explored what happens when we can no longer assume the person whose name is on something actually created it, and how that uncertainty changes our relationship with the work.
There is another layer to this. We’re already seeing what happens to our individual voices when millions of us run our words through the same handful of machines. And sometimes the change is most noticeable when we already know what someone’s fingerprint looks like. A colleague recently reminded me of a concept that gave me another way to think about what we’re experiencing: the uncanny valley.
The term dates back to 1970, when Japanese roboticist Masahiro Mori described something peculiar about our response to machines that resemble us. As a robot becomes more humanlike, we tend to respond more positively. But when it gets very close to human without quite getting there, our response can suddenly shift. Something feels off.[ii]
For decades, we’ve mostly talked about the uncanny valley in relation to robots, computer-generated faces and, more recently, synthetic voices. But what if there is an uncanny valley in language too? What if some of what we’re reacting to when we say this sounds like AI is that same sense of almost, but not quite?
I’m not sure that explains all of it. Sometimes the writing doesn’t feel strange at all. It simply doesn’t sound like the person whose name is on it.
The question I keep circling: is the person whose name is on the work still in it?
I’ve Heard This Writer Before
Once you start noticing the patterns, they become difficult to unsee. It’s gone way beyond the em-dash. There are certain constructions I encounter again and again, such as the perfectly balanced contrast, or the sentence that tells us something isn’t this, it’s that. Then there are the series of three examples and the short sentence dropped between longer paragraphs for emphasis.
None of these are bad writing. I would argue that some of them are very good writing. AI learned these patterns from us, making certain legitimate human habits look like evidence of AI.
That’s part of my own conundrum. I recognize some of my writing style in the very patterns I now associate with AI. So when those patterns begin appearing everywhere, what am I supposed to do? Change the way I’ve always written just to prove that I wrote it? I don’t have an answer to that yet.
There’s a cost to this that I don’t think we talk about enough. If polished patterns start to read as evidence of AI, then people who have always written that way become suspects. They can’t produce an alibi for their own writing process. In Part 8 of The AI Reckoning Series, I wrote about the way an AI-detection flag works like spectral evidence: the accused has no way to disprove it.[iii] Trust erodes in both directions. We stop believing the work, and the people who wrote it stop expecting to be believed.
That doesn’t mean AI has a single voice. It was trained on an extraordinary range of human language, and given enough direction it can change its vocabulary, cadence and style considerably.
That’s not necessarily how we use it in everyday life. We ask for help with an email, a post, an article or a response, and something polished appears almost instantly. Maybe we change a few words or rewrite a paragraph. Sometimes the language is good enough that we simply use it.
So much of it sounds like the same human.
I’ve noticed the effect most when it happens to someone whose writing I’ve been reading for a while. I already know how they communicate. I know how they tell a story, how they make an argument, the words they tend to use and the ones they don’t. In some cases, I’ve read enough of their work that I have some sense of how they think. Then suddenly, the voice changes. The writing may be more polished. The vocabulary may be more sophisticated. The ideas may still belong to the person. But I’m reading it and thinking, That’s not you.
I would never say that to the person. And I can’t know exactly how much AI was involved. Our instincts about these things aren’t proof. What surprised me was how strongly I reacted to the mismatch. What bothers me is the distance between the person I’ve come to know and the person I suddenly encounter on the page.
It has made me think differently about voice. We tend to describe a writer’s voice as though it were primarily a matter of style. I’m not sure it is. Over time, the way someone communicates becomes part of how we recognize them. Their language carries traces of what they notice, what they know, what they question and how they make sense of things. When that changes, the person can become harder to find in the writing.
The more I thought about the uncanny valley, the more I wondered whether this might be a version of the same phenomenon. There is some evidence that it is. In one study involving more than 1,000 people, participants tried to distinguish between human and AI-generated posts in social media conversations. They weren’t particularly good at identifying which was which. But the researchers still found evidence of an uncanny-valley effect in people’s responses to AI-generated dialogue.[iv]
I find the gap between those two things fascinating. People could have a reaction without being particularly good at identifying what they were reacting to.
With a humanoid robot, we can sometimes point to what feels wrong. The eyes don’t move naturally. An expression lingers too long. Something looks very close to human without quite getting there.
Language is harder. There may be nothing obviously wrong with a piece of AI-generated writing. It can be articulate, coherent, warm and funny. Sometimes the very things that make it good are the things that make me suspicious. It’s a sentence balanced in a way I’ve encountered dozens of times before, or an idea reframed with a familiar construction. There are few detours and not much left unresolved. The writing moves efficiently from question to answer, perhaps more efficiently than most of us actually think.
None of that proves that AI wrote it. A human can write that way, and AI can be prompted not to. So the clues keep moving. As AI gets better at human language, the boundary becomes harder to locate. At the same time, we’re spending more time reading, editing and sometimes adopting language produced by AI. The influence may be moving in both directions.
Machines are learning how we sound. I’m curious about how much they may eventually change how we sound. And that isn’t necessarily all bad. Could AI actually teach some of us to be better writers?
The original uncanny valley has another side. Once the artificial representation becomes convincing enough, the discomfort is expected to diminish. There is the possibility that eventually we won’t notice anything at all.
I wonder what happens when we get there.
I Have Seen This Before
In some ways, this isn’t a new problem.
A couple of decades ago, long before generative AI, I was working for an executive during a post-merger integration. She was responsible for developing a plan for our department, but she didn’t really know how to develop it or put it together. I ended up writing the plan for her, and it went forward with her name on it. I remember watching her present it, using words I had chosen, and feeling relief that it was good, alongside a quiet ache at being invisible in my own work.
I never told anyone that I had written it. I didn’t need to. Her boss and other members of the senior team came directly to me when they wanted to discuss the work or needed help taking it further. They knew.
They knew her. They knew me. They had seen enough of both of our work to recognize that the thinking and writing in that report didn’t match the name on the front of it.
People could have a reaction without being particularly good at identifying what they were reacting to.
I’ve thought about that experience again recently as I’ve watched people I know begin using AI more heavily in their writing. The feeling is remarkably similar. No one had a detector back then. No one had been tipped off. They simply knew the work didn’t sound like her. And when they wanted to understand the work, they came to the person who had actually done the thinking.
When the document stops telling people who you are, they look for another signal. They ask a follow-up question. They put you in a room and watch how you think. They wait to see whether you can take the idea somewhere it hasn’t been. The premium moves to whatever can’t be handed off, whether that is a live conversation, an unscripted moment, or a questions nobody prepared you for.
It has also made me think differently about what our writing communicates. Words aren’t only a way to transmit information. They tell other people something about us.
When we read someone’s strategy, we’re getting some sense of how that person thinks. An author’s essay gives us more than sentences on a page. The same is true when a founder writes to an investor, a job candidate answers a question or a colleague sends us a thoughtful response. The communication becomes part of what we know about the person.
Of course, people have always had help. Executives have communications teams. Authors have editors. Political leaders have speechwriters. Colleagues collaborate. My own story proves that none of this began with AI. What AI changes is how easy it has become to separate the person from the work carrying their name.
That report took me time to write. I had to understand the business, know what needed to happen and develop the plan. Today, someone could hand much of that task to AI in seconds. That doesn’t mean the result would necessarily be bad. It might be excellent. But there is a point somewhere along that continuum where the work may stop telling us very much about the person whose name is on it.
In Made by a Human, I wrote that when I buy a book, I don’t believe I’m purchasing only the words. I’m purchasing access to someone’s thinking. I’m realizing that applies to far more than books. We use communication as one of the ways we come to know one another. Over time, the way someone communicates gives us a sense of how they think and who they are.
That’s why I have found myself worrying about the reputation of people whose voices suddenly seem to change. The person I know and the person on the page no longer quite match.
If the work carrying our name becomes less reliable as a signal of the person behind it, I think we have a bigger issue than whether someone can spot AI writing.
When Better Starts to Look More Alike
What I’ve been noticing anecdotally is beginning to show up in research. A recent study examining more than 880,000 texts across several settings looked at what happens when large language models are used to polish and rewrite human writing. The meaning people wanted to communicate was largely preserved. What changed was the variation in how they expressed it. The writing became more alike.
Researchers found that AI assistance reduced differences in writing complexity and shifted some of the linguistic characteristics that can carry information about the person behind the words.[v]
Other research has found a similar tension in creative writing. In one study, people given access to generative AI produced short stories that were judged more creative, better written and more enjoyable, particularly among people who initially scored lower on creativity.
The AI-assisted stories were also more similar to one another.[vi] I think that’s where this gets interesting. The individual writer can benefit while, collectively, our writing becomes less varied. It also means sameness isn’t necessarily an argument against using AI. It gives us a reason to pay more attention to how we’re using it.
What AI changes is how easy it has become to separate the person from the work carrying their name.
If the first response is good enough, do we take it? If AI gives us a more polished way to say something, do we stop to ask whether it’s actually the way we would say it? When it removes an awkward turn of phrase, sometimes that’s exactly what we wanted.
Other times, it may have removed something that belonged to us.
Are We Learning to Perform Authenticity?
For years, technology gave us increasingly sophisticated ways to improve how we presented ourselves. We smoothed skin, corrected lighting, whitened teeth, removed wrinkles and erased whatever we decided didn’t belong in the picture. Eventually, we got so good at creating perfection that perfection itself became recognizable. We could see the filter.
Then being unfiltered became a signal. People began posting without makeup, showing the messy kitchen, leaving in the awkward moment, sharing the photograph that wasn’t quite perfect. Some of that was genuinely more authentic. Some of it became another aesthetic. The candid photograph could be carefully staged. “No makeup” could require quite a lot of makeup. Once we learned which signals communicated authenticity, we also learned how to reproduce them.
AI introduces a new version of the same problem. If polished writing starts to sound artificial, we can tell AI to make it less polished. We can ask it to vary sentence length, use contractions, leave a thought slightly unresolved, add humor or avoid the phrases people have begun to associate with AI. We can give it samples of our own writing too.
That’s where this gets harder. The characteristics we’re using to decide that something feels human can also become instructions for a machine. A typo doesn’t prove a person wrote something. An unusual sentence doesn’t either.
If enough of us start associating certain quirks with human writing, AI can learn the quirks. Which makes me question whether the rough edges were ever a very good measure of authenticity.
I’ll admit a version of this myself. I’ve stopped using em dashes, because people have learned to read them as a tell. I changed my writing to avoid looking like a machine, which is a small performance of its own.
Can Personalization Fix It?
Personalization looks like the fix for sameness. Given enough examples of my writing and enough feedback from me, AI should become increasingly capable of reproducing the characteristics of my voice. If generic AI smooths away some of our differences, personalized AI could preserve them.
It isn’t that simple. A person’s voice isn’t just sentence length, vocabulary, cadence or a collection of stylistic preferences. Those are the things we can observe on the page. Underneath them is everything that produced them.
My voice has been shaped by what I’ve experienced and studied, but also by the things I’ve changed my mind about. I notice things someone else might pass over and make connections because of experiences that belong specifically to me. There are also contradictions I haven’t completely worked out yet. All of that finds its way into how I write because it shapes how I think.
AI may eventually become very good at reproducing the patterns those experiences leave behind. A reader may eventually have no way of knowing whether I wrote something myself, used AI throughout the editing process, or handed most of the writing over after teaching it how I sound.
That takes me back to the executive and her report. If I had been able to write that report perfectly in her voice, perhaps no one would have noticed. But it still wouldn’t have reflected her thinking.
So, if AI eventually sounds exactly like me, have we solved the authenticity problem, or have we simply made it harder to see?
How Knowing May Change What We Experience
Research suggests that what we believe about the origin of something can affect how we experience it. In one experiment, people evaluated artworks that had actually been created by AI. Some were labeled “Human-created” and others “AI-created.” The works believed to be human were evaluated more positively on measures including liking, beauty, profundity and worth.[vii]
Poetry complicates the picture. In a 2024 study, non-expert readers were worse than chance at distinguishing AI-generated poems from poems written by well-known human poets. They frequently mistook AI poetry for human poetry and, on some measures, preferred it.[viii]
So our ability to recognize human creation isn’t necessarily very good, even when knowing or believing something is human can affect how we value it.
A person’s voice isn’t just sentence length, vocabulary, cadence or a collection of stylistic preferences.
I think about this in ordinary interactions too. A thoughtful comment beneath something I’ve written feels meaningful in part because I assume someone read what I wrote, thought about it and responded. If I later discover that the person barely read the piece and generated the response with AI, the words haven’t changed. My experience of the exchange probably has.
Of course, I don’t care equally about this in every interaction. If AI gives me accurate instructions for resetting my router, I’m probably not concerned with who wrote them. A personal letter from someone I love is different.
Why does knowing who is behind the words matter so much in one situation and so little in another?
The Authenticity Premium
When I first started thinking about an authenticity premium, I imagined it largely as a response to abundance. If AI can produce endless quantities of polished writing, images, music and video, things we know were created by humans might become more valuable simply because they are scarcer.
I’ve changed my mind. Human origin alone isn’t enough. Humans create plenty of things that aren’t original, thoughtful or particularly good. And work created with AI can contain enormous amounts of human experience, judgment and originality.
What I find myself caring about is whether there is some correspondence between the person and what they are presenting to me.
If I’m reading your ideas, I want them to be your ideas. If I’m listening to your experience, I want it to be your experience. If your name is on an argument, I want to believe you understand it well enough to stand behind it.
AI can be part of that process. It can challenge an argument, help with research, point out something I’ve missed or help me find a clearer way to express an idea. Sometimes it helps me get closer to what I was actually trying to say.
The harder question is how much distance we can create between ourselves and the work before something important changes. That may be where the authenticity premium actually comes from. I don’t need to know that AI never touched the work. I want to know that the person whose name is on it is still there, and that their thinking shaped what I am reading.
So here is where I’ve landed: the authenticity premium belongs to work its author can stand behind, however it was made.
What I Don’t Want to Lose
There is some irony in writing this article while using AI. I build AI. I use it regularly. And throughout the process of writing, I find myself doing something I probably did less consciously before. I listen for myself.
AI may suggest a sentence that is perfectly grammatical and sometimes quite good, and I’ll reject it. Sometimes I know it’s because I’ve seen the construction too many times, or it has a rhythm I’ve started associating with AI. Other times I can’t tell you exactly what’s wrong with it. I just know I wouldn’t say it that way. That doesn’t make my version better. Someone else might prefer the other sentence.
But I notice the half second before I decide. AI offers something smoother than what I had, and part of me wants to take it without checking whether it’s true to what I meant. Usually the checking is what makes the writing mine. I wonder how many people skip that step, and what happens to their confidence in their own voice when they do.
If all I wanted was polished language, AI could increasingly provide that for me. I want something different from the relationship. I want it to help me research more broadly, challenge what I think I know and notice what I may have missed.
And I still want to recognize myself when we’re done.
[i] https://www.thefutureoftrust.net/p/made-by-a-human-the-question-behind?r=3nbvtz
[ii] The Uncanny Valley: The Original Essay by Masahiro Mori - IEEE Spectrum
[iii] https://open.substack.com/pub/thefutureoftrust/p/when-seeing-is-no-longer-believing?r=3nbvtz&utm_campaign=post-expanded-share&utm_medium=web
[iv] [2409.06653] Human Perception of LLM-generated Text Content in Social Media Environments
[v] The shrinking landscape of linguistic diversity in the age of large language models | Nature Human Behaviour
[vi] Generative AI enhances individual creativity but reduces the collective diversity of novel content - PubMed
[vii] Humans versus AI: whether and why we prefer human-created compared to AI-created artwork - PubMed
[viii] AI-generated poetry is indistinguishable from human-written poetry and is rated more favorably | Scientific Reports