TL;DR: Tools for removing AI-ness can only reach the “pattern layer”: what can be generalized, enumerated, automated. Human quality lives in the non-pattern layer, where tools don’t reach and AI can’t follow. AI can learn your style; it cannot learn your essence. The closer writing sits to the statistical average, the more readable it becomes, and the less it sounds like anyone.

“What withers in the age of mechanical reproduction is the aura.” — Walter Benjamin

The stronger AI gets, the easier it is to produce smooth prose. But prose that is too smooth often sounds like no one in particular.

The reason isn’t hard to see. A stronger model picks the most likely word at each step, always choosing the safest option, until the whole piece reads without a single surprise. But what makes writing feel human usually lives in the unlikely choices: a pause, a detail given more weight than it deserves, a sentence only you would phrase that way. It isn’t in the layer that can be generalized. It’s in the specifics where rules run out and models fall short.

The end of “correctly good” is writing that sounds like no one

This is something like what happens with AI-generated faces. The proportions are right, nothing is technically wrong, yet you sense immediately that this isn’t a real person. Because what it renders is the average of all faces.

What makes a real face that person’s face is precisely where it departs from the norm: the slightly crooked smile, the asymmetric brow, an expression that only appears from one angle. Writing works the same way. An AI sentence is correct everywhere and belongs to no one. When everything is uniformly right, a certain plastic quality sets in.

This isn’t just a metaphor. Psychological research has found that faces people judge as attractive tend to be the ones closest to the population average; the nearer a face sits to the statistical mean, the more appealing it reads. The principle behind AI sentence generation is the same: pick the statistically most likely word at each step. The average face and the average sentence share a common mechanism: selecting what is most typical. Attractive, readable: both are byproducts of being too typical. And the end of typical is writing that sounds like no one at all.

AI can learn your style. It cannot learn your essence.

Two things are worth separating here.

Part of your writing can be generalized: the vocabulary you favor, the sentence structures you reach for, the rhythm that recurs. This part AI can learn. When it imitates someone’s “style,” this is the layer it works on. Feed it enough of your writing and it can produce something that resembles you.

But another part resists generalization: why you pause here, why you choose this metaphor, why you let this sentence stay a little rough. The choices behind those decisions come from your judgment, your experience, your sense of proportion in that moment. That is your essence. AI cannot learn it.

So, plainly: AI can learn your style, it cannot learn your essence. The line in Part 1, “the hardest AI-ness to remove is the part that already sounds like me,” is about style converging with the average. The layer that doesn’t converge is the essence.

The wider de-AIing tools spread, the more writing converges

Stripping the AI quality out cleanly, then, is not enough on its own.

When everyone uses the same set of de-AIing tools, all running on the same underlying model, the “clean” they produce is the same clean. There’s an irony here: the more widely de-AIing tools spread, the more the resulting writing resembles itself. It’s like cosmetic surgery calibrated to the same golden ratio: run enough people through it and you get a row of faces that look like each other.

Clean is the baseline. Whether writing is still recognizably yours after cleaning depends on what you put back in, not on the tool. The tool’s job is to remove what doesn’t belong to you. What does belong to you, you have to grow yourself.

The writers detectors wrong most often are the most careful ones

This also explains why AI detectors are unreliable.

They look for pattern-layer features: word choice that is too predictable, sentence length that is too uniform. The trouble is that carefully written, thoroughly revised human prose tends to be clean, orderly, and predictable in exactly those ways. So the people detectors most readily flag are often the most conscientious writers, not AI.

A 2023 Stanford study tested several leading detectors: native English writing was classified correctly almost without exception, but TOEFL essays by non-native writers were misidentified as AI-generated at a rate of around sixty percent. The reason is that non-native writing tends toward conservative vocabulary and simpler structure, which lands squarely in the “too predictable” feature zone. More pointedly, when those essays were revised to be more elaborate and closer to the vocabulary AI tends to favor, the false-positive rate dropped sharply. What the detectors were punishing, it turns out, was an authentic voice that hadn’t yet been smoothed away.

Non-native writers aren’t the only ones affected. Another study had 22 reviewers read one AI-written and one human-written piece blind. Only 12 identified the AI correctly, and 4 misidentified the human piece as AI. Even trained reviewers can’t always tell.

A detector gives you a probability score, not evidence. Delegating the judgment to a tool that routinely wrongs careful writers means trading your own perception for that tool’s blind spots.

What about you?

At this point you might want to ask: don’t you use AI heavily in your own writing?

Yes. But I’ve kept the parts that are mine to do. AI helps me remove rough edges, expand ideas, and run first drafts. What carries weight in an argument, how the structure fits together, where to place a piece of lived experience: those judgments I haven’t handed over. Writing as a craft is the practice of exactly those decisions.

The pieces on this site are produced that way: AI can generate a draft, but I’ve revised in ten or more rounds, making each choice myself.

The strongest evidence that writing is human isn’t that it looks like it was written by a person. It’s whether it can hold up under questioning: how many versions it went through, why it changed, which sentence stayed and which was cut, and whether the author can explain any of it.

Whatever can be learned was never the essence

One last question: what if AI eventually learns the “judgment” and “refinement” you’re describing?

Research is moving in that direction. There is evidence that AI’s tendency to converge toward the average can be partially resolved through specific prompting strategies, producing more varied output. The homogenization at the pattern layer is, technically, addressable.

My answer remains: whatever AI can learn was never the essence.

The essence is always “the part that hasn’t been generalized yet.” Once a technique gets formalized into a method and absorbed by AI, it becomes the new average. Then something new, still ungeneralized, takes its place as the human quality. Technology can keep pressing toward the edge of the pattern layer, but the part that can’t be caught always runs ahead. That line doesn’t stop, because writers don’t stop moving forward.

In the AI era, why write at all?

Let me close with something larger than whether your writing sounds like you.

Many people worry that AI will replace writing. But that anxiety overestimates our own minds. The human brain isn’t particularly suited to deep thinking; it’s suited to perception, movement, instinctive response. Deep thinking is costly, and most of the time we’re happy to route around it.

The risk in an AI-saturated environment is that something is very willing to spare you that effort. Platform algorithms learn your cognitive patterns and feed you content that is half-true and precisely calibrated to move you emotionally. The process is frictionless. You don’t notice it happening. Your judgment gets quietly collected.

In that context, writing changes its value. It isn’t only about producing a piece. It’s about exercising your cognitive muscle.

Working through something yourself and then putting it down in writing is the cheapest way to maintain cognitive complexity. It keeps you from gradually surrendering your own judgment to the feed.

Writing in the AI era isn’t for publication. It’s for staying awake.