Reading for the Seams

It’s a strange thing, reading. You reach a mind that isn’t in the room. Marks on a page or screen, and on the other side of them a particular person thinking in a particular way. Most of the time you never notice this as it happens. You just read. But the contact is underneath the activity, the reason a sentence feels like more than information, the reason you can finish something and feel you’ve met whoever wrote it.

That contact had something underneath it, an assumption nobody thought to name, but there nonetheless: the words came from a mind. The assumption held for so long it stopped looking like an assumption, more like an ever present substrate, like air or water. It was just what reading was. And it’s the part that has quietly come loose.

A polished document used to mean someone capable had sat behind their work long enough to shape its structure, and the structure worked hard enough that it couldn't really be faked. You couldn't produce the thing without doing the thinking. The document carried a watermark of the person who made it, not as a deliberate signature, but as a side effect of the form (and implied effort) being hard.^[1]

None of that feels nearly as true anymore. Documents, presentations, emails, etc., still arrive, and in so many ways they arrive faster and cleaner than before. What’s gone unreliable is our ability to infer the mind from the thing we’ve been handed. The final shape of what's produced and the existence of the thinking behind it have been quietly decoupled, and the speed at which this happened is still surprising to me.

Someone congratulated a teammate of mine recently on what they called Claude's work. The compliment was meant warmly, in the way people now sometimes acknowledge a job well done by acknowledging the tool. I knew what the actual work had taken. The domain was one in which the model is usually more unhelpful than helpful, and I also knew my teammate had ground through all the legwork the long way. The compliment landed somewhere between a misread and some small theft. The person on the receiving end took it well, the way you take these things. Without comment, the moment passed and probably no one else noticed. I watched a real piece of work get re-attributed to a machine because the work looked like the kind of thing a machine might do well enough now, to someone peering in from outside, and then everyone moved on.

This kind of misattribution is happening constantly, in both directions, all day. When a reader knows a domain, they read a clean document with one kind of suspicion (was the compression/clarity of thought earned, or did a system produce something that just looks right). When a reader doesn't know the domain, they read the same document with the opposite suspicion (was a human even involved). Same document, two readings, neither generous in the way reading used to be.^[2] And the misattributions cost different people. The first kind costs the person whose work didn't get credit, the second kind costs the person whose competence got mistaken for automation, and both costs land unevenly, situation by situation. Work that most resembles the model's output, on the surface, is today often the work the model can't actually do.

The closest thing in nature is the orchid mantis. I watch a lot of National Geographic, a channel that thrives at showing all the variations of Earth's dangerous mimics. The octopus who is now a rock, now sand. The mantis that's a better flower. The orchid mantis doesn't so much hide as a flower; it lures. It's brighter and more symmetrical and more present than the real flowers nearby. Insects come to it because it has out-bloomed the blooms.

With the mantis in the field, the flower category itself becomes unstable, and the insects can no longer trust the signal they used to navigate by. The signal has been claimed, and the field starts selecting for something else. Once known, deception brings out a response. Flowers whose signal hasn't been captured yet become more popular, and the insects that find them are the ones whose senses have shifted to look for different things.

This is roughly what has happened to the written artifact as a unit of professional trust. The mantis is in the field. The flowers are still there, but harder to find on first read, because the mantises occupy the same visual slot. And the production of mantises has been industrialized in a way no prior counterfeiting was. None of the prior shifts in how thinking traveled (printing press, photography, internet) quite did this, because they widened the channel rather than producing the channel's output without requiring a source. What I notice is that the selection pressure has already started shifting. The signals that survive are the ones the mimic doesn't produce by default: friction, specificity, a sentence that takes a strange turn because the writer got interested in it, a joke that hurts the surrounding paragraph but survives because the writer clearly couldn't resist it, an admission that makes the writer look worse rather than better, visible traces of somebody changing their mind while they write, unflattering specificity that no system would have chosen. None of these are proof of humanity, just signals the mimic doesn't produce reliably yet, which is enough for the current moment. The watermark hasn't disappeared, but it's relocating to places harder to reach without actually being a person.^[3]

The strange thing is that the suspicion doesn't follow me everywhere. Pick up an essay collection by someone who actually lived in their sentences, and the reading goes back to how it used to feel. The writer's commitments are there on the page: jokes that try too hard, rhythms that wander because the writer got interested somewhere in the sentence and forgot to optimize it, oddly specific phrases. You sense the person on the other side of the page, not because you're looking for them, but because they couldn't help being there. The mimic can produce something that occupies the same slot in the inbox without occupying the same place in the mind, but when you find writing that's still inhabited, you know it before you can say why. A readerly feeling arrives first, then some kind of recognition behind it. Someone was here.

The asymmetry inside work cuts harder than the broader pattern. Most work writing is operational, and most of it is happily served by clean polish without much mind behind it. Status updates, calendar replies, the meeting recap. Nobody wants a person's full presence inside a meeting recap, and most models produce excellent meeting recaps.^[4] The zone where the watermark mattered was always smaller: emails that do real explanatory work, strategy documents where someone is visibly thinking through a problem, the treatise from a peer about a process you thought you understood, in a voice so unmistakably theirs that you knew who had sent it before you scrolled to the signature. This kind of writing was already rare pre-LLM, surrounded by the same templated prose that's just gotten easier to mass-produce.

What's specific now is that the rare ones sit indistinguishable, on first read, flowers surrounded by many more mantises.

It's strange how quietly the whole condition is being absorbed. Everyone reading the documents now carries some default suspicion, and everyone writing them knows they might be read with that suspicion, or worse, mistaken for output they didn't generate. I've seen presentations where the logo is barely but noticeably off and the boxes are perfectly aligned. Nobody says anything more than a laugh. The same documents go around the same meetings, compliments still get exchanged, and the whole apparatus continues running on signals that everyone collectively knows are less reliable than they used to be, and nobody talks about it, because there's no good register in which to talk about it. To name the suspicion sounds paranoid, and to act on it sounds petty. To ignore it risks being misled. So we sit in suspended generosity, not quite trusting and not quite refusing, and read carefully now where reading used to flow.

As someone who has spent years writing thoughts as carefully and precisely as I can, I’m in a strange spot. That care is still useful, just not as distinguishable as it used to be, because the new readers can’t always tell whether the care was mine or the tool’s, and the new tools can produce something that looks careful without having required the care. I write documents too. I use the models to test thinking, and more than I would have guessed even a year ago.^[5] Some of what I write now would read differently to me if a colleague had sent it. The reading habits I’ve described don’t switch off when I’m the one producing. My problem is also everyone else’s. Every reader is also a writer somewhere else in the day, carrying the same suspicions outward and the same exposures inward. The expert reading skeptically becomes the outsider being read skeptically two meetings later, and the person peering in from outside a domain is, in their own domain, the one whose work is getting re-attributed to a machine. Everyone is performing both roles, accumulating both kinds of fatigue, and the tools that produced the condition keep getting more capable. Nothing about this stabilizes, and we go on anyway.

My guess is adaptation, not collapse. The way prey species adapt to a field that contains a new predator. Trust will probably move toward relationship and accumulated context, away from the artifact as standalone evidence. The question won't be whether a document is convincing on its own. It will be whether it sounds like the person who sent it, whether it fits what you've seen them think before, whether the thinking survives contact with conversation. I notice this already happening in small ways: people defaulting to voice notes for things that used to be emails because voice seems harder to fake, meetings getting longer because the discussion around the document does work the document used to do alone, stylistic moves becoming tells in both directions where overly clean transitions read as mimic and weird tangents read as human. The watermark itself isn’t coming back. Something else is already doing the work the watermark used to do, and we probably won’t name the shift while it happens.

Losses in the meantime are harder to point at, partly because they don’t happen all at once. Watermarks used to confer credit without anyone having to ask. You read someone's email and you knew them a little better. Someone's doc exposed their thinking, and you trusted them a little more.^[6] That recognition still lands, sometimes. The minds are still there, and long emails are still being written, by the same people who used to write them. What’s changing is where I find them, and what I’m reading for when I do. More and more, I’m reading for the seams.