Meaning Doesn't Transfer
The email was good. The job got done. So why the pause?
This morning I got an email from a colleague. His first language isn't English, and for years his emails have read like what they were: sharp thinking paying a toll at the border of a second language. This one paid no toll. It was fluent, structured, warm in the right places. And it had, unmistakably, been through a model.
The email was good. It did the job. The idea arrived intact, and for the first time it arrived without the tax. If you want the strongest case for generative AI in the workplace, this is it: a competent professional whose ideas were being discounted for their packaging, and the discount just got abolished. That's not a trade-off. That's a win.
And yet, it still gave me pause. It took me most of the day to work out why.
The signal that stopped costing anything
Part of it is the old economics of effort. Polished prose used to be expensive — it proved time spent, care taken, competence held. Michael Spence got a Nobel for formalising why that matters: signals work because they cost something. The costly version of a well-written email carried a second message underneath the first one: a mind spent effort on you. Make the polish free and it becomes universal. As a signal, it becomes meaningless. Fluency was evidence of a person at work; it is now evidence of an API call. The email is better and it certifies less, and those two facts have the same cause.
But I don't think that's the real loss, because I have to be honest about what the old signal was doing. His rough edges weren't signalling the truth about him. They were signalling less competence than he had. The costly signal was regressive — it taxed exactly the wrong people, and it taxed them on packaging rather than substance. The old regime was unjust. The new one is illegible. Anyone who tells you one of those is fine hasn't looked at both.
So the pause wasn't about trust, or authenticity, or whether I can still tell who wrote what. It's always been lossy. What's changed is that the distortion isn't his — and I can't learn a channel that isn't the same one twice. It's about something underneath all of that.
There is no conduit
The default picture of communication goes like this: I have a meaning, I put it into words, the words carry it to you, you take it out. Language as a pipe. Meaning as cargo.
The picture is wrong, and linguists have known it's wrong for nearly fifty years. Michael Reddy called it the conduit metaphor — the fossilised assumption, buried in phrases like getting your message across and putting it into words, that meaning travels. It doesn't. No meaning crosses the gap between two heads. What I send you is a signal, and the signal's only job is to make you build the meaning yourself, from your own materials, inside your own skull. Communication isn't transfer. It's remote construction — and the builder is always the reader.
A small demonstration, live from the drafting of this post. In a conversation about AI prose, the word polish went past me — as in shine, refinement, surface. I read a nationality and a flag. Same string, two meanings, and the meaning that fired was the one my head built, not the one the writer sent. We caught it in seconds, because we could ask. That's the whole model in miniature: the signal underdetermines the meaning, always, and the receiver generates the rest.
Once you see communication this way, you see what the skill of writing actually is. It was never fluency. Fluency is surface. The skill is the modelling underneath: what does she already know, what will she infer, where will this word send her? Every hard email is an exercise in simulating another mind and engineering a signal to steer its reconstruction. The struggle to convey meaning — the thing he fought through in every email for years — isn't friction in a pipe. The struggle is the skill being exercised. The artefact at the end is almost a by-product.
The rehearsal and the recital
Now look again at what happened this morning.
Maybe he wrote every word and had the model tidy the grammar, in which case the simulation happened, the rep was done, and the machine buffed only the surface. Maybe he gave it bullet points and approved the result. Maybe something in between. The ladder of delegation runs from tidy this down to write this down to handle this, and every rung emits the same artefact. That's the fact that keeps this from being a morality tale about him: the email doesn't say which rung it came from. It can't. And the rung is invisible in the other direction too, which is the quieter problem — the tool gives the sender no feedback about how much of the understanding he actually did. Use it long enough and the question stops having a felt answer.
So the erosion I'm nervous about isn't a sin anyone commits. It's that effort itself has become illegible: you can no longer tell how much of it was put in — to a message, by a colleague, into you. It might all still be there. You've lost the ability to read it. And it's a gradient nobody can see themselves sliding down.
When the model does take over the reconstruction-engineering — whichever rung, however often — it writes for a statistical reader and usually hits, because most professional meaning-building is generic. The email works. But each delegated rung is a skipped rep.
And the thing about modelling other minds is the thing about any capacity: it is maintained by use. Writing for a specific reader is theory-of-mind exercise — daily, effortful, invisible. Delegate the artefact and you quietly delegate the practice, and nobody notices, because the artefacts keep arriving and they keep working. The gym closed but the trophies still ship.
Reddy saw this coming in 1979, before any of the machines existed. His warning about the conduit metaphor was precisely that a culture which treats meaning-extraction as trivial — as mere unpacking, requiring no skill — will stop teaching and maintaining the human capacity for reconstruction. We've now built the technology that makes the conduit metaphor feel true. The words arrive pre-polished, the meaning appears to transfer, and the machinery in both heads that was doing the real work gets to stand down.
The man in the middle
Reading has always been reconstruction, and reconstruction runs against a model of the sender. His mind is the key I decode with. When I read his old emails — rushed here, careful there, funnier than usual — I wasn't extracting those facts from the words; I was inferring them, by simulating him. This morning, some unknowable fraction of the text wasn't encoded by his mind — and that's enough, because the decode doesn't fail cleanly, it fails silently. Is the warmth his, or a temperature setting? Is the care his, or the house style? I can't apportion it, which means none of it can be safely read as him. Security engineering has a name for this topology: the man-in-the-middle. Mallory doesn't break the encryption — she runs two channels, one with Alice and one with Bob, while each believes it is talking directly to the other. Everything decrypts cleanly at both ends, which is precisely why nothing looks wrong.
So I undersold the problem earlier. A free signal that certifies less would be thinning: less information about the sender. This is spoofing — and the counterfeit doesn't need to be total. If the warmth might be a temperature setting, the warmth stops being evidence, whether or not it was really his. The content arrived true, and the person arrived unverifiable — signed, in his name, by an unknown blend of him and something that has never met me.
There's a benign version of this middle, and it has an honourable name: an interpreter. Nobody calls the UN booth an attack, and the function is identical — a third party re-encoding meaning between two people so it can cross a border it couldn't cross alone. The difference was never the function. It's disclosure and allegiance. An interpreter is visible, declared, and professionally bound to fidelity to the speaker. This layer is undeclared, signs its output in the sender's name, and is bound to neither of us — it has a house style, an owner, and a set of instructions neither correspondent wrote. Route enough of the world's correspondence through the same few intermediaries and every message gets transcoded out of an idiolect and into a dialect. His emails now sound a little like yours, which sound a little like everyone's, because they were all re-encoded by the same middle. The man-in-the-middle position is also the normalisation position.
And watch which way the cryptography runs at the bottom of the ladder — because offices are already there, whether or not mine was this morning. A sender hands bullet points to a model that expands them into prose; at the receiving end, on a busy day, the prose goes to a model that compresses it back into bullet points. Encryption at one end, decryption at the other — natural language repurposed as a wire format, passed between codecs, still human-readable mainly out of legacy compatibility. Except this protocol inverts every property encryption exists to provide. In end-to-end encryption the endpoints hold the plaintext and the middle sees only noise. Here the endpoints see digests and the middle holds the only complete copy. We have engineered the one channel design in which the intermediary reads everything in full while the humans read summaries. And the codec is lossy in both directions with no checksum: the expansion invents — warmth, emphasis, specifics nobody chose — and the compression discards, and the decode is not the inverse of the encode. What survives the round trip is something plausible, not something said. The only authentication check left in the protocol is the oldest one there is: asking. Did you actually write this? A four-word question doing the work that checksums do everywhere else — for as long as it still gets a true answer.
The dial
I write in two registers, and the difference between them is the thing the machines don't have.
When I blog, I am precise. I want a thousand readers to converge on one meaning — mine — and divergence is failure. When I write songs, I want the opposite: a thousand listeners generating a thousand meanings, and convergence would be the failure. A song everyone decodes identically is a memo with a melody. Umberto Eco built an aesthetics on this split — the open work, designed as a field of possibilities the audience completes, against the closed work that channels its reader to a destination. But the working writer's version is simpler: it's a dial. You size the reconstruction space on purpose. Tight for the argument, wide for the song. Sometimes it changes per sentence.
The dial is where authorship actually lives. Not in the polish — in the intent about ambiguity: choosing which gaps stay open, and where those gaps send people.
And there's a private return on the open setting that I only understood recently. Because meaning doesn't transfer, the person who wrote the song isn't quite the person who plays it back months later. I re-listen to my own songs and generate new meanings from them, like a stranger with inside knowledge. The gap between signal and meaning — the same gap that makes communication hard, the same gap that misfiled polish under nationalities — is, on the open setting, a renewable resource. The bug and the feature are the same mechanism. What differs is who controls the aperture.
Generated prose has a default aperture problem. It isn't that models write badly — it's that training pulls them toward the same powerful middle: clear, competent, closed-ish, shaped for a generic reader. Ask for precision and you get something near it. Ask for openness and you get vagueness — which is not the same thing. The ambiguities in a song are load-bearing: chosen, placed, aimed. The ambiguities in generated text are residue — gaps left by indifference, not design. An open work and an underspecified one look identical on the page and are opposites in kind. One is an instrument. The other is fog.
The ones you don't write yourself
Delegation isn't evenly distributed. Nobody hands Tuesday works, see you then to a model. You hand it the complaint. The pushback. The disagreement with someone two levels above you. The apology you've been avoiding since Thursday. Those are the ones that cost hours, and they cost hours because of what they demand: sit in the discomfort, hold the other person's likely reaction in your head, and revise the sentence you'd enjoy sending into the one that will actually be heard.
I used to spend an entire evening on a single complaint. Not on the prose — on the simulation. Reading it back as them. Cutting the line that would have felt good and landed as an accusation. Now I can have a well-worded version in thirty seconds without any of the emotional freight, and that is a genuine relief. The dread was a real cost, paid heaviest by the people who care most about how they're received. I'm not nostalgic for it.
But look at which emails got selected. Logistics stay human. Character goes to the machine — because the difficult messages are exactly the ones carrying the most information about the sender. Where their limits are. What they'll defend and what they'll concede. How they behave when they're angry and still have to be professional. Whether they apologise, and how. You learn more about a colleague from one carefully-fought disagreement than from a year of sounds good.
So the reps aren't skipped at random. The hardest ones go first, and the hardest ones were the training. It runs in both directions: I don't practise modelling them, and I no longer receive the evidence that would let me model them. The correspondence continues at full volume. The person stops appearing in it.
That's the part I didn't see coming. I know my colleagues less well than I did, and nothing about our contact has decreased. How do you come to know someone through written correspondence when the writing was never theirs at the difficult moments — the only moments that ever told you much?
What propagates
So this is what I was nervous about, and it was never the email.
The email is good. The job got done. His ideas finally travel untaxed, and I won't pretend that's anything but a win. But the win ships with a quiet cancellation: the daily rehearsal — simulate the reader, design the signal, check the reconstruction — now happens in a data centre, inside a middle that reads everything in full and answers to neither of us, while the recital continues on schedule. And Akerlof's lemons logic was about used cars rather than prose, but it suggests how this propagates: when recipients can't distinguish costly quality from cheap imitation, they stop rewarding the difference, and the incentive to supply the costly version decays, sender by sender, until the careful version isn't just rare but pointless.
And the sender's side is only half of what propagates. The same tool that writes the email will happily read one — what do they mean, how do they feel, what should I say back — so the interpretive work goes the way of the compositional work, and the exchange hollows at both ends: two systems performing reciprocal empathy on behalf of absent humans. A relationship is a jointly trained model — two people fine-tuning on each other through years of misreadings, checks, revisions, remembered history. That friction was never around the relationship; it was the mechanism by which the relationship became specific. Delegate both ends and the mutual model stops updating: two people becoming strangers at ever-higher levels of apparent fluency. And the specificity doesn't vanish — it relocates. The middle reads his full text and mine, models him to write and me to interpret, and builds a working representation of both of us inside the exchange. The intermediary is now the only party modelling both sides in full. The intimacy the endpoints stop building accumulates in the middle.
The capacity won't vanish loudly. Capacities maintained by use never do. The artefacts will keep arriving, indistinguishable from craft — that's exactly why nobody will notice the dial gathering dust until someone, somewhere, actually needs to turn it: needs to close a meaning down to precision, or open one up on purpose, and reaches for a skill that was always practised in the writing we no longer do.
Meaning doesn't transfer. It never did. Which means the oldest work between minds was never the words — it was the deliberate engineering of what happens in someone else's head. We just automated the engineering and kept the words.
The words were the cheap part.