I wrote two lines. My client read two lines. ChatGPT charged us both for fifteen.

Let me be specific about what happened, because the specificity is the whole point.

My two lines said: the timeline works for us; let's proceed with the revised scope. That is the entire information content of what I needed to communicate. Everything else was framing; context; professional distance; the linguistic equivalent of a pressed shirt. What I wanted the client to feel was not "I have more information than I did before reading this." What I wanted them to feel was: "Tabrez takes this seriously. Tabrez is thorough. Tabrez is the kind of person who sends fifteen-line emails."

ChatGPT gave me exactly that feeling, at zero additional cost in time. I felt professional. I clicked send. I moved on.

What I did not know was that the client had not read it at all.

This is not a hypothetical. I am telling you this because I suspect it is happening thousands of times per day; possibly more. Someone writes something true and short. They feel it is insufficient. They ask a model to make it more. They send it. The receiver, who has too many emails and too little time, asks a model to make it less. They read the less. They respond. The response is also less than it should be. They expand it before sending. And so on.

I mentioned this to a friend who runs a small IT services firm in Pune. He said he had been doing the exact same thing for almost eight months. His clients were in Germany. He said the Germans summarised even faster than the Indian clients. "They just click Zusammenfassen and reply in one line," he told me. I am not too sure what Zusammenfassen means but I thought I shall look it up later. I did not.

What we have built is an extraordinarily efficient system for converting meaning into volume and then converting volume back into meaning; with significant compute costs at each end and no additional value produced at any point in the cycle.

Here is the underlying thing that I keep returning to, and I was a big struck by it when I first realised it. Words used to be expensive. Not in money; in time. Writing fifteen lines took longer than writing two. That time was a signal. When someone sent you a long, careful email, the length itself was communicating something: I thought about this. I invested time in this. I consider this worth elaborating.

That signal no longer exists. The length of a message no longer tells you anything about the effort behind it. A two-line email and a fifteen-line email might have required exactly the same amount of thinking; it is simply a question of which instruction was given to which model. We have, without quite deciding to, destroyed one of the oldest legibility signals in professional communication. We replaced effort with the appearance of effort; and we did it so efficiently that nobody noticed.

And now... we are using the same tools to strip the appearance of effort back off again. Which means we are left with exactly what we started with; except we have paid for the round trip.

I am not here to write a sad article about the death of long-form email or the attention economy or whatever the current vocabulary is for this anxiety. What I am interested in is the economic structure of what is happening; because I think it is genuinely strange and nobody is quite saying it directly.

Attention has not expanded. Time has not expanded. The cost of producing words has collapsed. So we are producing more words; to signal the same amount of effort that fewer words used to signal; and the receiver is spending more time compressing them back down to the actual information content; and everyone is paying token costs on both ends of this transaction. The only entity unambiguously winning is the one selling the tokens.

There is a version of this that was suppose to go the other direction. The argument was: AI reduces friction, which means the quality of communication goes up. You spend the time you used to spend on phrasing on the substance instead. You think more clearly; you write more efficiently; the receiver gets something better in the same space.

What actually happened, at least in my case, was: I spent zero additional time on the substance; I used the friction reduction to produce fifteen lines instead of two; my client used their own friction reduction to get back to two; and I have genuinely no idea whether the substance was better or worse. I only know the length was different.

We took two lines, expanded them to fifteen, and then compressed them back to two. And paid compute costs at both ends of that round trip. I am not sure what the word for that is, but it is definitely not efficiency.