Paul Graham has stopped reading his mail. Not all of it — just the part that arrives in what he calls a hard-hitting journalistic style, the polished register no founder under pressure has ever written in. He says he can tell these were written by AI, and that once he knows, he can't finish them. The word he reaches for is not boring. It is lied to.
That word is the whole story, but there is more to it.
The easy interpretation is that AI writing is simply bad, and Graham is reacting to bad writing. It is a comfortable position, because it flatters us humans: machines produce slop, discerning humans detect slop, taste wins. However, there are interesting edge cases worth considering in a thought experiment.
Take for instance two emails. The first was written by a founder who spent an hour with a model prompting, cutting, rewriting until nothing in it reads as machine-made, until it sounds like a person who had something to say. The second was banged out by a human in ninety seconds, ugly, with a typo in the second line. Graham finishes the first and never suspects it. He finishes the second too. Now consider a founder who spent that same hour and still shipped something that reads like the model's default style. That one he closes on sight.
So the variable is not effort, and it is not quality. The ugly human email is lower-quality than the polished machine one, and it passes. The thing being punished is recognisability: you might call it detectable defaultness. What gets rejected is not that a machine was involved but that nothing was chosen: the sender reached for the model, shipped whatever fell out, and signed it. The em dash, the tricolon, the confident throat-clearing, the hard-hitting journalistic style — these are not bad writing. They are the absence of a decision, and the absence of a decision is legible to anyone who has seen the default before.
For decades, the surest sign that a document had been written by someone who did not care how it looked was that it was set in Times New Roman. The font was not ugly. It was a superb piece of design, drawn for The Times of London in the 1930s and lovely at the width of a newspaper column. It became a punchline for exactly one reason: it was the default in Microsoft Word, and so it came to mean I opened the program and started typing and changed nothing. The beauty was never the point. The not-choosing was the point.
Then Microsoft moved the default. Calibri replaced Times New Roman in 2007, arriving as the fresh, rounded, modern alternative to all that austerity — and within a few years people were calling Calibri overused, the new sign that no one had bothered. In 2024 the default moved again, to Aptos, and the cycle reset for a third time. The tell was never any particular font. The tell was the default, whatever it happened to be that decade. Each new default arrived as the escape from the last one, then hardened, on schedule, into the next thing to escape.
Right now the tells are the em dash and the journalistic cadence and, a year or two ago, the word delve. They feel diagnostic. They are not durable. The models will be retrained, the registers smoothed, the obvious markers sanded off. A new set of defaults will settle in and begin to read, to the next Paul Graham, as the unmistakable sound of someone who did not choose. Avoiding this season's tell is not a strategy. It is a way of losing more slowly.
Which returns us to lied to. A signature is a claim. It says a person stood behind these words, and that the person spent some part of themselves on you in particular. The default betrays the claim. It reveals that the thing presented as costly was free, that the note which arrived addressed to you could have arrived, identical, addressed to anyone, because no one chose anything in it. That gap between the signature's promise and the prose's emptiness is the lie. It is not a lie about who typed the words. It is a lie about whether anyone was home.
This is also why Graham, who has spent years funding the companies that build these models, is not the hypocrite his tweets make him out to be. In the same thread, when someone pointed to the tension, he drew the line cleanly: use it, but use it correctly. The line between using it correctly and incorrectly is the line between an output you would stand behind and a default you would not. He is not against the machine. He is against the empty signature. Those were always different objections.
So, if the offence is detectable defaultness, then the winning move is not to write everything by hand. It is to make the machine's output undetectable. You have to put enough of yourself on it that no tell survives, which is to say, to choose. The future Graham is describing is not an inbox free of AI. It is an inbox where the AI is invisible, because the people will have learned to sand off the defaults. The slop that remains will be the slop that didn't try, and everyone congratulating themselves today for catching the em dash is grading on a curve that is about to get much steeper.
The question stops being was this written by a machine. It becomes did anyone choose anything here — which, long before any of this, was the only question that ever told you whether something was worth reading.
Unlock the Future of Business with AI
Dive into our immersive workshops and equip your team with the tools and knowledge to lead in the AI era.