HOW I LEARNED TO STOP WORRYING AND LOVE PANGRAM

since i’ve been getting some pushback here: it’s really weird that so many of the most anti-ai people have also decided to believe that pangram is fake, all detecting software is unreliable and there’s simply no way of telling for sure whether something is ai or not. i don’t know why you’d want to think that; if you’re trying to preserve human writing then a working ai detector is a pretty useful thing to have. maybe it’s just the usual left melancholy that refuses to believe in any kind of mitigation. maybe it’s because pangram itself uses ai and they don’t want to admit that it can be useful in any kind of situation ever. but for what it’s worth, it’s simply not true. as i keep pointing out, ai writing is actually incredibly unusual: it’s not bland, it’s not just an average of the training data, it follows very different rules to human writing and if anything the gap between the two is growing. if you think that ai writing is necessarily worse than human writing—which it is—you should accept that ai writing is distinctive, not in some wishwashy sense of lacking “soul” or whatever but distinctive in a way that can be measured statistically. this is exactly the kind of thing neural nets are extremely good at! let it digest enormous corpora of human and ai-generated writing and it will very quickly learn the differences and gain the ability to sort any new text into one or the other category. i’ve seen a lot of people claim that it will give false positives because ai tells like “it’s not x, it’s y” are actually just chatgpt repeating formations that are already in its training data. in fact, that’s incredibly unlikely to happen, precisely because those formations are in the training data. that’s taken into account. the detector doesn’t work by applying rules, it works by noticing billions of statistical correlations and tangling them together in a black box of probabilistic parameters. it just knows, in the same way that when i’m reading an obviously machine-generated text like the one below i just know. later i can point out specific features, but in the moment the process of recognition is opaque to me. when i put it into pangram and it came back 100% ai that was just the machine echoing my own judgement. but you really can trust it. ai is already very good at distinguishing human writers: if i give any chatbot a piece of my own unpublished writing it will instantly finger me as the author. i have a pretty distinctive style but this seems to work on much less voicey people too, and anyway the rote and stereotyped nature of llm prose means that it’s even more clockable than i am. you can accept this without ceding any ground to the machines. if i ask claude to write something in the style of sam kriss, it consistently spits out some cold dogshit i would never write in a million years. if i feed that cold dogshit back into claude and ask who wrote it, it doesn’t say me. last time i tried, it said marina hyde or nick cohen. it can easily identify will self, but when i fed it the “will self” article in the telegraph it said it was aris roussinos. it works. we should be incredibly grateful that it works, and it’s possible to use ai against itself, because that means there’s still some kind of future for the individual literary voice. pangram’s cluster charts are one of the most hopeful things i’ve seen in recent weeks. anyway, when i say that this text was written by ai and published under will self’s name i say that with absolute certainty. it’s possible that he’s doing some kind of bit here, and in a few months he’ll come out with a big essay on how robo-will self suckered all the morons in the reading public; that would be nice. it’s possible he’s been incapacitated and some unscrupulous family member is churning out slop under his name; that would still be better than the remaining alternative. but it is totally possible to objectively identify slop, and this is slop.