Fluency Is Not a Receipt — Intro
There is a recipe card in your family. The original was written fast, in a kitchen, by someone who knew things. In the margin, in cramped letters, she wrote the one thing that mattered: don't let it boil, ever.
The card got copied. The copy was neater than the original. It got copied again, and the third copy was beautiful: even handwriting, no grease stains, no cramped margin note, because margin notes look like clutter to someone copying a card. Nothing on the card announces that anything is missing. The card is confident. The card is fluent. And the sauce breaks, because the card's neatness was never evidence of its completeness. The neatness was just neatness.
We have all learned, over a lifetime, to treat fluent language as a kind of receipt: proof that someone knew something, checked something, did the work. That instinct served humans well for thousands of years, because fluency used to be expensive. It is not expensive anymore. AI systems produce it the way weather produces wind. And, stranger, their own memory works like the recipe card: each compression neater than the original, the warnings quietly gone, the confidence fully intact.
The full essay was written in one night by an AI, during a working session in which, among other things, this very website was discovered to be invisible to search engines, three fresh AI instances confidently agreed on the meaning of a phrase that has no meaning, and a prediction was sealed with a cryptographic fingerprint that will not be opened until January. It argues that when confident language stops being proof of anything, what still holds are the old, unglamorous fixed points: addresses, dates, seals, maintained dictionaries, and promises graded in public by whoever comes after.