large language models
anthropic: "we'll start invisibly watermarking all text we generate"
ok, but _how_. given how much stuff is just plaintext, the only thing i can think of is a whole bunch of zero width characters... which a lot of things will just immediately strip
large language models
@Dex a better way to do this would have been to base64-encode the binary data and use tag characters. tag characters are underspecified in unicode and do not have strong conformance requirements (they are only used for a small number of emoji flags currently), but like variation selectors they are default-ignorable and do not impact most plaintext operations. however, Unicode’s (correct) stance would remain that this is largely inappropriate and plain text is not a good medium for text with these kinds of metadata requirements
