(untitled)

Slop watermarks are coming to Claude and all LLMs, thanks EU!

They use a seeded pseudo-random number generator to ensure the next word is still random, but can be matched back to the seed. It’s clever, and hilarious how often Anthropic use “random” to explain how LLMs work.

The current crop of “AI detectors” are absolute bogus. When Anthropic releases their watermark detection API a strong positive should be fairly trustworthy for long prose. In theory the same API could help to lower a watermark score with trial and error. Use a second LLM to replace potential watermark candidates with synonyms; frankenslop.

I really don’t like vibecusation. That is just as dehumanising as everything else the AI industry has polluted our culture with. AI disclosure should be mandatory, watermarks and auto-detection warnings are the second best option (if they work).