posted in Technology
Anthropic just rolled out a tool that'll decimate some people's dreams of writing AI novels undetected
tech.yahoo.com/ai/claude/articles/anthropic-just-rolled-tool-thatll-040904467.html
Maybe I’m out of the loop, but how would I watermark a text without sounding obvious? By using weird phrases? They get edited. By using e.g. an exact combination of starting letters over a large paragraph? One changed word and it’s broken. And even if not, I could happen to write the same myself, and then?
How am I hiding a signature in Plain text? Anyone got a better idea than my silly ones?
By being just a little weird, but in a pattern, and not to you. A pattern like: every n tokens raise the temperature (randomness) for one token. Then to dectect AI, you tokenize and calculate how expected each next token is. Then you try fitting the pattern to that.
Wouldn’t that potentially affect output quality particularly in code/tool calls?
To my surprise almost if not all LLM are set not to be deterministic and have one unique input result in always the same output.
They all are set to have a “temperature” setting so that they sound more natural.
Personally I think quality of the output tokens of LLM is surprisingly not as much their priority as the quality and truthfulness of the result.
I would 100% prefer a LLM that is purely deterministic and repeats the same answer exactly to the same question. Instead LLM are constantly choosing the next likely token more TL appear human rather than being accurate.
These LLM are designed as sycophants and set up and trained as such.
So a fingerprinting in the output seems quite realistic. An LLM is not giving you it’s best most likely answer. It’s taking one of the most likely answer and adds a sprinkle of uncertainty and randomness on top of it so it looks natural…
I would 100% prefer a LLM that is purely deterministic and repeats the same answer exactly to the same question
Then you don’t like your LLM to solve problems, because a lot of how they solve problems is trying, testing, failing, then trying again - getting a different answer and testing that…
During inference ?