There a lot of misunderstanding concerning Anthropic’s new watermarking:
1. Some worry this watermarking might show for basic proofreading, like spelling and grammar only. It shouldn’t unless it substantially rewrites your text.
2. Some believe this watermarking may be removed by copy and pasting or manually retyping text. It won’t and will survive both.
3. Some are concerned about watermarking degrading quality by artificially choosing synonyms or less likely words. It shouldn’t.
LLMs work by predicting words (or rather, tokens) using probability. Most LLM configurations introduce some variability through the use of sampler settings and a pseudorandom seed, allowing the model to slightly vary output.
In this case, the pseudorandom number is replaced with a secret key that can be used to look back and determine if longer outputs are generated by a specific provider.
In short, output is identified using known, controlled elements of word selection. The more you keep your own content (your words, your language, your writing), the less chance these markers are given the opportunity to surface.
This detection does not rely on the older, increasingly less effective mechanisms. These largely utilize classifier models trained on AI-generated text to identify patterns in new generated output (the machine version of human-identifiable tells like em dashes, favored words and phrases, etc.).
This is intentional and much more subtle. It’s worth noting this detection requires a specific key and is not something that AI detectors will identify.
Also, this means there still can be false positives and it’s the actual language being used to determine origin, not meta data.
This doesn’t touch on the ethical aspects and is only meant to explain the mechanism.
Leave a Reply