Old maps (and perhaps new ones) used to add fake little alleys so a publisher could quickly spot publishers infringing on their IP rather than going out and actually mapping. I wonder if something similar is possible with LLMs.
i have come across this one for example
https://github.com/sentient-agi/OML-1.0-Fingerprinting> Welcome to OML 1.0: Fingerprinting. This repository houses the tooling for generating and embedding secret fingerprints into LLMs through fine-tuning to enable identification of LLM ownership and protection against unauthorized use.
Would be interesting to see if this kind of watermarking survives the frankenstein types of editing they are presumably doing. Per the linked account, they took a model, changed tokenizers, and added layers on top. They then presumably did some form of continued pre-training, and then post-training. It would have to be some very resistant watermarking to survive that. It's not as simple as making the model reply with "my tokens are my passport, verify me" when you ask them the weather in NonExistingCity... Interesting nonetheless.