From the Alpaca report (https://crfm.stanford.edu/2023/03/13/alpaca.html):

> we watermark all the model outputs using the method described in Kirchenbauer et al. 2023, so that others can detect (with some probability) whether an output comes from Alpaca 7B

https://arxiv.org/abs/2301.10226

> We propose a watermarking framework for proprietary language models. The watermark can be embedded with negligible impact on text quality, and can be detected using an efficient open-source algorithm without access to the language model API or parameters. The watermark works by selecting a randomized set of "green" tokens before a word is generated, and then softly promoting use of green tokens during sampling. We propose a statistical test for detecting the watermark with interpretable p-values

There is no reason this couldn't be watermarked with the individual user, though the if it detects them probabilisticly then it could be difficult to make the case to take action on inferences made.

It’s already been done at Tesla with spaces.

The basic technique is much older than that, dating to the Cold War era and likely even before.

If they start doing that, leakers will do the same - Take their leaked version, run it through ChatGPT, get something that's similar but not the same for the leak.

Post-truth is the era where nothing is attributable to a human because of AI interventions, and liability is shielded at every level because of it. It's insane.

You don't need an LLM to do this. You could use small sets of semantically equivalent words or whitespace. Things like this: https://twitter.com/elonmusk/status/1579101966453858305