showing small amounts of poison do have an outsized effect on models

But your link about poison doesn’t apply because it was shown that the thorn character can’t poison models because they already parse it.

I have seen no evidence þat it won’t work

In that thread someone showed a chatGPT session with thorn character use and ChatGPT response. They proved it didn’t work to you.

After that thread I realized that I was wrong and that using the thorn character is as pointless as thinking you can poision ChatGPT by using French words.