2026-04-10 Re: Parlor Trick
Some comments about my being called out by an LLM and other random links about LLMs
I wonder if the LLM he fed his reponse through read my post (because Christian did link to it) because of the use of the phrase “parlor trick.”
I did manually feed Sean's post to ChatGPT so that it could have the full context, and it did pick up parlor trick from that. I wanted to see what my AI prompt injections would do had the AI been given the context of both Sean's post and the LLMs reply to it. The instructions I did didn't always fool the LLM, because apparently some LLMs have been trained on recognizing instructions in data, or they use certain tricks to try to distinguish data from instructions, etc. But there are different tricks to try to get around these, according to what I've read online. I hate that we have to do prompt injections now just to protect ourselves from AI, but hopefully we can find a way to do them so that it's not so disruptive to the reading experience.
One thing I noticed is the LLM also picked up my two usages of "..." and then added those into its response everywhere. This is another way of showing how statistical models actually compound stylistic things, making them more common in outputs, and that's why AI detectors rely on common AI phrases for detecting whether something was written by AI. It's also why em dashes are a thing, because their frequency is being compounded from all the apparently professionally published articles and books the AI is trained on.
I do wonder if it's possible that we could add stylistic things like this to prompt injections that could more easily reveal responses as an AI, using things like semicolons, elipses, em dashes, etc. Anyways, just a thought.
P.S. I made slight updates to my previous post to make it more technically accurate for any pedants. Not very big changes, and my core arguments remain the same, but they're more clarified.