In proofread feature, instruction given to the AI appears in the output

leo97g

The issue I mentioned in the title is appears below the output. I've only used it in English, and it might be very unique to certain situations that I don't fully understand the reason. I don't know if the same problem exists in other languages. Here is the sample text where I received instructions in the output:
'Each time the user will provide you with an English word or phrase they want to memorize. Your only task is to respond in a way that helps the user permanently memorize the meaning of the text they entered. The user speaks Turkish, so your thought process and language should be in Turkish.'
As you can see from the text, this is actually an AI instruction, which might be the reason, but there's no such instruction would cause this error in the text either.

Simply just seeing corrected text in the output

kirkmc

If you choose proofreading, why would you expect it to behave otherwise?

leo97g

kirkmc seems like you misunderstood you can see what the problem is from the screenshot below

kirkmc

It's proofreading. Why do you think that within its proofreading routine it is also doing LLM stuff?

leo97g

kirkmc From the output above. As you can see, there is an instruction to the ai. So it is probably performing proofreading through the AI.

kirkmc

The proofreading process isn’t querying the AI. It’s using the translation API which is different.

fxgn

kirkmc Kagi Translate is a wrapper around an LLM pipeline. Proofread used LLMs.

kirkmc

fxgn It uses LLMs, but it's not interpreting anything in the text as a query to an LLM. It's using a different API.

fxgn

kirkmc yes, but LLMs are always prone to prompt injection. You can see that on the screenshot @leo97g provided. Since his text looks like an LLM prompt, the LLM assumes that it has to also put its own prompt in there. They didn't type the bottom paragraph (the one circled in the screenshot). The LLM just added that itself to their provided text.