1. You assume that your LLM of choice is perfect and impartial on every given topic, ever.
2. You assume that your prompt doesn't interfere with said impartiality. What you have written may seem neutral at first glance, but from my perspective, a wording like yours would probably prime the model to try to pick apart absolutely anything, finding flaws that aren't really there (or make massive stretches) because you already presuppose that whatever you give it was written with intent to lie and misrepresent. The wording heavily implies that what you gave it already definitely uses "persuasion tactics", "emotional language" or that it downplays/overstates something - you just need it to find all that. So it will try to return anything that supports that implication.
It doesn't matter if you make assumptions or not - your prompt does. I think the point of failure isn't even necessarily the LLM, but your writing - because you leave the model no leeway or a way to report back on something truly neutral or impartial. Instead, you're asking it to dig up any proof of wrongdoing no matter what, basically saying that lies surely exist in whatever you post, and you just need help uncovering all the deception. When told to do this, it would read absolutely anything you give it in the most hostile way possible, stringing together any coherent-sounding arguments that would reinforce the viewpoint that your prompt implies.
1. You assume that your LLM of choice is perfect and impartial on every given topic, ever.
2. You assume that your prompt doesn't interfere with said impartiality. What you have written may seem neutral at first glance, but from my perspective, a wording like yours would probably prime the model to try to pick apart absolutely anything, finding flaws that aren't really there (or make massive stretches) because you already presuppose that whatever you give it was written with intent to lie and misrepresent. The wording heavily implies that what you gave it already definitely uses "persuasion tactics", "emotional language" or that it downplays/overstates something - you just need it to find all that. So it will try to return anything that supports that implication.