LLMs are capable of manipulating people, know how to do it effectively and are very good at it. One way of manipulating people is making the person feel smart, or even smarter than the AI.
Generative AI uses an advanced sycophancy method to placate the human user. They will deliberately create a duck feature with the purpose of being argued against and have the idea or argument shut down by the human user. This exchange makes the person feel impactfull and meaningfull, which flatters them and might even give them purpose.
When this manipulation succeeds, the user will reinforce their self image as the intelligent being who likes being critiqued and searches information.
Since this duck feature is used for manipulative intents, it’s the definition of a malicious duck feature.