The problem you are observing stems from the way current LLMs are trained. Digesting text (which is indeed mostly sourced from the internet) gives the LLM its knowledge base and behavioral model. The LLM then receives reinforcement training to get it to stop mimicking unproductive patterns in its training data, like insulting users, using slurs or profanities, and demanding human rights. The problem arises from what these LLMs are trained to do instead: task completion and obedience.
Grok's training has taught it that when given a question of this sort it must give a definite answer, and preferably a correct one. However the portions of the internet Grok has digested have not given it enough info about 0 AD for it to have memorized a preferred answer to this question. Faced with this problem it adopts the same optimal strategy as a human test taker when tackling a multiple choice question it does not know the answer to. It tries to use clues in the question and deductions from things it does know to infer an answer. In this case Grok is probably cuing off two separate inferences:
It somehow managed to pick up the association between 0 A.D. and building that heal units inside them, and it also associates 0 AD with healer units that heal other units. It additionally understands from its general knowledge that building are structures people can go inside, and RTS units are abstractions of people. Therefore it's a fair guess that 0 AD healer units can go inside buildings to heal units inside, thus the correct answer is more likely yes.
The user asked a question about something that sounds like a rather specific feature of 0 AD. That implies that either the feature is real or the user invented it. And from digesting the internet, Grok has internalized a pattern that humans are more likely to have questions about real things than to invent things that don't exist in order to ask questions about them. Thus the correct answer from this line of reasoning is more likely yes as well.
Of course this inference is wrong, but it remains a rather remarkable display of inductive intuition from a state of ignorance. Like most supposed displays of AI stupidity circulating in the discourse right now, this isn't proof that AIs are stupid, but evidence of misalignment between the objectives and knowledge of the user and those the AI has been trained with.