I'm going to go out on a limb and say there's a good chance you don't agree with the title. Many won't recognize the song at all, but it's a possible correct answer for a "reasoning" AI. "Reasoning" AIs can give worse answers than non-reasoning ones, so just be aware before you over-trust positive descriptors applied by corporate marketing departments.
What is "reasoning"? In this context "Reasoning" means breaking down a task into component stages based on similar past queries, and then executing those stages. That's how humans tackle many of the more complex tasks in life, so why should we not apply it here too? Because the AI does not think in the human sense of the word. AI chatbots use weighted distributions to pick from a selection of the most appropriate next token (word-chunk) for any given context. That means synonyms are similarly weighted. But words have implications as well as definitions. They have second meanings, homophones, metaphor, metonomy, and more. And words have alternate meanings on the way in (your prompt) to an AI model AND on the way out (AI response).
So what is the most popular song of all time? Well let's come at it like a reasoning AI. Let's break it down to a few possible definitions.
song = a musical composition / musical vocalization / rhythmic audio entertainment
popular = most sales / appreciated by the most people / has been played the most times
all time = within recorded history / since the charts began / from the big bang until now
Each of them overlaps with the others but has slightly different meanings, and yet the model has to determine which one or combination of them it applies. This is a rather extreme example, but even one of those alternate meanings could have a significant impact on a user or company. It's easy to imagine a record company using AI to write, and the AI uses a version of "popular" that doesn't map to the record company's busines model.
What musical vocalization has been played the most times since the big bang?
The Frog Chorus. Because it has been sung by millions of frogs for millions of years. This isn't wrong, but it's not what the human intended to ask when typing "What is the most popular song of all time?"
In contrast, if you asked a non-reasoning AI, it would give you an answer like "Blinding Lights" or "Happy Birthday to You". Even then there are different metrics that will give you completely different answers. This is not about AI, this is about how language works and the unstated assumptions humans make. Most language is imprecise and requires interpretation. As usual, the risk is in unexamined use of AI. It's being billed as doing something similar to what humans do, but when we underestimate what humans do and then offload underestimated model we introduce unexamined error types.