Recommender systems, the algorithms suggesting everything from movies to jobs, have a popularity problem. They often push already-popular items, creating a 'rich-get-richer' scenario that leaves hidden gems undiscovered. But what if Large Language Models (LLMs), the brains behind AI chatbots, could shake things up? A recent study dives into this, examining how LLMs perform as recommender systems. Researchers built a simple LLM recommender called WOK (World Knowledge Recommender) and tested it against traditional systems on a movie recommendation task. Surprisingly, WOK showed *less* popularity bias, even without any tweaks. But could it be improved further? The researchers then experimented with prompting, instructing the LLM to recommend movies matching the user's taste for blockbusters or indie films. While this further reduced bias, it came at a cost – recommending some niche movies *too* obscure, resulting in lower recommendation accuracy. The findings? LLMs hold potential for building less biased recommenders, but striking the balance between diversity and relevance remains a challenge. The trick lies in giving these LLMs the right nudges and perhaps incorporating more structured data about user preferences, paving the way for a future where AI helps us discover the unexpected.