LLMs in recommenders risk amplifying extremist content without careful prompts
New study shows LLM reranking can inadvertently boost conspiratorial material...
Deep Dive
Researchers used LLM-assisted reranking on YouTube sidebar recommendations and found that naive prompting amplifies exposure to extreme and conspiratorial political content. However, adding lightweight prompt constraints reduced promotion of extreme content and increased ideological diversity with only modest relevance loss. The work highlights that LLM-assisted personalization needs value-aware design, not just accuracy optimization.
Key Points
- Naive LLM reranking on YouTube sidebars increased exposure to conspiratorial and extreme political content for users with such histories
- Constrained prompts with topical and ideological regularization reduced extreme content promotion and boosted diversity with only slight relevance loss
- LLMs appear to rerank via statistical language regularities, not true semantic ideology recognition
Why It Matters
Reveals that LLM prompts in recommenders are value-loaded defaults that need careful design to avoid amplifying extremism.