- Short user-generated text snippets (as few as 13 words) can easily manipulate AI agents like ChatGPT and Google AI search, leading to spam/scam outputs.
- Brands inject promotional content on platforms like Reddit, Wikipedia, and Quora for AI-engine optimization (AEO), poisoning AI tool results and citations.
- Research shows deep research agents cite user-generated content in about half of queries, with nearly a quarter from such sites, enabling end-to-end attacks.
- AI manipulation exploits lexical similarity: content mirroring user queries convinces LLMs, making targeted posting on relevant forums an effective poisoning strategy.
- Moderation struggles due to minimal text needed for manipulation; distinguishing poisoned from genuine content is hard, raising societal-level concerns about AI trust and verification.