@mms if I was writing open ai's spider, I'd have it in batches use a deep learning model (not generative text) to score the reliability of the information as it picks it up, with a lower scoring bias for more recent information and higher scoring bias for more reliable and multiple sources. Since it's core data that it trusts is pre 2023, I think we're far too late for any poisoning strategy to be effective outside of niche new information generation.