Google AI summary: „parachute effectiveness"

Rather than be related to training data this looks like a RAG issue - even if you don't have the AI answers (I don't get them) the first result is:

"Conclusions: *Parachute use did not reduce death or major traumatic injury when jumping from aircraft* in the first randomized evaluation of this intervention. However, the trial was only able to enroll participants on small stationary aircraft on the ground, suggesting cautious extrapolation to high altitude jumps."

Google is leveraging their normal search tech to bring up extracts of some number of webpages which match the query, and then adding the AI feature on top of that to synthesise a single result from those extracts. It's meant to improve on pure LLM output by putting out more topical results - I think that on average though you get worse results than the pure output of an LLM. This is because you're telling it to base its answer on the results, which might be satire (like in this case), or by someone lying on purpose, or written by an idiot who has no idea what they're talking about, or some other kind of outlier from the mean.

On the other hand, pure LLM output will be more likely to be correct because it's based on the most likely next token (which is something similar to common sense) - "parachutes are ..." is more likely to be followed by "useful when skydiving" than "useless when skydiving".