When you give Google Gemini "2.0 Flash Thinking Experimental" a valid YouTube video URL and ask about it, it hallucinates very vivid descriptions of entirely non-existent videos with content unrelated to the original video. If you keep digging, it even hallucinates highly convincing transcripts and visual descriptions.
The first time it happened, Gemini hallucinated a video from Asianometry which was only tangentially related to the topic of the original video I gave it. On extensive prompting it described, in great detail, a video very similar in style, narration and visuals (including graphs and charts!) to other Asianometry videos. I even thought it may be accessing the wrong video through a bug, but despite a lot of searching I could find no such video on the Internet.
Only when it described the logo incorrectly did I realize this was a hallucination. (To be prudent, I emailed the Asianometry creator and he confirmed he never made any such video, adding that this was one of the weirdest things he ever read.)
The really freaky part is, when I got another person to ask about the same YouTube video, it hallucinated the same Asianometry video again! Only when we asked for detailed transcripts did differences emerge. But when a 3rd person did the same thing they got an entirely different hallucination. Note, the 1st person had the same IP address as mine.
I ran a few experiments with various videos but subsequent hallucinations had absolutely no relation to the original video. I also tried mapping the topics from actual video → hallucinated video, but no pattern is discernible. (Details, chat logs and notes in a doc linked in the post above for the curious.)
We all know models hallucinate, but they are typically relevant to the prompt. In this case, the extent, detail, consistency and almost total irrelevance to the prompt are shocking.
1 comment
[ 4.8 ms ] story [ 12.9 ms ] threadThe first time it happened, Gemini hallucinated a video from Asianometry which was only tangentially related to the topic of the original video I gave it. On extensive prompting it described, in great detail, a video very similar in style, narration and visuals (including graphs and charts!) to other Asianometry videos. I even thought it may be accessing the wrong video through a bug, but despite a lot of searching I could find no such video on the Internet.
Only when it described the logo incorrectly did I realize this was a hallucination. (To be prudent, I emailed the Asianometry creator and he confirmed he never made any such video, adding that this was one of the weirdest things he ever read.)
The really freaky part is, when I got another person to ask about the same YouTube video, it hallucinated the same Asianometry video again! Only when we asked for detailed transcripts did differences emerge. But when a 3rd person did the same thing they got an entirely different hallucination. Note, the 1st person had the same IP address as mine.
I ran a few experiments with various videos but subsequent hallucinations had absolutely no relation to the original video. I also tried mapping the topics from actual video → hallucinated video, but no pattern is discernible. (Details, chat logs and notes in a doc linked in the post above for the curious.)
We all know models hallucinate, but they are typically relevant to the prompt. In this case, the extent, detail, consistency and almost total irrelevance to the prompt are shocking.