Gemini randomly dumped 15 weeks of some stranger's personal Japanese writing into my homework chat, and Google says it's just a hallucination
So this was weird. I was working on an assignment in a Jupyter notebook and sent Gemini a totally normal English prompt. Instead of helping with my code, the response had a big chunk of Japanese text in it. It looked like someone's weekly personal column from a while back, about 15 entries in a row, all written the same way.
What got me is that it wasn't random-sounding filler. Each entry reads like a real person writing about their week. They mention kids who come to visit a community space they run, a lecture they attended and what stuck with them, a local event for seniors, a walk around a historical site with a volunteer guide, and the weather and how tired everyone is at the end of summer. Then they tie it all into their own reflections on faith. The entries also refer to actual events that really happened and actual places in a real town in a country that's named in the text. There are no personal names anywhere, but it's close enough to personally identifiable that I'm pretty sure if I put in the time and effort I could figure out whose writing this is. I'm not going to, obviously.
I reported it to Google and they just said it was a hallucination. I really don't think it is. Making up text that's this consistent and this tied to real places and events seems a lot less likely than the model repeating something it actually saw.
I did search online for a bunch of distinctive sentences and couldn't find it anywhere. It could be from some deleted public thing that's no longer around, but I doubt it.
My suspicion is that it's either someone else's session leaking into mine, which I don't think is very likely, or more likely some kind of error in how the model generated its output that caused it to spit out a coherent block of its training data. Basically a glitch that made it regurgitate memorized text instead of answering my prompt. Either way, my concern is that this stuff is still sensitive. Even if something was public once or ended up in training data, I don't think anyone would want their own writing and details about their life coming out like this.
idk
For context, I was using Gemini 3.1 Pro with high thinking in Antigravity when this happened.
