Non-latin characters in quick answer

bert

Query: was beta amyloid research overturned?

Quick Answer: Research into beta-amyloid as a cause of Alzheimer's disease has faced challenges and এসেছে controversies, but has not been entirely overturned. Here's a summary: [and then several bullets that I'm not bothering to copy here]

Not sure what that non-Latin (Devanagari?) text is doing in the middle of the answer.

Don't include random words in other languages in Quick Answer output.

Hanbyeol

Although it's not exactly like the OP's issue, I also see a similar problem where Chinese text is appearing between English text. (Chinese : 疑慮)

References:
[1] https://www.nbcnews.com/science/science-news/alzheimers-theory-undermined-accusations-fabricated-research-rcna39843
[2] https://www.acs.org/pressroom/reactions/library/alzheimers-research-retracted-paper.html

MattM

Right, so this issue has been there for a while and deeply investigated. It's actually coming from the underlying model (at time of writing, gemini flash 2.0) -- and it's not even related to the presence of bengali characters in the actual search page, it's just the model going into hallucination for some weird reason.

Right now the fix for it will be when we change the core model (very soon!) but we're aware of it, and thanks a lot for reporting it nonetheless.

fxgn

Quick answer randomly included japanese characters

It shouldn't do this

bert

Here is another instance of this happening.

Query: “short float?”

marek

Searching with AI using English (but with region selected as Poland) generated correct results, but there was a Cyrillic word in one place. Looks like the output is as expected non-deterministic: retry makes the issue to dissapear.

Only Latin alphabet