Description of the bug:
Hi - I am not sure where this should go, and to be honest, I do not have time to go hunting down a place to provide this information back to Google.
I have added issues in here before that we have either resolved, or have been dismissed by your teams. I get that - nature of the interaction.
BUT....
I reported previously on the issues of the multimodal models returning speech in foreign languages.
The response from your team was to 'tell the model to only respond in English'
The issue is - this is occurring in your PUBLIC models - it happens when the audio input is noisy or not elevated - your models will return an assumed foreign language - that is totally rubbish.
The next issues is apparent in most google products - the assumption of location and language.
So in the above issue, even if the model thinks my rubbish English input quality is actually Japanese - it then looks at my location and starts responding in Thai.
This location based practice in Google products needs to stop. It is wrong. It is a Biase.
This issue is with ALL Google sites providing me pages in Thai - changing the language is also not simple WHEN YOU CAN'T READ THE WORD FOR LANGUAGE.
This is most notable in Deep Research - using either the ai studio interface or calling into a model with this mode - even though the request does not state anything about Thailand - your system is then spitting out a whole heap of useless junk about my local Thai area into the research result - even though the source data of the research request is totally unrelated.
These models are great and I've for the most part really enjoyed working with them and my interactions with the staff in various teams - but the details matter. The smaller details matter the most.
Someone NEEDS to start fixing those issues.
Happy to provide links or any further evidence I can.
Actual vs expected behavior:
No response
Any other information you'd like to share?
No response
Description of the bug:
Hi - I am not sure where this should go, and to be honest, I do not have time to go hunting down a place to provide this information back to Google.
I have added issues in here before that we have either resolved, or have been dismissed by your teams. I get that - nature of the interaction.
BUT....
I reported previously on the issues of the multimodal models returning speech in foreign languages.
The response from your team was to 'tell the model to only respond in English'
The issue is - this is occurring in your PUBLIC models - it happens when the audio input is noisy or not elevated - your models will return an assumed foreign language - that is totally rubbish.
The next issues is apparent in most google products - the assumption of location and language.
So in the above issue, even if the model thinks my rubbish English input quality is actually Japanese - it then looks at my location and starts responding in Thai.
This location based practice in Google products needs to stop. It is wrong. It is a Biase.
This issue is with ALL Google sites providing me pages in Thai - changing the language is also not simple WHEN YOU CAN'T READ THE WORD FOR LANGUAGE.
This is most notable in Deep Research - using either the ai studio interface or calling into a model with this mode - even though the request does not state anything about Thailand - your system is then spitting out a whole heap of useless junk about my local Thai area into the research result - even though the source data of the research request is totally unrelated.
These models are great and I've for the most part really enjoyed working with them and my interactions with the staff in various teams - but the details matter. The smaller details matter the most.
Someone NEEDS to start fixing those issues.
Happy to provide links or any further evidence I can.
Actual vs expected behavior:
No response
Any other information you'd like to share?
No response