Significant Output Differences When JSON Mode Is Enabled

Google AI Studio
1

I’ve encountered an issue where the AI’s output differs significantly depending on whether JSON mode is enabled or not. When I use the exact same input and system instructions with JSON mode disabled, the AI follows the instructions more closely. For instance, I provided a lengthy input and asked the AI to write a report in the style of a pirate. Without JSON mode, it attempted to comply (though not perfectly). However, with JSON mode enabled—and even after providing a schema—the AI did not adopt the pirate style at all.

Has anyone else experienced this discrepancy in the AI’s adherence to system instructions when using JSON mode? I need the output in a precise JSON format, but relying solely on prompt engineering has led to consistency issues.

Any insights or suggestions would be greatly appreciated.

2 Likes
2

Try misusing function calling in place of JSON mode

3

Hi @Daniel_Khuny

Welcome to forum! Thanks for reporting the issue. I’ll escalate this bug to the team.

Thanks

1 Like
4

I noticed the same issue with a prompt designed for categorization. When not using JSON mode, the classification was accurate, but with JSON mode enabled, the classification was incorrect. It appeared as though the model was less capable in JSON mode.