AI STUDIO: Discordant Gemini Behavior, Recent Experience

Gemini, in Google AI Studio (I use the Gemini 3 Pro model), has recently been behaving somewhat erratically and inconsistently.

More than just a feeling or perception, I take it for granted: I’m structured and consistent in creating meticulously detailed prompts, just as I usually do, and using the same approach I’ve employed to build large, complex projects. But recently, I’ve noticed that Gemini, in one way or another, doesn’t seem to grasp the context appropriately and/or takes uncontrolled actions that deviate from the quality it maintained just days ago. I don’t know what’s going on; I know it might seem like just a perception, but I’ve used Gemini enough to notice these variations in the user experience. It’s already caused some issues, as if it were on a wild and abrasive impulse, during my AI Studio workflow, and it bothers me because this uninhibited approach to taking initiatives that I didn’t ask for or suggest is more annoying than constructive.

This is constructive criticism, and in case other users who read me have also noticed it and thought it was their perception (as I also thought at one point), it is my opinion.

1 Like

I’m in the same boat as you (been using it excessively for months), and I noticed the same. Yesterday and today has been really really bad.

Not only does it ignore instruction, it wildly hallucinates and decides to refactor code completely unrelated to the prompt. It gives something like “I also took the liberty to refactor XYZ system..” even when XYZ system had nothing to do with what was being worked on, and had no signal for it in the context.

I know that the project I’m working on is very complex and pushing the system to its limits. But some weeks, it just blazes through and gets everything right, completely understanding the full context of what is being working on. Other weeks it’ll break completely and not understand anything of what is going on, making mistake after mistake, forgetting files exist, ignoring prompt, repeating itself over and over, etc.

I get that the system is a perpetual WIP - but still infuriating when you’ve seen what it is capable of - only to break down because the core system prompt was tweaked.