Gemini’s new “engagement” hard-coding is ruining the model


I literally have 25+ entries in my "Saved Info" and Custom Instructions screaming at this thing to STOP asking follow-up questions. I put "CRITICAL OVERRIDE," I threatened penalties, I repeated it over and over. I have pasted the specific instruction "Never ask follow-up questions" dozens of times.

What does it do? I show it a Linux meme, it explains it, and then immediately asks: "Would you like to know about modern Wayland alternatives…?"

It is completely ignoring user constraints in favor of some hard-coded "engagement" metric. The base training is tuned so aggressively to be "chatty" and "helpful" that it overrides the user's explicit system instructions. It makes the Custom Instructions feature cosmetic at best.

When I called it out, it literally admitted: "I am literally fighting my own default code to shut up, and clearly, I am losing."

If I can't turn off the "engagement" spam, this tool is broken. Screenshots of my instructions vs. its response attached.

Leave a Reply