Gemini felt dumb on February 2, 2026.

I’ve been hammering the same highly specific custom instructions into the model, and while some replies get a noticeable boost, others turn completely bizarre. The inconsistency is frustrating—I want the tool to fire those instructions only when the context truly matches, but right now it’s all over the place. I’m looking for a way to lock the triggers so the model behaves predictably.

I tried using Google’s search AI and was shocked to see a glaring typo—“if1one” instead of “if one.” The mistake broke the flow of the philosophical quote and made the output look careless. It felt annoying and undermined my trust in the tool, especially since it’s supposed to be powered by a Gemini model.

I tried feeding Gemini dozens of URLs, hoping it would fetch the actual pages, but it just spun out fabricated content and never provided any citations. It acted as if it had read the site, yet the info was off‑topic or pure hallucination. Seeing ChatGPT pull the same links and actually cite sources made the contrast stark, leaving me irritated and doubtful of Gemini’s reliability.

I fed a YouTube transcript into the AI with a prompt to act as a technical documentation expert. It stripped out all the filler and gave me a concise, numbered markdown checklist of exact actions—clicks, commands, shortcuts. The result turned a 40‑minute React tutorial into a 15‑line list, letting me get the app up and running in minutes instead of watching the whole video. The tool felt empowering and dramatically sped up my learning.

I fed six months of venting chats and notes into Gemini 3 Pro, asking it to act like a behavioral data scientist and spot hidden triggers. The AI quickly identified that my anxiety spikes right after I skip sleep on Sunday nights. That insight let me tweak my schedule, and the panic vanished. The whole process felt like turning personal drama into clear, debuggable code.

I kept trying to get Gemini to do something specific, but it kept spitting out random images I didn't ask for. When I finally managed to get the image I wanted, it was a laughable result. The whole experience was irritating, turning a simple task into a confusing back‑and‑forth with the AI.