Gemini felt dumb on February 24, 2026.

I fed Gemini 3 a long list of just “(” or “)” characters—thousands of lines—and the model started hallucinating a prompt out of nowhere. The weird output was unexpected and confusing, making the tool feel unreliable for simple tests like this.

I was exploring Imagen 3’s archive, loving the photorealistic 3D results, but when I tried the new generation flow I was let down. Clicking ‘Repeat’ now yields almost identical images most of the time, and there’s no way to disable the seed in chat. The lack of variety feels frustrating and makes the tool feel limited.

I tried the Gemini 3.1 Pro on a 2026 AIME 1 problem, expecting it to handle the math like the earlier 3.0 version did. Instead it spit out 273 while the correct answer is 83. The result was a let‑down; the model clearly missed the core reasoning and gave a wildly incorrect number, confirming it’s still far from reliable on advanced contest math.

I was shocked when Gemini gave me a completely unrelated transcript after I asked a simple question. The response was off‑topic and couldn’t be edited, which made it even creepier. I’m left puzzled and uneasy about what the AI actually did, because it seemed to hallucinate something totally irrelevant to my prompt.

I was amazed when I tried Gemini 3.1 after the frustrating 3.0 pro experience. Previously any tiny tweak rewrote the whole file, grinding to a halt on larger scripts. This new version feels lightning‑fast, editing only the bits I change. Watching the code screen show targeted replacements was a relief, turning what used to be a tedious chore into a smooth, enjoyable vibe‑coding session.

I kept getting an infinite loading spinner after each response, which made using the tool feel sluggish and irritating. Every time I tried to get an answer, the UI never finished loading, breaking my flow and forcing me to wait or restart. The constant hang‑up was frustrating and stopped me from getting any real work done.