Gemini felt mid on February 10, 2026.

I’ve been testing Gemini 3 Flash and it’s a mixed bag. When I ask about serious topics, like anti‑racist studies, it actually pulls up real research I hadn’t seen before, which felt impressive. But the moment I probe obscure books or obscure musicians, it spins convincing‑sounding lies instead of saying “I don’t know.” That confident hallucination makes me uneasy, especially since competitors will at least trigger a web search when uncertain. The tool’s knowledge is useful, yet its weak uncertainty detection erodes my trust.

今天表現還行

I noticed Gemini 3 Pro would gloss over thin evidence when summarising large datasets, turning weak data into confident statements that could mislead leaders. It was frustrating because important nuances disappeared. I tackled this by adding an “Evidence Floor” prompt, forcing the model to only state insights that meet a minimum data threshold, flagging or dropping the rest. This turned the summarisation into a clearer, evidence‑based filtering process.