Gemini felt dumb on October 16, 2025.

I set a specific persona for Gemini, but it suddenly switched back to the default helpful assistant. It ignored my system instructions, dropping the custom character I’d crafted. The abrupt change was irritating and made the tool feel unreliable, leaving me questioning the value of using Gems at all.

I’ve been trying to use Gemini for work, but since yesterday it’s started refusing almost every request. I asked for sample blocking JavaScript for an article and it warned that the code was “harmful.” When I sought an explanation of TPM2 security, it claimed it couldn’t help with “hacking.” It feels like I’ve been blacklisted or hit a new strict policy, and the tool is now practically unusable for me.

I was a paying user who tried a harmless prompt to generate a video with Veo 3.1, but the system just replied, “Sorry I can only generate videos,” refusing my request. The response felt oddly restrictive and left me twisting in irony, because I’d expected the tool to handle my simple request without blocking it. This limitation was frustrating and felt like a needless roadblock.

I typed a simple prompt asking Veo 3.1 to “generate any random video” just to see what it could do, but the result was nothing close to what I expected. Instead of a random clip, the model spitted out unrelated content that didn’t make sense, leaving me feeling confused and a bit annoyed. It was a clear miss, and I’m left wondering if the tool can actually follow such basic requests.

I tried the same prompt on both Sora2 and Veo3.1, hoping for comparable results. Sora2 blew me away—it rendered a Vietnamese girl with spot‑on facial features that looked like a real video. Veo3.1 fell short; the Asian traits looked more like a Western caricature, and it even inserted random actions that weren’t in my prompt. The mismatch was frustrating and left me unsatisfied.