Gemini felt dumb on December 7, 2025.

I asked Gemini for some code, and instead of text it handed me an AI‑generated image of the code. It was ridiculous—nothing I could copy or run, just a picture that looked like it belonged on a meme page. I felt annoyed and a bit amused, because the tool completely missed the point of my request, turning a simple coding task into a visual joke.

I’ve been grinding on a single architectural visualization for days, feeding the gemini-3-pro-image-preview model the same floorplan and detailed prompts, but 90% of the generations are completely off—either exact copies of the input or bizarre hallucinations. After burning through hundreds of attempts and paying for each one, the results are still unusable, leaving me frustrated and worried the model has just gotten worse.

I tried using Gemini’s image updates, hoping it would refine a picture I’d already generated. Instead, if the first output isn’t perfect I have to start a brand‑new chat. When I point out missing elements—like 7 and 8—it just repeats the same image over and over, refusing to build on the original. The tool’s behavior was frustrating and felt stuck.

I tried Gemini 3 Pro’s Deep Research on a supplement, uploading a product doc and label and asking for a 50‑citation APA bibliography. The report itself was detailed, but only the uploaded doc showed up as a citation. It hid 100+ sources under “Sources reviewed but not included,” leaving me clueless about which studies actually informed the content. The missing references made the output practically useless for my needs.

I tried switching from Claude Opus 4.5 to Gemini 3 after hitting a rate limit, hoping to finish my work. Instead, Gemini 3 started rambling, never grouping its thoughts, and kept backtracking with “wait, no wait” prompts. It even got stuck in an endless “done” loop, forcing me to kill the chat. The experience was frustrating and unproductive.

I tried using the AI to crank up the heel height on a meme image, expecting a fun, exaggerated result. Instead, the model just spat out the original picture, refusing to make the shoes any taller—even when I fed it a photo of just the heels. It was flat-out uncooperative, which was pretty frustrating given how much I usually enjoy using it.

I use Gemini every day and love its top‑tier response quality, but I’ve hit a few pain points. It can be sluggish, and I’d rather wait a bit than lose its nuanced reasoning. The voice input often cuts off when I pause, making lengthy dictations a hassle compared to GPT’s smoother handling. Finally, the Deep Research mode only spits out full reports; I’d love a middle‑ground chat mode that gives deep, citation‑rich answers without creating a formal document.These tweaks would make the tool feel even more polished.

I tried moving my ChatGPT history over to Gemini 3.0 during the free trial. At first it was amazing—the conversation continued seamlessly—but soon I hit two annoying roadblocks. I can’t scroll back far enough, so the early days of my chat vanish, and Gemini won’t read a .json file once I start a new conversation. It feels limiting and a bit frustrating, especially after the initial wow moment.

I tried using Antigravity and it just spewed a endless “Reply. Done.” loop, completely hijacking my notebook. The tool kept repeating the same phrase instead of answering my prompt, so I ended up with a massive block of nonsense that polluted my workspace. It was frustrating and unusable, leaving me unsure what went wrong.

I was relying on Gemini’s AI Pro subscription, which promised a 5‑hour token refresh rate. After the AntiGravity update, that refresh window vanished despite the changelog saying Pro and Ultra users would keep it. I’ve been stuck watching my tokens drain far faster than expected, and it’s driving me nuts—feeling powerless and irritated that a promised feature disappeared without notice.

I tried to push Nano Banana’s image‑to‑image feature beyond simple tweaks, asking it to totally re‑imagine a photo in a cubist style. Instead it just blurred the picture, kept the same composition, and added vague dark edges. Every time I changed the prompt it clung to the previous “idea,” forcing me to start a new chat for any real variation. Even with pixel art, it refused to produce smooth vector results, leaving me frustrated and unimpressed.