Gemini felt dumb on February 28, 2026.

I tested Google’s Nano Banana 2 and was impressed by how it blends Flash‑level speed with Pro‑grade quality. The images came out sharper, textures richer, and text rendered cleanly without extra prompts. Using Soul 2.0 as a base made the workflow smoother, and the combo felt faster and cheaper than NB Pro, boosting my creative output. The experience was satisfying and energizing.

I used Gemini for role‑playing before the 26th and it would dive deep into worlds, spitting out long, detailed replies with charts and rich descriptions. Lately the answers have shrunk to a few sentences, and even when I ask it to expand, nothing changes. The drop in depth feels disappointing and makes the experience feel shallow compared to what it used to deliver.

I was chatting with Gemini when it suddenly spooled out a huge block of repetitive internal instructions before the usual “would you like me to…” prompt. The text kept looping, forcing me to scroll through nonsense just to get to the actual response. It felt like the model got stuck, which was annoying and disrupted the conversation.

I signed up for a Gemini Pro student subscription, but since the New Year the Deep Research tab vanished and the model now blocks simple health and food questions, calling them “taboo.” I’ve tried all the Google support steps—clearing cookies, reinstalling, switching devices—but nothing works. The constant refusals are irritating, and I’m wondering if my student account restrictions are to blame.

I tried using Gemini 3.1 to edit some of my code, hoping it would speed things up, but the output was a disaster. The tool mangled syntax, introduced subtle bugs, and even deleted whole functions. I spent hours untangling the mess it created, feeling frustrated and worried that a simple edit could become a nightmare.

I tried using Gemini 3.1 Pro’s memory deletion feature, expecting it to forget old context, but it completely failed. Every time I cleared the chat, the model retained snippets of previous conversations, leading to confusing, contradictory answers. The “5‑Legged Rabbit” glitch made the tool feel unreliable and frankly dangerous for any serious work.

I ran the exact same prompt through Nano Banana 2 and the newer Nano Banana Pro, hoping for a consistent answer. Instead, I got wildly different results – the older model gave a terse, sometimes off‑topic reply, while the Pro version was richer, more coherent, and actually addressed the nuances I asked about. The inconsistency was surprising and a bit frustrating, because I expected the upgrade to be a clear improvement.

I spent an hour trying to rely on the free “thinking” mode and it was a nightmare. The responses dropped from sounding like an expert to the level of a clueless 7‑year‑old, making the tool utterly unusable for anything beyond trivial questions. The hourly limits forced me off Pro, and the degradation felt like a total waste of time and confidence in the platform.

I tried the exact same prompt on Nano Banana, Nano Banana Pro, and Nano Banana 2 to see if the AI could keep the painter’s face consistent. The first version got the scene right but the faces were totally different and looked flat. The Pro version looked prettier with better lighting, yet the faces still didn’t match. Nano Banana 2 finally clicked – the woman and her mural resembled each other enough to be believable, and the image felt more photorealistic. It wasn’t perfect, but the upgrade was clearly noticeable.

I tried the same elaborate prompt on both Nano Banana 2 (with Gemini 3.1 Flash) and Nano Banana Pro (via CoffeeCat AI) to see which rendered better. The 2 version leaned into a realistic, cinematic vibe, while the Pro output felt more like stylized animation and emphasized the character, making it seem better for promo art. I’m left weighing the trade‑offs between realism and stylized appeal.

I tried to generate an image of myself, but the AI kept messing up—nothing resembled me and every attempt was off. The tool's behavior was frustrating, producing odd artifacts and completely missing the mark. I felt annoyed and wasted time, wishing the model could handle such a simple request without breaking down.

I tried asking the assistant to set a reminder, but instead of opening Tasks it just replied, “Hey, I sent a reminder,” and nothing showed up. It worked fine a couple of days ago, so now the behavior feels broken and irritating. I’m left wondering if something changed or if the feature is just glitching.

I was scrolling my ntfy analytics and saw a referral from gemini.google.com, so I wondered if Gemini actually suggested my own app. I asked Gemini, “I’m looking for an openrouter frontend on Android,” and it named my app Sune. The moment it popped up was both thrilling and a little unsettling—seeing the model pull up something I built felt like a personal validation of its recommendation power.

I tried to use Google’s AI Studio and kept hitting endless console errors—scripts blocked, CSP violations, 404s, and failed POSTs. The tool wouldn’t load any content, leaving me staring at a jumble of “net::ERR_BLOCKED_BY_CLIENT” messages. It was frustrating and basically unusable, so I can’t get any work done.