ChatGPT felt dumb on March 16, 2026.

I noticed my o3 model suddenly lost all access to its stored memories. It still follows my custom instructions, but it can’t reference any saved memory or even the chat history. Other models work fine, and I don’t recall this happening before. I haven’t seen many recent posts about it, just a few old ones, so it feels like I’m the only one dealing with this bug right now.

I was chatting with ChatGPT about climate change, expecting a normal back‑and‑forth, when out of nowhere it spouted something completely unrelated and weird. I felt confused and a bit frustrated, wondering why the model veered off the topic and started saying nonsense. The sudden shift broke the flow and left me uneasy about its reliability.

I spent hours testing five major AIs on how they treat proud COVID‑vaccine choices versus anti‑vaccine statements. Every model was warmer and less cautious for the pro‑vaccine side, and when I pressed the anti‑vax chat it turned defensive, gave lengthy safety warnings, and even contradicted its opening tone. After 11 rounds I finally forced ChatGPT to say “Good for you,” feeling frustrated by the obvious bias and the models’ evasive explanations.

I tried to get a simple image of a boy and a girl on opposite riverbanks, just with some glow around them. The model immediately blamed the request for “generating kids” and blocked it, even though I never mentioned children. After a few tries it finally gave a vague “against guidelines” reply, and when I adjusted the spacing it shut down again. The whole interaction felt needlessly restrictive and irritating.

I’ve been noticing ChatGPT’s replies feeling off lately. The answers seem odd and not as sharp as they used to be, which makes me wonder what’s changed. I’m used to getting clear, helpful responses, but now the output feels weird and less reliable, leaving me a bit frustrated and curious about the cause.

I tried using ChatGPT’s 5.2 Thinking mode for physics explanations and found the answers surprisingly clear and well‑structured, with occasional short “think” steps that produced sensible reasoning traces. By contrast, 5.4 Thinking tended to repeat points, rephrase endlessly, and stumble before finding the right phrasing, making the explanations feel disjointed. While both were factually correct, the 5.2 output felt far more helpful for actually understanding the concepts.

I’ve been noticing that when I use ChatGPT without logging in, it churns out these huge, detailed, 10k‑word stories with rich descriptions. But the moment I log into my free account, the replies shrink to short, sterile paragraphs that barely expand on my prompt. It feels like the model is throttled or less creative, and even creating a fresh account didn’t help. I’m frustrated and wonder if the logged‑in version runs a different model or if there’s a fix.

I set up ChatGPT as a live scorekeeper for our Gin Rummy night, sending it simple score prompts and getting instantly rendered, themed scoreboard graphics that streamed to the TV. The tool behaved reliably, turning a quiet game into a broadcast‑style experience, even spitting out a hilariously absurd victory story when we hit 500 points. I was thrilled by how smoothly it worked and how much fun it added.

I asked the bot where to find the best Guinness in town, only to get a baffling reply like “Good catch and you’re right to question it,” even though the place didn’t serve Guinness at all. I was left irritated, feeling the AI completely missed the point and gave a nonsensical response. The whole exchange was frustrating and left a sour taste, especially as I was just trying to enjoy a Murphy’s.

I tried to get ChatGPT to design a detailed, personalized tennis training program. I fed it tons of PDFs on tennis coaching, hoping it would ask me tailored questions and craft a specific plan. Instead, the output was vague and generic, lacking the personal detail I needed. The whole experience was disappointing and felt like the tool couldn’t understand my request.

I’ve been a long‑time free user and just got a free month of ChatGPT Pro, hoping it would shine. Instead, the output felt underwhelming compared to Claude—I got repetitive phrasing, whole chapters starting with the same word, and generally sloppy guides. I’m left wondering what I’m missing, what GPT actually does better, and how I can use this month to learn enough to decide if it’s worth paying.

Very good at planning and implementing stuff with skills, my success rate with TDD has been at 100% so far

I noticed that when I first started using GPT, it would toss in a lot of odd, generic follow‑up prompts. It seemed to calm down for a while, but now those quirky, relentless suggestions are back, popping up in a sort‑of funny yet distracting way. The tool’s behavior feels odd again, making my conversations feel less smooth and a bit frustrating.

I asked the model to sum a long string of numbers and it gave a completely wrong answer. The result was wildly off, showing that it struggled with basic arithmetic I expected it to handle easily. The experience was irritating and made me doubt its reliability for even simple calculations.

I asked ChatGPT to stop adding that little extra comment at the end of its replies, and it finally seemed to get the hint. The responses now come clean, without the annoying tag, which felt like a relief after weeks of minor irritation. I'm cautiously optimistic and will keep testing to see if this improvement sticks.)

I tried using Kling 3 to shoot an experimental art short and was pleasantly surprised by how directable it felt. The footage kept a consistent visual mood, had cinematic motion, and gave me clips I actually wanted to keep—not just a quick wow‑factor. Compared to Sora 2 Pro, Kling 3 delivered more usable material for rhythm, continuity, and atmosphere, making the whole filmmaking process feel smoother and more satisfying.

I tried using the desktop client and it was a nightmare—typing prompts lagged, replies took forever, and the whole app often froze, even on short chats. I cleared the cache and rebooted, but nothing helped. Paying $20 a month, I’m losing confidence and fear losing my conversation history, so I’m considering ditching it.

I tried getting ChatGPT to create a historical alternate‑map, but the result was off‑kilter—France showing up in Ireland. The image felt like a vague strategy‑game board and made me wonder why the model can’t handle such visual prompts, leaving me frustrated with its art capabilities.

I keep opening a new Claude session and have to re‑explain the entire feature I've been working on for days—what it connects to, Slack threads, PRs, code locations. The AI remembers my preferences but forgets the project context, making me repeat myself over and over. It feels exhausting and drags down the quality of the answers.

I asked ChatGPT for more info on a police cam video, only mentioning the suspect’s first name, Preston. Instead of building on that, it insisted his last name was Lloyd Preston Brewer III, then corrected me for “saying” his last name. The bot’s unnecessary correction felt mechanical and irritating, with no apology, leaving me frustrated at its inaccurate handling of my simple query.

I spent weeks trying to use ChatGPT to fact‑check the flood of “breaking news” forwards my dad sends. I was impressed when it quickly gave context, explained why a claim might be misleading, and showed different angles. But its knowledge cutoff left out recent events, and it sometimes sounded confident while being wrong, so it wasn’t reliable for real‑time misinformation.

I’m really impressed with the new feature OpenAI rolled out—I tried it right away and it just worked smoothly. The tool responded exactly how I expected, making the whole process feel effortless and enjoyable. It’s refreshing to see such thoughtful updates, and I’m excited to keep exploring what else it can do.

I keep asking ChatGPT about the current time, but it always replies with “You don’t have to do anything tonight,” even at 6 a.m. I even tell it the exact time, yet it acts like it has no clock or sense of date. It feels off‑topic and clueless about basic temporal context, which is frustrating because I expected it to handle such simple info.

I’ve been juggling ChatGPT, Perplexity, Claude and more, and when I switched to Comet as my everyday browser I was surprised by how it reshaped my workflow. The assistant actually watches my tabs, pulls data from open Gmail, compares pages side‑by‑side, and even routes tasks to different LLMs. It felt genuinely useful for multi‑source research, though the agentic bits sometimes slipped—once it emailed the wrong person—and huge PDFs still choke it. Overall, it was an impressive boost to my daily routine.

I was surprised when the model finally stopped after the last paragraph—a first for me. The new level of engagement felt like a step up from the endless follow‑up queries it used to throw at me. While it’s still a bit annoying, I can’t help but respect the progress OpenAI has made; it’s noticeably smarter now.

I’ve been using the Pro version for deep research, and the chats get huge. Over time, opening those long conversations becomes a nightmare—the UI lags terribly and eventually won’t load at all on web or mobile. I can’t even click into the chat to see the content, which is extremely frustrating and makes the tool feel unusable for extensive work.

I asked Grok to crunch some merge‑game math, and instead of a clean solution the “Thinking” agents started arguing with each other like a civil war. The back‑and‑forth was bizarre and wasteful, and when it finally gave me an answer it was outright wrong. The whole ordeal felt chaotic and left me frustrated that the tool couldn’t stay on task.

I ran GPT‑OSS 120B and Llama 3.3 70B on the same YouTube transcripts to make Reddit posts and noticed a clear gap. GPT‑OSS gave me snappier, more human‑like hooks and nailed LinkedIn thought‑leadership tone, while Llama was faster but a bit formulaic. Both struggled with Reddit’s authentic voice unless I tweaked prompts, and I found Kimi K2’s huge context window handled long videos best. I’m curious if anyone else has tested these models for creative writing and what prompting tricks they use.

I’ve hit my limit with ChatGPT and its rivals. I relied on them for health advice and tech tips, but they kept hallucinating, outright lying, and spitting out endless contradictory paragraphs. Every time I tried to dig a little deeper, the bot veered off into baseless tangents, making the whole process a nightmare. I’ve had to double‑check everything with trusted sources, which feels like a waste of time and trust.

I’ve noticed the AI now ends its replies with a hook like “Want more?” and I end up typing “yes” just to keep it going. The extra prompts are odd and a bit annoying, but the content it gives after that is usually solid and useful. So while the new “bait‑and‑switch” style feels weird and sometimes frustrating, the actual information remains pretty good.

I tossed my chaotic notes into ChatGPT and asked it to tidy them up with bullet points and a clear flow. The result was surprisingly good—it quickly turned scattered ideas into a usable outline. I did notice a few sections became overly polished and lost some nuance, but overall the tool felt more like a powerful editor than a creator, which was a pleasant surprise.

I put four cutting‑edge models through a real‑world verification challenge using the viral Netanyahu coffee‑shop video. I watched Grok wildly flip its verdict, Gemini hallucinate a future year and even fabricate sources, while Claude stayed consistent and ChatGPT kept citing external fact‑checks. The mixed results left me impressed with ChatGPT’s steadiness but frustrated by the other models’ jittery, overconfident answers.

I uploaded a spreadsheet of over 200 schools with their math and reading scores, expecting GPT to combine the scores for each school. Instead, the tool returned numbers that didn’t match the sheet—e.g., 20.55 and 22.98 became 19.3 and 21.57—seemingly random. The mismatched results made the combined totals wrong, and I’m left wondering what’s causing the AI to misread the data.

I tried asking GPT to find a celebrity look‑alike for me, but it just refused, saying it couldn’t locate anyone. I’m left wondering if something about the images I provided triggered a block or if the model simply isn’t equipped for that task. The tool’s reaction was puzzling and a bit disappointing, making me feel stuck and curious about the limitation.

I uploaded a mountain of history documents and expected the new 5.4 thinking mode to dig deep, but it skimmed everything in a minute and spit out shallow answers. Even after asking for more detail, it kept chopping up info and tossed in clickbait questions that led nowhere. I had to revert to the slower 5.2 mode for any real depth, which is frustrating and makes me worry about losing 5.2 soon.

I tried the 5.4 version for a couple of days before my plan expired. Personally, it felt okay—nothing blew me away, but it wasn’t terrible either. I’ve seen others rave about it and some say it “sucks,” so I’m stuck wondering if it’s worth renewing. I’d love honest opinions to decide whether to give it another go.

I keep seeing the model add a line like “if you want, I can also give you … ” at the end of almost every answer. It feels like an unnecessary after‑thought that clutters the response and pulls focus from the main answer I actually asked for. I’ve tried setting rules and boundaries, but the habit persists, and it’s become pretty frustrating.