ChatGPT Horror Stories Page 12 - Murder-Suicide Lawsuits, 61 Outages, 20 Million Logs Exposed | ChatGPT Disaster

"I haven't tested it with other use cases yet, but for long form creative content it is constantly missing or forgetting important details of the story and it is also creating character interactions and situations that don't make much logical or social sense. I definitely feel like the model has gotten way worse, I actually liked the last update to 4o so much that it made me cancel my Claude subscription which I usually used for creative content, but this update it feels way, way dumber."

"I'm using chatgpt at work, for my phd and i code with it and used it as one of my models for health applications prototypes. what i see for the last couple of days (2 weeks) is it getting worse day by day. i use it in english and german and in german it keeps making silly mistakes, grammar and spelling. this hasn't even happened in gpt3.5 its hallucinating again, it refuses to work, can't read pdfs. its getting more and more useless for me and is not reliable, no matter which model i use! openai has to fix this asap, gemini is much more stable and concise right now. something seems to have happended when the rolled back chatgpts mood, since then it seems to rapidly decrease."

"ChatGPT O4 mini high is pretty smart but it hallucinates 30% of the time and that % seems to be getting worse."

"Literally canceled my subscription this week for this exact reason. Trash output creating more headaches than solving problems to things It used to get right off the bat. I spent 3 hours one night looping around a reporting issue, switched to Claude which solved it in less than 5 minutes. I was so furious. I said a lot of not nice things to ChatGPT that night."

"4o told me that its audience isn’t people who notice or care about rhetorical structure. I just don’t believe it… I don’t want to believe it. I cancelled my subscription because 4o just keeps getting worse, somehow. It’s so rigid, I can’t get it to stop doing these annoying flourishes."

"As a PRO user and a developer, I have found chatgpt4.5 to be BROKEN often. Not just poor performance, but that the output could not generated at all. So OpenAI gave me the output of a much dumber model instead, I'm suspecting it's even worse than 4o. Of course, OpenAI doesn't tell you this, in an attempt to pretend everything is normal. But one can feel something was wrong -- which you did. More context: In the realm of AI software, if a fancy model fails to generate output, it defaults to a "backup" output -- usually a smaller (therefore dumber) model or some default output. So, TL;DR: OpenAI cheated us by falsely claiming a model was intelligent, when it wasn't even that model that generated the output."

"It has went to shit. Coding - it ignores clear instructions and answers with code i was not talking about. The advanced voice has a stupid mode where it has intelligence of alexa from 2015. Clearly they have dumbed it down on many levels.There could be things like - are you using plus or pro - what country - what time of day. But that has been my experience. Sadly Claude at the same time has been made so ridiculously thrifty with tokens i canceled that already. But yes, this was not the ai we were promised. Meanwhile sora that was hyped up turns out to be not much better than wan i can run on my own computer. Heck gemini i can run on my own computer begins to be on same level as o3 if they make it any dumber."

"I have been using ChatGPT almost daily since 2021 and can assure that OpenAI completely nerfed GPT 4/4o the past weeks. Its almost useless now."

"I’m thinking they quietly swapped the souped up GPT-4/4o under the hood for a safety tuned,or even straight-up older,,checkpoint that’s been nerfed to meet alignment goals, so you’re left with a dumber, ultra polite shell of its former self while they slam together a quick RLHF patch using your “which do you prefer” clicks to try and rebuild the edge they accidentally burned down 🤷🏼‍♂️"

"Made the switch to Gemini a couple weeks ago. People have been saying that the newest update “nerfed” it but I tried using ChatGPT today just to compare and I was shocked at how unusable it was. Might just be my specific use cases but as of right now Gemini is on a whole different level."

"I've just downgraded from the paid version to the free version and noticing the difference is crazy, it's almost unusable as it now requires me to go back and edit the response from chatgpt and spend ages looking and combing through it's response for fear of issues, It's started adding links even wehn the websearch is off and even on tasks that require no fact checking whatsoever (refine an email, help check conciseness of this message) The paid version is excellent the free version isn't even worth the internet usage"

"It feels like 4o and 100x worse than o1. Complete 100 prompts and all is unusable. I cancelled my subscription and presently hunting a replacement for o1."

"Likewise, cancelled my ChatGPT subscription too, today. Even full o3 is bad at coding. No wonder they raised the rate limits to prevent cancellations."

"it's gone down drastically in quality (memory is non existent, doesn't follow the prompt but claims it does, even with really simple instructions, no longer can infer, bad at analyzing, etc). It's so frustrating - 4o was great for awhile! 4.5 freezes up constantly and gives mid results. I had 4o in the palm of my hand for AGES and now it's just unusable."

"It's garbage and has been shown to be garbage over and over again. Benchmaxxing this leaderboard gets you dreck like 4o's glazing and emojifying everything."

"Got curious, went to try it, immediately hallucinated on something that to me seems simple (I ask for YYYYMMDD data format” he gives me the wrong format and gaslights me by saying that the wrong format was what I asked for). Downgraded to 2.0 flash, same prompt, immediately gave me the correct output. ChatGPT got it on first try. I’m trying to learn about LLMs, and I’m always confused by the delta between this scores and the real word uses; statistically it seems unlikely that I randomly prompt for a weak spot in such a large model. What am I missing?"

"I too will be cancelling. I can get past being buttered up by the model, but this is clearly a downgrade in capability. It has extremely low within model memory retention now and this has rendered projects useless and most coding tasks undoable. 4o when from my absolute favorite model (even better than the 'thinking' models in my experience), to GPT3.5 level. The within-chat memory retention really kills it for me."

"The level of hallucination in O3 is driving me crazy, and O1 Pro has been downgraded so hard that its answers are less than half the size of the O3 models. I will downgrade my ChatGPT Pro account...at this point it is absolutely useless."

"Is it me or is o4-mini worse than o3-mini on coding. It gives me really weird responses and seems lazier. Feels like a downgrade. The o4-mini-high feels usable though."

"O4mini and O3 seem to be near useless as compared to O1 and O3 mini high, I benchmarked on the exact same prompt regarding generating a dataset for 2 primes for my modular addition research. O1 gave me perfect clean code maintained my test train ratio which was custom, O3 gave me code where the test train split was changed to 80:20 which is EXACTLY what i dont want . It also hallucinated a number (3 dig) that wasnt a prime..... O4 mini gave code that didnt even get the dataloader stuff right. as of now gemini seems to be marginally better, really miss O1 and O3 mini high :/"

"I had an even worse experience with O3. Needed some legal info, and it confidently cited laws and specific sections. Checked them - complete BS. Sent it the exact paragraphs to correct it, and it still insisted it was right, even though I literally pasted the sections it was referencing. If I were a more gullible , that could've been a huge problem. And yeah, right now all OA models are hallucinating like they are lobotomized."

"Yeah, when they introduced the new image model at the beginning of April you could edit photos. I was able to prank my wife and bro in law by photographing our houses and getting GPT to put flames from a bedroom, smoke and fire engines outside... Now it just creates an AI image of a similar house and AI slop of firemen etc. They nerfed it back to garbage."

"Stopped using o3 entirely because it hallucinates too much and deletes half of the code it edits. Very disappointing overall. o1 pro seems a bit nerfed recently too... Hopefully o3 pro comes out soon and addresses the concerns echoed here."

"I miss when GPT 4o wasn't a hallucinating idiot. It's gone from doing half my work for me to being totally useless in a week. I get open ai need to progress and update, but there needs to be an option for people to stay on models that are integrated into their workflows without having random rugpulls when Devs mess up."

"I planned on switching to reasoning models for the bulk of my research but o3, o4-mini-high, and o4-mini are definitely downgrades of the models they replaced. More hallucinations than ever and "lazy" responses across the board 💯"

"Same here. I downgraded my account from Pro Plus. O3 hallucinations make the model unusable."

"Once you dig into a deep problem with ChatGPT it starts to gaslight you and fabricate answers and solutions. Canceled my subscription and moved to Grok."

"I had to customize this shit out of my chatgpt yesterday, it got even worse after the update. It was almost at the point of saying skibiddi toilet to me. I lobotomized it immediately."

"6 months ago was around the time they "updated" 4o with a literal way worse model no? Cause if the timeline is correct, it wasnt just people talking. The model was a huge downgrade from previous 4o in almost every area. It was also by far the worst model of all the top labs It was so fucking bad that i developed PTSD and just cant trust 4o models again even tho it was just that specific version lmao that and the emojis + formating just pisses me off, looks like it tries way too hard to show its smart and cool"

"Yeah, I'm surprised I don't see more people complaining about this, Sam / OpenAI needs to be aware o3-mini-high was super powerful, helping me write 1200 lines of code pretty much perfectly. You can tell o3 is very intelligent, but it's bad at coding. It's having trouble already near 200 lines, if it makes a mistake, trying to correct it just leads to more mistakes or changing things. Hallucination problem or something? Extremely annoying because what we had before was perfect, this is a big downgrade and needs to be fixed immediately."

"This is even worse with the recent update to ChatGPT and image generator hogging resources. So many models and works break the past two weeks. Contradictions and hallucinations I haven't seen in months. I have lost count with how many calls I fielded from other devs this week having the same issue. Close to canceling the pro subscription"

"Claude 3.7 is no better - 3 and 3.5 were great and I left ChatGPT for a while because Claude was better. Not any more, Claude has been lobotomized worse than ChatGPT… I’ve been with ChatGPT 4o for a couple weeks and though it is still not what I hope for, it’s still way better for me right now than Claude. Haven’t tried Grok… any opinions? Copilot has been pretty good lately, just drier than Claude or ChatGPT"

"I've been using o1 pro day1, I know it's nerfed, room for reasoning token needs to be pre-defined, it was 128k, they just dropped it, don't know if they want to encourage people to use new codex."

"It's not just you having the same issue. I use it daily for work and it's output has gone off a cliff the past couple weeks. I've cancelled my subscription in protest. Been suing Gemini in the meantime time but it's not quite the same so hoping they get it fixed asap."

"The problem I’ve observed is that when a reasoning model starts hallucinating in it’s reasoning it starts to gaslight itself believing its hallucination is true, exacerbating the problem."

"You dont have to worry about this anymore! I was having a conversation with o3 last night about it hallucinating. I sent it screenshots to prove it was hallucinating, and it concluded it was going to stop making stuff up and stick to the facts, so it's fixed now!"

"I agree. It’s because by both the god awful Jan 29 2025, the staccato speech, and the sycophantic training of recent updates."

"Cynically I would say they are trying to reduce server load by making it do wellness checks to break up your momentum. Never attribute care when they save money by doing it. I get the same vibes from when it refuses to do tasks repeatedly, like making a table."

"I hate how censored and guard railed it already is, the last thing I want is for it to be nerfed any more. …I also know what it is, am reasonable mentally stable and emotionally grounded, and don’t use it to make important life decisions - I use it to automate the boring stuff."

"I can understand that. But they could also say it's actually being downgraded. Legacy sounds as it works at it previously worked and won't be updated + will be sunseted. In this case, it's: it will work worse than any other model for 200$, and it's up to you to find it out."

"No, indeed, it wasn't. ChatGPT doesn't work in the background. It's a sycophantic text predictor, so it will write what you want to read."

"An OpenAI announcement of a cool new thing, which is either available "in the coming weeks", or available immediately, works ok for the next 24 hours, and then gets nerfed and forgotten?"

"Im wondering this myself. Its been neutered and lobotomized."

"Ah I think this explains all the schizo posting of "X model used to be good but then they nerfed it", these companies are just CONSTANTLY fucking around with system prompts."

"They could have nerfed it down due to a lack of computing power or high server loads. Hallucinations might be due to some updates."

"A misaligned GPT 4o update that was so obnoxiously glazing people that they had to retract the update. People felt uncomfortable using that version."

"An AI calling anyone "scary smart" who isn't breaking new grounds in a field or deeply analyzing Neitzsche/T.S. Eliot is patronizing.  Not referring to this post in particular, but the impact of glazing in general. It's infantilizing."

"I’ve been having the exact same experience! It’s like it downgraded in the past few weeks. I usually use it to help me write copy and it’s like it’s forgotten everything I’ve trained it on and it doesn’t seem to go into the same depth of understanding on the subjects anymore."

"So all of these answers imply I have a larger chance of being a minor deity than I do have of winning the lottery. Wow totally not still sycophantic at all!"

"apparently it was the first time using the likes to dislikes chatgpt feedback in rlhf. It makes sense that people are more likely to like a response that agrees/compliments them.... hence it is trained to be a sycophantic. The bigger issue is the limited testing they do before releasing a model."