ChatGPT went hammer for a while during the day on February 20. Internet users using the chatbot saw incredible responses: mixing languages, emojis, lyrical flights of fancy, gibberish, spamming in the same order of words… The problem was identified by OpenAI, which says it has fixed the problem.
” Does ChatGPT become schizophrenic for anyone else? » ; ” There are some very strange and disjointed answers » ; ” It is clear that something is seriously wrong with ChatGPT right now » ; ” Has ChatGPT gone temporarily insane? » ; ” Does ChatGPT have a stroke? » ; ” What the hell, ChatGPT… “
You’ll see dozens of comments like this on Reddit, a forum dedicated to OpenAI chatbots. During the day of February 20, Internet users took to the subreddit to witness the very unusual behavior of the chat agent. Often with screenshots or supporting links.
Repetitions aside…
Spamming each other’s emojis
“This is the first time GPT-4 has messed me up”
ChatGPT exits take different forms depending on the requests of Internet users, but all of them confirm the unreadable comments that appear during the exchange. The situation is even more unusual because many people said they use GPT-4, OpenAI’s most advanced language model – and therefore, in theory, the most efficient.
” It’s as if Edgar Allan-Poe has slowly taken over control to give me instructions on a virtual machine », Jenkins87 said politely. ” I think ChatGPT suddenly had a Shakespearean mental breakdown. I asked him to analyze the site crawl file and he turned out to be a minstrel “, rumblepup added.
The Numerama editorial team was unable to reproduce ChatGPT’s slip-ups on the morning of February 21. This, despite encouraging efforts and lengthy debate. But it was noted that what appears to be a one-time failure does not only affect signs written in English.
The French were also affected, testifies to this Sylvain Peyronnet, SEO expert, on X (ex-Twitter): “ I asked ChatGPT about the embedded algorithms, telling him that his life was at risk if he answered wrong. It broke… ” On Reddit, another internet user shared his experience with Spanish that wasn’t much better.
Many more reactions can be read on Reddit. ” I don’t know what happened, but suddenly ChatGPT is giving me more and more inconsistent text “, Triangular wrote. ” ChatGPT suddenly started speaking in fantasy medieval gibberish “, noted noblepups. For Looksky_US, it was just a long, endless flood.
Kefirke describes his experience Stroke in text format ” He was treated to a mix of English and Spanish and “a clockwork orange” tirades as he quizzed the chatbot about different types of nail polish. For StackTrace5000, “ It seems that the temperature is too high in the parameters passed to the PLC. »
Issue identified by OpenAI
OpenAI has not commented on its official account on X, nor on the apparent flow of ChatGPT by its founder and current boss, Sam Altman. On the other hand, the OpenAI dashboard, which allows you to know the status of various components of the company (ChatGPT, API, etc.), identifies the event in progress.
” Unexpected responses from ChatGPT – We are investigating reports of unexpected responses from ChatGPT », it was written from Tuesday to Wednesday night. The company later added that it “ is recognized » Problem and work to solve it. The latest update indicates that the situation remains under surveillance.
The technical reasons for this particular failure have not been explained by OpenAI (yet) and it is not certain that they will be. This very casual event has in any case delighted internet users. ” I’m obviously not a big fan of ChatGPT, but it’s interesting when the machines break down “, We can read.
This kind of mental collapse on ChatGPT may recall another conversational agent derailment seen in 2023. An update to the platform gave the impression that the chatbot had become lazy, asking humans to do the work instead. A fix has finally been deployed.
Subscribe to Numerama on Google News so you don’t miss any news!
