one year on
ChatGPT speaks gibberish overnight, OpenAI investigates unexpected responses
Users flood Reddit and X with screenshots of the chatbot producing Spanglish word salad and repetitive loops as OpenAI says it is investigating.
OpenAI’s ChatGPT is speaking overnight in tongues, producing nonsensical responses that mix English, Spanish, and surreal ramblings. Users on Reddit and X post screenshots of the chatbot answering simple prompts with phrases like ‘a mouse of science, an easy draw of a sad few’ and repeating the same sentence until the screen fills.
OpenAI acknowledges the issue on its status page at 6:40 p.m. ET Tuesday, saying it is ‘investigating reports of unexpected responses.’ Seven minutes later, the company says the issue has been identified and is being remediated.
Cognitive scientist Gary Marcus calls the episode ‘a warning’ and argues that the systems have never been stable, lacking safety guarantees. Users share the screenshots widely on Reddit and X, with some describing the output as ‘word salad.’
The record
Cognitive scientist Gary Marcus described the behavior as 'a warning' and argued that generative AI systems have never been stable and lack safety guarantees.
One Reddit user noted that responses begin normally then 'devolve into nonsense,' while another asked 'Is my GPT having a stroke?'
One year later — open only if you can handle spoilers
The bug was a config error in the inference kernel, not a rogue AI. It became a defining 'wtf' moment in ChatGPT's early history, often cited in discussions about model reliability.
The Weekly Replay · free by email
This week, one year ago — every Sunday.
One email each Sunday: the week's replayed AI news, with the one-year-later annotations included. Written like it's breaking — dated like it isn't.
Free · double opt-in · unsubscribe anytime · privacy