What Happened
On February 20, 2024, ChatGPT users began reporting that the model was producing bizarre, nonsensical outputs. Instead of coherent responses, ChatGPT was generating garbled text that mixed multiple languages, repeated phrases in loops, produced strings of seemingly random characters, and in some cases generated responses that appeared to be fragments of training data surfacing in raw form.
The outputs were unlike normal hallucinations, which at least maintain grammatical coherence. These were visibly broken at the text generation level, suggesting a deeper technical issue than typical model behavior.
Timeline
Reports began appearing on social media and the OpenAI community forums during the evening of February 20, 2024. Screenshots of the bizarre outputs went viral quickly, with users comparing the behavior to a “digital stroke” or “AI having a nightmare.” OpenAI acknowledged the issue within hours and attributed it to a bug. The issue was resolved by the following day.
Impact
While the technical impact was short-lived, the psychological impact was notable. The glitch outputs were genuinely unsettling to many users, particularly those who encountered them without knowing about the bug. The viral nature of the screenshots (which looked like something from a science fiction horror film) temporarily amplified concerns about AI stability and predictability.
The incident also demonstrated how quickly AI failure modes become viral content, with millions of people viewing screenshots of the glitch text within hours.
Response
OpenAI’s status page acknowledged the issue and attributed it to a bug affecting how the model generated text. The company stated that the issue was identified and resolved quickly, and that no user data was compromised. The technical root cause was related to the model’s token generation process rather than the model itself.
Lessons Learned
The glitch text incident illustrated that AI systems can fail in unexpected and alarming ways that go beyond the well-understood category of hallucinations. It demonstrated the importance of monitoring systems that can detect when model outputs deviate not just from accuracy but from basic coherence. The viral response also showed that the public is primed to react strongly to evidence of AI instability, making robust production monitoring essential for maintaining user trust.