ChatGPT’s Meltdown Resolved

ChatGPT system experienced glitches and was generating gibberish and nonsensical responses from Feb. 20 to 21.

Between February 20 and 21, OpenAI’s well-known ChatGPT artificial intelligence (AI) system experienced a minor public meltdown during which it confused and perplexed users by babbling random words and other bizarre things, including spontaneous pseudo-Shakespeare.

On February 21, at 8:14 p.m. pacific standard time, it appeared that the issue had been resolved. According to the most recent information available on OpenAI’s Status page, “ChatGPT is operating normally.”

This suggests that the issue was resolved eighteen hours after OpenAI’s initial report of it. The exact reason for the issue is now unknown and OpenAI has not yet replied to our request for comment. A quick glance at the reported outputs suggests that there was some sort of tokenization confusion with ChatGPT.

Because of the opaque nature of extensive language models constructed using GPT technology, OpenAI scientists may not be able to pinpoint the exact cause of the issue. If this is the case, the team will probably concentrate on precautionary steps such as adding more barriers to stop lengthy stretches of what appears to be gibberish.

Social media sentiment seems to suggest that the chatbot’s only real harm was that it wasted users’ time while they were expecting intelligent answers to their questions. But this example shows how generative AI systems might send erratic, surprising or hallucinogenic messages at any time.

Unwanted reactions like these can have negative consequences. For instance, when a court ordered Air Canada to pay a partial refund to a customer who had gotten inaccurate information regarding booking policies from a customer support chatbot, the airline realized it couldn’t place the blame on the algorithm.

Investors in the bitcoin space are increasingly creating portfolios and executing trades using automated methods based on LLMs and GPT technologies. The recent collapse of ChatGPT shows that unexpected failure can occur at any scale even in the most robust models.