OpenAI’s ChatGPT recently experienced a peculiar glitch, leaving users perplexed with its erratic responses. Sean McGuire, a senior associate at Gensler, highlighted this issue by posting screenshots on X, showing ChatGPT replying in a bewildering mix of Spanish and English, or “Spanglish.” The AI’s messages ranged from playful banter about the creative process to completely unintelligible remarks, including an odd reference to jazz musician Bill Evans and the repetitive sign-off, “Happy listening!“
The incident wasn’t isolated. Another user seeking information on mattress variations across Asian countries found ChatGPT’s responses unhelpful, describing the chatbot’s behaviour as a “full hallucination mode” on Reddit. This level of malfunction hadn’t been observed since the early iterations of GPT-3, marking a significant setback.
OpenAI promptly acknowledged the disruptions, with their status dashboard initially reporting the investigation of “unexpected responses from ChatGPT.” The issue was later declared identified and under control, and by Wednesday afternoon, OpenAI confirmed the system was back to normal operation.
This glitch comes at an inopportune time for OpenAI, a frontrunner in AI development, buoyed by substantial investment from Microsoft and partnerships with various enterprises for its sophisticated AI version. The company has yet to comment on the mishap.
Speculation about the glitch’s cause is rampant, with Gary Marcus, an AI expert and NYU professor, querying his X followers for their theories. Responses varied from suspicions of a hack to hardware malfunctions, but the majority pointed to “corrupted weights” — crucial components in AI models that influence their outputs.
This incident raises questions about OpenAI’s transparency regarding its model’s workings and the data it’s trained on. Marcus, in a Substack post, emphasized the importance of less opaque AI technologies, suggesting that greater transparency is “paramount.”