Unexpected Multilingual Responses from ChatGPT
Recently, English-speaking users of ChatGPT were quite surprised when they noticed the AI inserting Arabic words into its replies. This peculiar trend sparked a number of viral posts on social media.
“It happened twice on my phone and once on my work laptop. Funny enough, I’m not even in an Arabic-speaking area,” shared one puzzled individual, highlighting their bewilderment.
In one instance, a recipe included an ingredient listed in Arabic, which caught many off guard. This wasn’t the only unexpected interaction; another AI enthusiast recounted how ChatGPT switched to Arabic while assisting with a prompt for a logo.
When confronted with these multilingual slip-ups, one user expressed confusion, noting, “Slipped in? That’s a whole different alphabet!” They wondered aloud if others were experiencing the same issue.
Numerous commenters on platforms like Reddit chimed in, recalling similar encounters where the AI unexpectedly responded in languages such as Armenian, Hebrew, and Spanish.
Some users attributed these language shifts to so-called “AI hallucinations” or even suggested a decline in the AI’s overall performance. However, others pointed out that this behavior likely stems from the way AI systems manage language. Essentially, large language models use tokens corresponding to the data they’re processing—this can include everything from images to text.
To illustrate, shorter words can often be represented as a single token, while longer words might require multiple tokens. This efficiency in tokenization can lead the AI to favor certain words, including foreign ones, because they are shorter and less resource-intensive.
One Reddit user even translated the Arabic word from the recipe, revealing it meant “low,” as in “low-fat yogurt.” Another discussion noted that an Arabic phrase related to the Health Insurance Portability and Accountability Act also made logical sense in context.
This isn’t the first time ChatGPT has been caught slipping into another language. Just last year, it appeared to have a breakdown of sorts, responding in Spanglish and producing some rather nonsensical replies.
In one memorable exchange, a user asked for recommendations of jazz albums by Bill Evans, only for the bot to begin repeating “Enjoy listening” incessantly, reminiscent of a stuck record or, in a more ominous light, like the malfunctioning HAL-9000 from *2001: A Space Odyssey*.
It’s clear that, while these mix-ups can be amusing, they also highlight some of the quirks and limitations inherent in AI language models.
