Several ChatGPT users are reporting a strange new bug: the model randomly inserts Arabic words and characters into otherwise normal English responses. Reports began surfacing around March 24, with affected users saying it has happened multiple times across separate conversations.
This isn't the older, well-documented issue where Arabic text displays with broken right-to-left formatting. That's a rendering problem. This is different. Users chatting entirely in English are getting responses with Arabic script spliced into the middle of sentences, unprompted and unrelated to the conversation topic.
The bug appears intermittent. Not every response is affected, and it's unclear which models or account tiers are involved. OpenAI hasn't publicly acknowledged the issue yet.
LLMs occasionally produce garbled or off-language output when something goes wrong during inference (the step where the model generates its response). It can happen when the model's internal probability distribution gets skewed toward tokens from the wrong language, sometimes triggered by specific prompt patterns or by backend changes during model updates. OpenAI has been shipping updates frequently, so a regression in token sampling isn't out of the question.
For now, it's a curiosity rather than a crisis. If you see it happen, starting a new conversation usually clears it. But if you're using ChatGPT for client-facing work or automated pipelines, it's worth spot-checking outputs until OpenAI addresses it.