ChatGPT Quality and Behavior Changes

Something started showing up in the ChatGPT subreddit around March 23rd, and it was strange enough that people kept posting about it for the next ten days. Users reporting that mid-conversation, without any prompt or context shift, ChatGPT was dropping words or phrases in Arabic, Hebrew, Persian, Cyrillic, or Thai into otherwise English replies. Not translating. Not code-switching for a reason. Just inserting fragments — one user flagged the Persian word for "regions" appearing in an unrelated response, another described Arabic script appearing mid-sentence like a hiccup in the model's output stream. By March 25th, multiple independent posts were describing the same pattern across different accounts and use cases, which is the point where coincidence starts to feel less convincing. Then, around March 26th, a different complaint emerged alongside it: that some responses felt tonally off, almost like the model was generating its own task prompts rather than responding naturally to users. By early April, a third thread arrived with a separate but related concern — that ChatGPT had become noticeably more conservative and evasive in its answers. Whether these are three symptoms of the same underlying change, or three unrelated grievances landing in the same window, is genuinely unclear.

If confirmed, here is what this means. Random multilingual injection in a model trained on English prompts suggests something went wrong at the token level — possibly a training data contamination issue, a sampling parameter change, or a bug introduced during a quiet model update. OpenAI deploys changes continuously and without announcements, which means millions of users can be running on a meaningfully different model without knowing it. The conservative behavior shift is the more strategically significant signal: if the model has been nudged toward refusal or hedging without public disclosure, that affects everyone from developers building on the API to businesses using ChatGPT as a customer-facing tool. Second-order, it raises the question of what users are actually running when they open ChatGPT — and whether the product consistency that enterprise customers are paying for is holding.

Watch for OpenAI's system status page or any acknowledgment from staff accounts on the subreddit, and pay attention to whether the foreign-language injection reports drop off suddenly — a quiet fix with no announcement would itself be telling.

NewsHive monitors these sources continuously. All signal titles above link to the original reporting.

Intelligence by NewsHive. Need help navigating what this means for your business? Contact GeekyBee →