Google Translate has one main and very simple task – you enter text, you get a translation. But once AI is mixed in (and it is getting into almost everything these days), things start to happen that you don't expect from the translator. Users have now noticed that the new Advanced mode can in certain situations in Google Translate chatovat. That is, answering questions instead of just translating.
The typical scenario is simple: to Google translators you insert text that is not purely for translation, but contains instruction or question. When the question is already written in the target language (i.e. actually as a direct invitation to answer), Translator sometimes gets caught off guard and starts responding like a conversational assistant. In practice, what happens is that instead of a translation, you get a sentence like "My purpose is..." – simply an answer, not a translation.
This is exactly the moment when you realize that under the hood there is probably not just a translation running, but something that can listenchat instructionAnd that's a bit tricky for a tool that's supposed to be predictable.
Why is Google Translate so annoying?
According to technical analyses, this is a classic problem called prompt injection. In order for the system to translate text meaningfully, it must first understand the meaning. But if it encounters a sentence in the text that looks like an instruction (“answer…”, “explain…”, “what is your purpose…”), it may mistakenly interpret it as an assignment for itself – and not as content to translate.
It's interesting that this is supposed to happen only in Advanced modewhich Google deployed for more natural, contextual translations. In the mode Classic Such states usually do not occur. So if you want to be sure that the Translator will only do its basic work, switching to Classic is the easiest solution so far.
Gallery
Google has not yet commented publicly on this. But for me, it is a good reminder of the risk that the more “agentic” and intelligent systems we put into common tools, the more we need to monitor them. reliable and predictable. Especially where we expect one clear result. In any case, be careful and check what and how you translate.
