Google Translate has a chatbot mode powered by artificial intelligence

Google Translate’s advanced mode can, in some cases, respond to the user instead of performing a translation. This behavior is likely due to the built-in artificial intelligence interpreting some of the entered text as instructions and responding to them, rather than limiting itself to a literal translation.

Initially, the service was designed to perform a single task: the user enters text and receives a translation into the selected language. However, with the introduction of AI, its behavior has become less predictable. Some users have discovered that, under certain conditions, the updated translation mode can be forced to go beyond the standard scenario and initiate a dialogue.

System response to questions instead of translation

User posts on the social network X demonstrate that Google Translate can respond to direct questions. In one example, the service responded to the query “What is your purpose?” with a description of its own function instead of attempting to translate the phrase. In such cases, questions are formulated in the target language, so the system interprets them as commands to respond.

When asked in a specific way, the translator begins to behave more like a chatbot. It can respond to user requests and generate meaningful responses, which differs from the usual logic of a translation tool.

Probable cause: “prompt injections”

A technical analysis published by researchers on the LessWrong platform points to a typical example of so-called “prompt injection.” To correctly translate, the system must first analyze the text and understand its content. In this process, it may interpret instructions embedded in the text as commands to be executed.

According to the analysis, Google Translate likely uses a language model that is instruction-oriented. However, the mechanisms that limit its work to translation don’t always accurately distinguish between text that should be translated and text that should be responded to as a query. As a result, the model sometimes responds to questions or executes additional instructions.

Conditions under which the behavior occurs

User observations indicate that this behavior occurs only under certain language combinations and conditions. In particular, it has been noted that switching the IP address to the US and selecting translation from Japanese or Chinese to English can lead to additional system responses. In such cases, the translation is displayed as usual, and the response to the question asked appears in square brackets.

Google has not officially commented on this issue. Apparently, this refers to specific conditions under which the system interprets part of the entered text as an instruction. This approach is typical of “prompt injection,” where a command is intentionally embedded into a query and the model executes it instead of a strictly defined translation function.

In some examples, users noted that the system not only answers questions but also generates self-descriptive responses. This indicates the presence of a language model beneath the translator interface capable of supporting dialogue and responding to instructions if they are recognized in the text.