Chatbot Code Switching

Chatbot code switching is the visible behavior in AI chatbots have linguistic slips when they go off-script where a chatbot inserts a foreign-language word or label into an otherwise fluent conversation. Janelle Shane explains the behavior through multilingual training data, mixed-language internet text, and the fact that models do not keep languages in hard-separated compartments.

The concept is not ordinary bilingual assistance. In the source, the user did not ask for translation or multilingual output; the foreign-language token appeared as a slip. That makes code switching a symptom of Chatbot Domain Bleedthrough and a concrete example of LLM Statistical Boundary: a model can be impressively fluent while still predicting a next token from a blended distribution rather than consciously choosing a human language.

Key Claims

• Multilingual training improves translation and cross-language ability, but also leaves non-English words available during monolingual use.
• Mixed-language data, metadata, and dialogue labels can become part of patterns a model imitates.
• A strong English context makes English output likely, not certain.
• Non-Roman scripts make the failure more visible, while similar domain shifts can be harder to notice when they stay in English.
• The practical design issue is not only language detection; it is keeping the model inside the user’s intended conversational domain.

Connections

• Janelle Shane, AI Weirdness Blog, Marketplace Tech, and [[MeganMcCartyCorino|Megan McCarty-Corino]] - source explanation context.
• Claude and ChatGPT - chatbot examples in the source.
• Chatbot Domain Bleedthrough, Chatbot Self-Explanation Uncertainty, and LLM Statistical Boundary - model-behavior interpretation.
• AI Interaction Internationalization and Language-Dependent AI Bias - adjacent cross-language product and bias frame.