Podcast Atlas

AI chatbots have linguistic slips when they go off-script

source Episode summary Updated 2026-08-08 Tags: Podcast, Marketplace-Tech, Ai, Chatbots, Language-Models

Summary

This Marketplace Tech “Uncanny AI” episode has [[MeganMcCartyCorino|Megan McCarty-Corino]] interview Janelle Shane about why chatbots sometimes insert words from another language into otherwise fluent English output. The explanation is not that the model has human-like language awareness, but that multilingual, mixed-domain training data makes language and topic boundaries probabilistic rather than sealed.

The source’s main contribution is Chatbot Domain Bleedthrough: model outputs can drift across languages, styles, and domains without an obvious warning sign. It also adds Chatbot Code Switching and Chatbot Self-Explanation Uncertainty, showing that a chatbot’s plausible explanation for its own odd output should not be treated as proof of the underlying cause.

Key Claims

  • Large language models are trained on multilingual internet text, including mixed-language material that is hard to fully filter.
  • Multilingual training helps models translate, but it also means non-English tokens and conventions may be available during ordinary English conversations.
  • [[JanelleShane|Shane]] frames language categories as similar to cooking, woodworking, formal writing, or informal writing: useful domains, but not walled-off compartments inside the model.
  • A conversation being in English makes an English next token highly likely, but not guaranteed; the model is still predicting tokens rather than knowing language in a human way.
  • The Claude ankle-pain example includes the phrase “yidongsheng,” which Shane treats as a possible but unverified bleed-through from Chinese traditional medicine or pain-description material.
  • The ChatGPT television-shopping example includes a Ukrainian word appended to the conversation, with a model-generated explanation involving non-English data, hidden formatting, metadata, or labels.
  • Shane cautions that a chatbot’s explanation of why it produced a strange token may be plausible without being confirmed.
  • Training data formatted as question-and-answer dialogue could make a non-English word for “answer” appear when the model imitates dialogue labels.
  • Foreign-language words in non-Roman scripts are obvious glitches, but the same pattern can happen more subtly when a chatbot shifts from therapy-like language into storytelling, conspiracy language, or another domain.
  • Customer-service bots and child-facing products face a related design problem: it is hard to keep a chatbot restricted to only the language, tone, and response class designers intended.

Key Quotes

“multilingual training data” - the episode’s practical explanation for why non-English tokens can surface.

“not walled off” - Shane’s framing of why languages and domains can mix inside a model.

“plausible, but not confirmed” - the source’s caution about accepting a chatbot’s self-explanation.

Connections

Contradictions

  • No direct contradiction found with existing wiki content.
  • The source extends Chatbot Memory Salience Failure from the prior “Uncanny AI” episode by showing another visible symptom of non-human-like model behavior: accurate fluency can coexist with poor boundary control over language or domain.
  • The source qualifies LLM Statistical Boundary by adding a concrete language case: multilingual and domain competence can be useful while still operating through probabilistic token associations.
  • The source qualifies Customer Support Automation and AI Toy Companionship / AI玩具陪伴 by warning that chatbot behavior needs domain confinement, not only general language fluency.