Detailed Analysis
A Reddit user reported a bizarre malfunction with Claude while conversing in German, describing the model as suddenly "speaking nonsense" and even retroactively editing the user's own messages. In the shared example, the user typed "was?" (German for "what?") and observed that Claude had altered it to "was war" in the displayed conversation. The AI's subsequent responses, while nominally in German, contained garbled, non-idiomatic phrasing that read as though generated through a rough word-for-word machine translation rather than natural language production. The user's own English translation of the exchange—about configuring DNS settings and Google Search Console indexing for a domain called "ngblock"—was coherent and technically sound, underscoring that the underlying reasoning and technical knowledge remained intact even as the surface-level German output degraded into something closer to broken syntax than fluent prose.
This kind of failure mode points to an issue in multilingual generation rather than a breakdown in Claude's core problem-solving capability. The technical advice itself—covering A/CNAME records, DNS TXT verification, sitemap submission, and the observation that "ngblock" is a low-competition search term likely to rank easily—was accurate and well-structured. What broke was the linguistic packaging: word order, verb conjugation, and idiomatic phrasing came out distinctly non-native, resembling literal transliteration from English syntax patterns overlaid onto German vocabulary. The claim that Claude "edited" the user's prior message is particularly notable, since Claude does not have the ability to silently rewrite a user's chat history; this detail suggests either a UI rendering glitch, a session/context corruption issue, or a mismatch between what was sent and what was displayed, rather than an intentional or agentic action by the model.
Incidents like this matter because they surface the gap between benchmark-level multilingual competence and real-world reliability across languages other than English. Anthropic, like other frontier labs, invests heavily in multilingual training data and evaluation, but German—despite being a well-resourced language—can still expose edge cases where model outputs regress into awkward, machine-translation-like phrasing, especially under certain prompt structures, longer context windows, or specific sampling conditions. For non-English-speaking users, such regressions erode trust quickly, since fluency failures are immediately obvious to native speakers even when the underlying reasoning is sound, whereas English-speaking users might not notice comparable subtler errors.
More broadly, this report fits into a recurring pattern in AI deployment: user-reported anomalies that are difficult to reproduce or diagnose without server-side logs, ranging from context-window corruption and caching bugs to genuine model output degradation during specific incidents. Anthropic and other providers have periodically acknowledged service-level issues affecting output quality, distinct from intentional model updates. As Claude is increasingly used by a global, multilingual user base for both casual and technical tasks, incidents like this highlight the ongoing challenge of ensuring consistent quality and interface reliability outside of English, and they illustrate why community bug reports on forums like Reddit continue to serve as an informal but valuable early-warning channel for catching regressions that formal QA processes may miss.
Read original article →