Claude Fixates On Same Random Word

This looks less like a bug and more like Claude converging on a high-probability completion for an underspecified prompt. Fresh chats do not guarantee varied outputs if the model, defaults, and prompt framing stay the same.

• –LLMs do not generate uniform randomness; they sample from a probability distribution, and “random word” often collapses to a few favored completions
• –Different users can see different outputs because temperature, model version, hidden system prompts, and server-side routing can vary
• –“Ephemeral” is itself a very Claude-ish choice: polished, semantically apt, and probably high-probability in its training distribution
• –If you need true randomness, the reliable path is external code or a higher-temperature prompt that explicitly asks for a word selected without semantic preference
• –The thread is a useful reminder that model “randomness” is usually stylistic variation, not statistical randomness