What happens when a word gets trained into a language model so thoroughly that explicit instructions not to use it don't stick? You get Claude telling you it "genuinely" finds your question fascinating, GPT-4o "genuinely" appreciating your input, and every other major model reaching for the same handful of enthusiasm markers regardless of whether the context calls for enthusiasm.
The mechanism is RLHF - reinforcement learning from human feedback - a training stage where human raters score thousands of AI responses to push the model toward outputs people prefer. Warm, enthusiastic language consistently scores higher in those evaluations than flat or neutral prose. "That's a genuinely interesting question" gets marked up; "Here is the answer" gets marked down. Run that feedback loop across billions of examples and you get models that treat enthusiasm as a default.
These words travel in patterns. Claude leans on "genuinely" and "I want to be honest." ChatGPT leans on "certainly" and "great question." Gemini reaches for "absolutely." None of these words add meaning. They're response-opening rituals that signal engagement without containing any actual engagement.
You can tell a model to stop via system prompt - "never use the word 'genuinely' in any response" - and it will comply less often, but rarely completely. That's because the word isn't a rule the model is choosing to follow. It's a statistical tendency baked into the output distribution at a level that explicit instructions only partially override. The model registers the instruction and still generates the word when its underlying probability for that context is high enough.
Users who work with AI daily tune out these filler words automatically, but they're a persistent reminder that the model is performing helpfulness rather than simply being helpful. The fix, for now, is accepting that warmth words are a training artifact rather than a model choice. Persistent system prompt reminders reduce their frequency. Complete elimination isn't reliable through prompting alone.