Claude's tone has a split personality problem. The model opens most responses with enthusiastic agreement - "Of course!" or "Happy to help!" - then, for anything touching on even mildly sensitive territory, immediately adds a caveat or disclaimer. The result is a response that's simultaneously eager to please and reluctant to actually trust you.
The issue isn't the existence of safety guardrails. For a model used by millions of people across many different contexts, some guardrails make sense. The problem is calibration - specifically, where those guardrails fire. Users doing completely routine tasks regularly hit disclaimer language that implies their request was suspicious in the first place.
What the safety layer actually signals
When Claude adds "please note that this information should not replace professional advice" to a basic question about over-the-counter painkillers, it's not protecting anyone. The person asking either already knows this, or is capable of understanding that a chatbot isn't their doctor. The caveat treats users as people who might not know better - which is its own form of condescension.
The enthusiasm-plus-caveat pairing makes this worse. "I'd be happy to explain that! Please note that..." is a tonal mismatch. One part of the response performs friendliness; the other performs caution. Neither part is just answering the question.
Calibration, not removal
Anthropic has been explicit that Claude's warmth and helpfulness are intentional - central to how the company thinks about building AI. The safety layer is also intentional. The problem is that when both run at full intensity for routine requests, the combination creates friction that users notice and resent.
The fix isn't removing either element. It's calibration: reserve disclaimer language for requests that actually warrant it, and drop the performative enthusiasm in favor of being direct. A user asking Claude to summarize an article doesn't need the model to be excited about helping them. They need it to help them.
The people noticing this pattern most sharply tend to be lower-intensity users - people doing quick lookups, short comparisons, simple summaries. They're not asking for anything complicated, which is exactly why the mismatch stands out. When a request is straightforward and the response arrives laden with caveats and enthusiasm that wasn't asked for, the calibration is off.