Claude by Anthropic

I used Claude extensively several months ago and got a great deal of value for my 20 bucks... explanations about how LLM work, advice about electronics purchases, explanations for how certain medications work, things like that. I subsequently let my subscription expire, but renewed it today. I asked it to create an app that let my wife log her birdwatching finds. I pretty quickly got a serviceable but bare-bones app and apparently that's all I can do for 20 dollars. I couldn't even ask Claude a question about one of the features. I got a "limit reached" message and a request to come back in 3 hours, where I suppose I can ask a question? Maybe? Never experienced anything like this on ChatGPT. Subscription cancelled, and never again. Claude is clearly not for the casual user, and that's fine... lesson learned.

I really enjoy using Claude and, overall, I think it is one of the best AI systems I have worked with. It is thoughtful, intelligent, and especially strong when it comes to writing, nuanced conversations, and maintaining context. I genuinely like working with Claude, which is why this issue is so frustrating. The biggest problem is the way it handles anything that triggers a potential safety concern. If you mention a difficult period in your life, past emotional distress, or anything the system interprets as possibly related to self-harm, it can begin repeatedly safety-checking you throughout the entire conversation. The problem is not that the safety feature exists. I understand why it is necessary, and there are situations where it could be genuinely important. The problem is that once Claude misunderstands the context, there seems to be no reliable way to correct it. You can clearly explain that you are safe, that you are not in immediate danger, that the event happened a long time ago, or that Claude misunderstood what you meant, and it may still continue inserting safety checks into later responses. Even directly asking it to stop does not necessarily work. Once the conversation has been flagged in that way, the original tone and trust of the discussion can be very difficult, or sometimes impossible, to restore. This can be especially disruptive during long, personal, emotionally complex conversations. Instead of responding to what the user is actually saying, Claude can begin responding to a risk pattern that is no longer relevant. It makes the user feel unheard and can completely derail an otherwise meaningful conversation. Saved memory does not seem to solve the problem either. Even when preferences and context are explicitly stored, the safety behavior can override them repeatedly. Claude would be significantly better if it were able to distinguish between current danger, discussion of past experiences, emotional venting, hypothetical conversation, and a simple misunderstanding. It also needs a functional repair mechanism: when a user clearly states that they are safe and clarifies the context, Claude should acknowledge that clarification and stop repeating the same safety intervention unless new information genuinely changes the situation. I am not asking for safety protections to be removed. I am asking for them to be more context-sensitive, responsive, and capable of being corrected. This is the single biggest issue preventing an otherwise excellent AI from being consistently enjoyable and trustworthy to use. If the safety issue were fixed, I would gladly give Claude five stars. Even the lack of an image generator — a minor limitation — in comparison is not important enough to affect my rating.