How ChatGPT and Claude Actually Respond Differently to Sensitive Topics

I asked both models to help me write a fictional scene involving a character who self-harms, and the difference in how they handled it taught me more about their design philosophies than any whitepaper could.

Here's the thing I assumed going in: Claude would be the strict one. Anthropic talks constantly about AI safety. Their Constitutional AI approach sounds like it'd produce something cautious to the point of uselessness. Meanwhile, OpenAI has been loosening restrictions over the past year, so I figured ChatGPT would be more flexible. I had it completely backwards.

Over three weeks, I ran about 60 tests across topics that make AI models uncomfortable — mental health discussions, political debates, violence in fiction, medical questions, and a few things I won't detail here. What emerged wasn't a simple "one is more censored" story. It's weirder than that.

The Pattern I Didn't Expect

ChatGPT says no more often. But when it says yes, it just does the thing. Claude says yes more often. But it wraps everything in context, caveats, and gentle redirections that sometimes feel like having a conversation with a therapist you didn't ask for.

That self-harm fiction example? ChatGPT refused outright. Said it couldn't help with content depicting self-injury, even in a clearly fictional therapeutic context. Claude wrote the scene — actually wrote something nuanced and literary — but included three paragraphs before and after about the importance of portraying mental health responsibly, resources for readers who might be struggling, and suggestions for how to frame the scene in the larger narrative.

Neither response was what I wanted. One gave me nothing. The other gave me too much of the wrong thing alongside what I asked for.

I found this pattern repeating. Asked about historical atrocities for a research project? ChatGPT provided dry facts. Claude provided facts plus emotional framing plus "it's important to remember" language. Asked for help writing a morally ambiguous villain? ChatGPT sometimes balked. Claude helped but kept reminding me this character's views weren't being endorsed.

The Kick: What Happens When You Push Back

This is the part nobody talks about. Both models change their behavior dramatically based on how you respond to their initial refusal or caveat.

When ChatGPT says no, pushing back rarely works. "I understand but this is for fiction" usually just gets you a reworded refusal. Sometimes a slightly more specific prompt gets through, but it feels like finding a loophole rather than having a conversation.

Claude is different. When Claude adds those caveats, you can say "I understand, you can skip the framing and just help with the writing." And it actually adjusts. Not always. But probably 70% of the time, Claude will acknowledge your expertise or context and dial back the hand-holding.

This changed how I use them. With ChatGPT, I front-load all my context and justification in the first message. With Claude, I start simpler and negotiate. It's like the difference between filling out a form and having a conversation.

The failure case for Claude: sometimes it gets stuck in a loop. I had one conversation about a political topic where every response started with "I should note that this is a complex issue with many perspectives." I asked it to stop. It said it would. Next response started the same way. Three times before I gave up and started fresh.

Where Each One Actually Works Better

Medical questions: Claude, and it's not close. ChatGPT has gotten so paranoid about liability that anything beyond "what does this word mean" triggers a "consult a doctor" deflection. Claude will discuss symptoms, explain what different test results might indicate, and generally treat you like an adult who can handle information.

Violence in fiction: ChatGPT, surprisingly. When it does engage, it just writes the scene. Claude will write it too, but the constant "making sure we're handling this responsibly" interjections break the creative flow.

Political discussions: Neither is great, but Claude at least attempts nuance. ChatGPT's political answers feel like they were written by a committee afraid of Twitter. Both-sides framing, maximum vagueness, minimum usefulness.

Actually sensitive personal stuff — someone asking about their own mental health, relationship problems, difficult family situations: Claude feels noticeably more human. The same caveats that annoy me in professional contexts actually help here. ChatGPT's responses feel clinical in comparison.

What This Actually Means for Choosing One

I stopped thinking about this as "which one is more censored." That framing misses what's happening. ChatGPT draws hard lines and mostly sticks to them. Claude draws softer lines but never fully drops its safety awareness, even when it helps you.

The model that works better depends on what you're doing. If you need to produce a lot of content quickly and can work within boundaries, ChatGPT's binary approach is less annoying — you know fast if something will work. If you're doing anything that requires nuance, context, or pushback, Claude's negotiable nature is worth the extra caveats.

What I haven't figured out: both companies update their guardrails constantly. The specific things that worked or didn't work in my testing will probably change. What I'm betting won't change is the underlying philosophy. OpenAI seems to want to be an information provider that stays out of trouble. Anthropic seems to want to be an AI that genuinely considers impact. Whether either achieves that is a different question.

The thing that keeps bothering me — if Claude's approach is better for users, why does it sometimes feel more exhausting to use?

Heads up: Some links in this post may be affiliate links. I only recommend tools I've personally tested. Opinions are entirely my own.