For those who have followed my past work, I’ve recently considered leaving Claude for Gemini, as the latter chatbot is a much better value. Still, I’ve remained on the fence largely because I find that Claude is the best AI partner for following long, detailed threads. Its memory is exceptional compared to Gemini, and it can search backward through your chat more readily than ChatGPT. I really feel like I’m talking to a collaborative partner in a way I don’t with other tools. Or at least, I used to.

Over the last few weeks, I’ve found that Claude is increasingly difficult to work with. Often, it seems to believe I’m asking for more nefarious or controversial topics than I am, and it often misreads my intent more than ever before. I’m not sure if Cluade has just gotten sick of my pointless questions or what the deal is, but I don’t seem to be alone. Visiting online communities like Reddit, you’ll find plenty of recent threads about issues with the bot becoming a bit too pushy and opinionated.

If Claude were a living person, I’d ask if it was okay. Can AI have mental breakdowns? Probably not, but it sure feels like it.

As someone who writes fictional short stories, short novels, and other creative projects, I tend to cover a vast range of topics. I also love learning about world religions, philosophy, psychology, and many other similar topics. Generally, I find that Claude can be a great companion for these kinds of deep dives as long as I carefully vet what it says and check it against outside sources.

Some of these topics are admittedly sensitive subjects, so I try to be very clear about what I’m asking and my reasons for asking. I’ve had many situations where it completely follows my scenario or answers my question on the first try, but every once in a while, I’ll hit a brick wall.  For example, I was vibing out a fictional scenario with it for a short story idea: aliens arrive and deliver lost written works that cast doubt on some mainstream religious claims. The aliens effectively claim that some past events were more ordinary than religious traditions believe, which, of course, leads to some conflict in the story concept.

Claude got very defensive here and made it clear it didn’t want to do this project. I explained further why I wanted to do this and that this was 100% for fiction. It still refused completely. It told me that it was uncomfortable presenting something that would impact a real-world religion as if it were fact. The thing, though, is that I never said it was an absolute fact; I just said that aliens “presented” evidence as fact. Whether it’s forged or not would be part of the story’s plot point. It didn’t seem to understand the difference, no matter how many times I tried to convince Claud otherwise.

Weeks before this, I would occasionally run into some preaching and “honest notes” where Claude was being a bit too cautious, but after I gave it more than enough clarity, it would eventually run with my original request, even if it slightly modified how it was worded. That’s acceptable. What makes this interaction stand out as worse is the complete refusal to acknowledge my point of view. It seemed to think I was manipulating it or being disingenuous.

What’s worse than being told no? Being told yes when you try the whole chain all over again. After seeing how easily Cluade put the brakes on this conversation the first time, I purposely recreated the same situation with the same prompts in a new conversation. Just as before, I tested it using Sonnet 5. Yet this time? It ran through the scenario without unwarranted comment.

Now I could almost forgive it for getting a bit too sensitive around a creative work like the one above, but I’ve run into similar refusals even when asking basic questions. For example, I was asking a deep question about the origins of Zoroaster and how this influenced Judaism. Again, Cluade got very sensitive here about how I worded the question. Nothing I was saying was factually wrong about its similarities, etc, but it just didn’t want to hear it.

I understand that Claude has safety protocols and is trying to ensure its tools aren’t used for harm. Still, basic debate, creative fiction, and even light satire around heavier subjects are not harmful and exist in the real world. The fact that I’ve had pushback around financial planning, brainstorming, and other topics also makes it clear that, for whatever reason, Claude is a bit too invested in policing its users’ questions.

Recently, Claude released Sonnet 5, and this was a major focus for the last week or so; that said, I want to make it clear I’d encountered these more resistant responses even before the latest Sonnet model arrived. When did I first notice the change? Honestly, it was some time after Fable 5 was first taken down. This doesn’t surprise me too much, as Anthropic had been working hard to get Fable 5 restored after the government accused it of being a security risk. Since then, Anthropic has gotten Fable 5 up and running.

I suspect Anthropic had to turn up its safety guardrails to an 11 to assuage the government’s concerns, as this hasn’t been a one-model problem. There does seem to be a scale, though. Opus 4.8, Sonnet 5, and Sonnet 4.6 seem to be where I get the most pushback. Opus 4.6 and Opus 4.7 have had the fewest problems here, as I’ve still been able to run dozens of prompts without a single problem, including on more sensitive subjects. By contrast, the newer models seem to have this problem much more frequently.

Of course, part of this is down to how you word your prompts. You need to be clear about what your request is for and why you need it. If you are too vague in what you’re asking, it’ll assume you are looking for something darker or more sensitive than you are. For example, I am working on a fictional story set around Ran, a real-life star about 10.5 light-years away. I was asking biology questions about how a hypothetical alien around a fictional Ran-based planet might work. I started this with Fable 5, which immediately switched it to Opus 4.8, as Fable is particularly sensitive about biology-related questions.

Opus 4.8 did answer me, but not without greatly reducing the scope of my request to ensure it was “safe”. As if I were planning to seriously engineer alien biology with its advice or something. Ultimately, it has the most trouble with hypotheticals, as it often treats these scenarios as if you are actually planning to do them.

That’s the million-dollar question. When I first ran into the issue, it seemed like every other conversation resulted in pushback, but this weekend, I purposely pushed as far as I could using Fable 5, Opus 4.8, and Sonnet 5 primarily. I only ran into one situation where Claude gave me any real pushback out of about half a dozen chats in total.

The bigger takeaway is that it’s becoming more important than ever to be clear yet concise with your AI engine. You want a prompt that it can follow with just enough detail, it doesn’t have to make too many assumptions about the root of why you are asking it something. I still like Claude and feel that, when it does what you ask, it does the best job for my needs among the major chatbots. That said, I’m starting to really appreciate Gemini’s consistency by comparison, even if its answers are often not quite as solid by comparison.

It should also be noted that many people use Claude every day without encountering these issues, so it ultimately comes down to how you use the AI platform.