I loved Claude, but the latest models are slowly ruining him.

For those who have followed my previous work, I have recently considered leaving Claude for Geminisince this latest chatbot has a much better value. Still, I’ve remained hesitant largely because I find Claude to be the best AI partner for following long, detailed threads. His memory is exceptional. compared to Gemini, and you can search back through your chat more easily than ChatGPT. I really feel like I’m talking to a collaborative partner in a way that I don’t with other tools. Or at least I used to.

In recent weeks, I have found Claude increasingly difficult to work with. He often seems to think I’m calling for more nefarious or controversial topics than I really am, and he often misunderstands my intent more than ever. I’m not sure if Cluade has gotten tired of my useless questions or what the problem is, but it doesn’t seem like I’m alone. Visiting online communities like RedditYou will find many recent threads about problems with the bot becoming too aggressive and opinionated.

If Claude were a living person, I would ask him if he was okay. Can AI have mental breakdowns? Probably not, but it sure seems that way.

As someone who writes short fiction, novellas, and other creative projects, I tend to cover a wide range of topics. I also love learning about world religions, philosophy, psychology and many other similar topics. Overall, I find that Claude can be a great partner for these types of deep dives as long as you carefully analyze what he says and compare it to outside sources.

Admittedly, some of these topics are sensitive topics, 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 hit a brick wall. For example, I was imagining a fictional scenario with him for a short story idea: aliens arrive and deliver lost written works that cast doubt on some dominant religious claims. The aliens effectively claim that some past events were more common than religious traditions believe, which of course creates some conflict in the concept of the story.

Claude became very defensive and made it clear that he did not want to do this project. I explained more why I wanted to do this and that it was 100% for fiction. He still refused completely. He told me that it was uncomfortable to present something that would impact a real-world religion as fact. But the fact is that I never said It was an absolute fact; I just said that the aliens “presented” evidence. as fact. Whether it is faked or not, it would be part of the plot of the story. He didn’t seem to understand the difference, no matter how many times I tried to convince Claud otherwise.

Weeks before this, I would occasionally come across some sermons and “honest notes” where Claude was being overly cautious, but after giving him more than enough clarity, he would eventually agree with my original request, even if he slightly modified the way it was worded. That’s acceptable. What makes this interaction worse is the complete refusal to acknowledge my point of view. He seemed to think I was manipulating him or being fake.

What’s worse than being told no? Being told yes when you try the whole chain again. After seeing how easily Cluade stopped this conversation the first time, I deliberately recreated the same situation with the same prompts in a new conversation. As before, I tried it using Sonnet 5. But this time? He went over the scenario without unwarranted comments.

Now I could almost forgive you for becoming overly sensitive around creative work like the above, but I’ve encountered similar rejections even when asking basic questions. For example, I was asking a deep question about the origins of Zoroaster and how this influenced Judaism. Once again, Cluade got very sensitive about how I phrased the question. Nothing he was saying was really wrong about their similarities etc, but I just didn’t want to hear it.

I understand that Claude has safety protocols and is trying to ensure that his tools are not used to cause harm. Still, basic debate, creative fiction, and even light satire on heavier topics are not harmful and exist in the real world. The fact that he has had criticism around financial planning, brainstorming, and other topics also makes it clear that, for some reason, Claude is too interested in controlling his users’ questions.

Recently, Claude released Sonnet 5, and this was the main focus for the last week; That being said, I want to be clear that I found these responses more resilient even before the latest Sonnet model arrived. When did I first notice the change? Honestly, it was some time after Fable 5 was first removed. This doesn’t surprise me too much, as Anthropic had been working hard to restore Fable 5 after the government accused it of being a security risk. Anthropic has since launched Fable 5.

I suspect Anthropic had to turn its guardrails up to 11 to assuage the government’s concerns, as this has not been a one-model problem. However, there seems to be a scale. Opus 4.8, Sonnet 5, and Sonnet 4.6 seem to be where I get the most rejection. Opus 4.6 and Opus 4.7 have had the fewest issues here, as I’ve still been able to run dozens of messages without a single issue, even on more sensitive topics. On the contrary, newer models seem to have this problem much more frequently.

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

Opus 4.8 did respond to me, but not without greatly narrowing the scope of my request to ensure it was “safe.” Like he was planning to seriously design alien biology with his advice or something. Ultimately, you have more trouble with what-if scenarios, as you often treat these scenarios as if you were actually planning to do them.

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

The biggest takeaway is that it’s becoming more important than ever to be clear and concise with your AI engine. You want a message that you can follow in enough detail, that you don’t have to make too many assumptions about the root of why you’re asking something. I still like Claude and feel that when he does what you ask, he is the best job for my needs among the top chatbots. That said, I’m starting to really appreciate Gemini’s consistency in comparison, even if their responses are often not as strong in comparison.

It’s also worth noting that many people use Claude every day without encountering these problems, so it ultimately comes down to how the AI ​​platform is used.