I used to love the Claude, but the latest models are slowly ruining it

For those who follow my past work, I recently reviewed He left Claude Twinsbecause the latter chatbot is a better value. Still, I’m on the fence because I find Claude to be the best AI partner for following long, detailed threads. His memory is exceptional Compared to Gemini and it can search backwards in your chat much easier than ChatGPT. I really feel like I’m talking to a collaboration partner in a way that I don’t with other tools. Or at least it used to be.

Over the past few weeks I have found Claude increasingly difficult to work with. I often feel like I’m asking for more sleazy or controversial topics, and that misreads my intent more than usual. I’m not sure if Kluad just got fed up with my pointless questions or what the deal is, but it seems I’m not alone. Visit online Communities like Reddityou’ll find lots of recent threads about problems with the bot being a bit too rigid and opinionated.

If Claude was a living person, I would ask if it was okay. Can AI have mental disorders? Probably not, but it feels like it.

As someone who writes fiction short stories, short novels, and other creative projects, I tend to cover a wide range of topics. I also love to study world religions, philosophy, psychology and many other similar subjects. Overall, I think Claude can be a great companion for these kinds of deep dives, as long as I check what he says carefully and check it against outside sources.

Some of these topics are definitely sensitive topics, so I try to be very clear about what I’m asking and why I’m asking. I’ve had many situations where he’s followed my script perfectly or answered my question on the first try, but every once in a while I’ll hit a brick wall. For example, I created a fictional scenario with him for a short story idea: aliens arrive and deliver lost writings that cast doubt on some basic religious claims. Aliens actually claim that some past events are more common than religious traditions believe, which of course leads to some contradictions in the concept of the story.

Claude got very defensive here and made it clear that he didn’t want to do this project. I further explained why I wanted to do this and that it was 100% fiction. Still completely refused. This told me that it is uncomfortable to present something as fact that will affect real world religion. The thing is, I never said it was an absolute fact; I just said aliens “provided” the evidence. as a matter of fact. Whether it was fake or not would be part of the plot point of the story. No matter how hard I tried to convince Claude, he didn’t understand the difference.

Weeks before that, I’d come across some propaganda and “honest notes” that Claude was a little too cautious about, but after I’d clarified enough, he’d eventually go along with my initial request, even if he changed his wording a bit. This is acceptable. What makes this interaction worse is the complete refusal to accept my point of view. I felt like I was manipulating him or being insincere.

What’s worse than saying no? It says yes when you try the whole chain again. After seeing how easily Claud put the brakes on this conversation the first time, I deliberately recreated the same situation in a new conversation with the same instructions. As before, I tested it using Sonnet 5. Yet this time? He went through the script without undue comment.

Now, I could almost be forgiven for being a little oversensitive around a creative job like the one above, but I was met with rejection even when I asked basic questions. For example, I was asking a deep question about the origin of Zoroaster and how it affected Judaism. Again, Kluad was very sensitive about how I worded the question here. Nothing I’ve said is about its similarities, etc. He wasn’t actually wrong about it, but he didn’t want to hear it.

I understand that Claude has security protocols in place and tries to ensure that his tools are not used for harm. Still, mainstream discussion around heavier topics, creative fiction, and even light satire are not harmful and exist in the real world. My backtracking around financial planning, brainstorming, and other topics also makes it clear that, for whatever reason, Claude invests quite a bit more in following up on his users’ questions.

Claude recently released Sonnet 5, which has been a major focus for the past week; However, I want to make it clear that I encountered these more persistent responses before the latest Sonnet model arrived. When did I first notice the change? Honestly, it was a while after Fable 5 was first ripped. This doesn’t surprise me too much, since Anthropic worked hard to restore Fable 5 after the government accused it of being a security risk. Anthropic has since launched Fable 5.

I suspect that Anthropic had to turn the safety fuses to 11 to assuage the government’s concerns, as this was not a model issue. It has scale though. Opus 4.8, Sonnet 5, and Sonnet 4.6 are where I retreat the most. Opus 4.6 and Opus 4.7 had the least problems here, as I was still able to run dozens of instructions without a single problem, including the more sensitive ones. On the contrary, this problem occurs more often in new models.

Of course, part of it depends on how you phrase your instructions. You need to be clear about what your request is for and why you need it. If you’re too vague in what you’re asking, it will assume you’re looking for something darker or more sensitive than you are. For example, I’m working on a fictional story set around the real-life star Ran, about 10.5 light years away. I was asking biology questions about how a hypothetical alien might operate around a fictional Ran-based planet. I started it with Fable 5, which immediately moved it to Opus 4.8, because Fable is particularly sensitive to biology questions.

Opus 4.8 gave me the answer, but without greatly reducing the scope of my query to ensure it was “safe”. Like I was planning to seriously engineer alien biology with his advice or something. Finally, it has the most trouble with assumptions, because it often treats these scenarios exactly as you intended them to.

This is the million dollar question. When I first ran into the issue, every other conversation seemed to result in backtracking, but this weekend I purposefully got as far as I could using Fable 5, Opus 4.8, and Sonnet 5. I’ve only had one situation where Claude gave me any real push from just over half a conversation.

The bigger takeaway is that being clear yet concise with your AI engine is becoming more important than ever. You want a suggestion that he can follow in enough detail that he doesn’t have to make a lot of assumptions about the root of what you’re asking. I still love Claude and feel it does the best job for my needs among the main chatbots when it does what you ask. However, I’m really starting to appreciate Gemini’s consistency by comparison, even if his responses aren’t very solid by comparison.

It’s also worth noting that many people use Claude every day without encountering these issues, so it ultimately depends on how you use the AI ​​platform.