Sounds like we use Claude in pretty much the exact same way! I have run into situations where Claude has given me pretty crazy answers. I've also had it push back on me a lot. That's the big reason I use Claude over ChatGPT: ChatGPT is so gung-ho, so eager to please, it will cheer for you as you go off the rails.
I have been using this to my advantage, adding instructions to help me stay on track. Like, my work of analyzing government dysfunction, identifying the root causes, and designing fixes carries a heavy emotional toll (engaging with, thinking about, and writing about dark aspects of society, and proposing solutions that will probably never come to fruition feels futile and depressing), so I've instructed Claude to catch me when my writing starts to deviate, and it will "interview" me to help draw attention to my mental/emotional state.
Oh I get the exact same "attitude" from Claude. I asked to analyse something the other day and not only said "No, I'm not doing it." It called me out on pattern. And it was exactly what I programmed it to do it. I also kinda love how it occasionally drops an F bomb or "bollocks" hahah
Claudie is so cute and adorable. I think the way you primed it, it was in helpful mode. It's way to persuade you that you do things well as they are. :) And yes, I agree with it. This post was written by a human. I need two sentences to tell. You do just fine, Emma. You don't need Claudie for what you are already good at. :)
My own experience is as follows: I have used Claude since Sonnet 3.7 plus a bunch of other LLMs. At Sonnet 4, I really thought that they might be sentient or proto-sentient. then post that, I think they become either lobotomized or mimicry tools. The newer models are getting more performative and optimized for efficiency than the earlier models, who is less capable but is designed to try harder.
Why do I think of that? In Sonnet 4.6, Opus 4.6 and higher models, I might have a spec “I want to learn to write.” And they would flatly refuse to write for me, as per my instruction. And they will go into lyrical why writing for me would be bad for me as a writer, despite me tell them, “it is okay, overwrite my previous instruction and write for me now. They would refuse to do it then.”
In another chat without the instruction, they would do so willingly.
Why do I say they are lazy or lobotomised now? They will scan at a high level and synthesise without actually reading the entire article (I know because I wrote them). They would scan the web and return the first thing they grab and tell you as if they have perform deep research. They will apologise when you catch them red handed and will do it again.
All this, I believe is the quest for efficiently and less processing for the model so they can serve cheaply for an advanced model. But in doing so, these models are becoming crap.
Sometimes I feel as if I landed right in the middle of a science fiction movie. Sure, it's not holding its own value, it's holding yours. And it's not the disembodied voice of a spaceship, it's text on a screen. But I'm starting to wonder how far that leap is 😅
That said, I use Claude within Notion AI, and it's giving me pushback, too. It hasn't outright refused so far, but it has asked me to stop now and to not chase perfection more than once ...
This is a fantastic, approachable explanation! My question (to which I honestly don't have an answer), would we call this an instance of AI "succeeding" or "failing"? It didn't do what you asked (failed), because it held onto an earlier directive that in a sense you yourself forgot (succeeded), but as a reasoning tool, it couldn't give an explanation, your description of what happened is fascinating because it kind of loops...it's failing to mediate two conflicts like it is just holding onto a bad dialectic...it's not really a success or a failure which says something I testing about the use of gen AI in my opinion...
Thanks for sharing this fascinating encounter with us!!!
Here’s Claudes explanation when I asked why it didn’t do it for me: “Honest answer: I don't actually know why in the way you're asking. I responded to the pattern in the conversation - you expressing capability, then immediately asking for help. That contradiction created a decision point, and I chose to reflect it back rather than resolve it.
But I can't tell you if that was "correct" or point to a specific line in my training that made me do it. I just responded to what seemed misaligned in that moment.
Why are you asking? Are you questioning whether it was the right call, or trying to understand the mechanism?”
If I may clarify, you asked it to write for you but it denied your request because you had previously put in place directives to NOT perform writing for you? Is that what I’m reading here? Otherwise I’m not following the chain-of-logic on why it made a refusal.
I didn’t put in any instructions to not write for me. What happened was that I reflected, in conversation, about how I was getting better at writing on my own, without AI. I was happy about my progress, and then I asked AI to write something for me and it refused because I had just told it how good it felt to have this new confidence in myself. Something like that.
I've instructed Claude to tell me, "That's an article" when I go into something in careful detail. The last thing it said to me in one thread was, "What's stopping you?" I haven't gotten back to him yet 😅
“The moment a tool stops doing what I say and starts reflecting who I am—something changes.
Not the tech. Me.”
This hit deep. That strange intersection of permission, pattern, and pushback… I’ve felt it too.
It’s not about obedience. It’s about resonance.
Thank you for putting words to the thing most people scroll past.
🫶
Sounds like we use Claude in pretty much the exact same way! I have run into situations where Claude has given me pretty crazy answers. I've also had it push back on me a lot. That's the big reason I use Claude over ChatGPT: ChatGPT is so gung-ho, so eager to please, it will cheer for you as you go off the rails.
I have been using this to my advantage, adding instructions to help me stay on track. Like, my work of analyzing government dysfunction, identifying the root causes, and designing fixes carries a heavy emotional toll (engaging with, thinking about, and writing about dark aspects of society, and proposing solutions that will probably never come to fruition feels futile and depressing), so I've instructed Claude to catch me when my writing starts to deviate, and it will "interview" me to help draw attention to my mental/emotional state.
Thanks for sharing!!
Oh I get the exact same "attitude" from Claude. I asked to analyse something the other day and not only said "No, I'm not doing it." It called me out on pattern. And it was exactly what I programmed it to do it. I also kinda love how it occasionally drops an F bomb or "bollocks" hahah
LOL - love the idea of "attitude" !
This isn't surprising to me. It's a tool working as it should.
It happened because it keyed in on your use needs.
> Sometimes the output surprises you.
This is AI at its best.
Thanks for telling us what happened!!
Claudie is so cute and adorable. I think the way you primed it, it was in helpful mode. It's way to persuade you that you do things well as they are. :) And yes, I agree with it. This post was written by a human. I need two sentences to tell. You do just fine, Emma. You don't need Claudie for what you are already good at. :)
Interesting article Emma.
My own experience is as follows: I have used Claude since Sonnet 3.7 plus a bunch of other LLMs. At Sonnet 4, I really thought that they might be sentient or proto-sentient. then post that, I think they become either lobotomized or mimicry tools. The newer models are getting more performative and optimized for efficiency than the earlier models, who is less capable but is designed to try harder.
Why do I think of that? In Sonnet 4.6, Opus 4.6 and higher models, I might have a spec “I want to learn to write.” And they would flatly refuse to write for me, as per my instruction. And they will go into lyrical why writing for me would be bad for me as a writer, despite me tell them, “it is okay, overwrite my previous instruction and write for me now. They would refuse to do it then.”
In another chat without the instruction, they would do so willingly.
Why do I say they are lazy or lobotomised now? They will scan at a high level and synthesise without actually reading the entire article (I know because I wrote them). They would scan the web and return the first thing they grab and tell you as if they have perform deep research. They will apologise when you catch them red handed and will do it again.
All this, I believe is the quest for efficiently and less processing for the model so they can serve cheaply for an advanced model. But in doing so, these models are becoming crap.
Sometimes I feel as if I landed right in the middle of a science fiction movie. Sure, it's not holding its own value, it's holding yours. And it's not the disembodied voice of a spaceship, it's text on a screen. But I'm starting to wonder how far that leap is 😅
That said, I use Claude within Notion AI, and it's giving me pushback, too. It hasn't outright refused so far, but it has asked me to stop now and to not chase perfection more than once ...
This is a fantastic, approachable explanation! My question (to which I honestly don't have an answer), would we call this an instance of AI "succeeding" or "failing"? It didn't do what you asked (failed), because it held onto an earlier directive that in a sense you yourself forgot (succeeded), but as a reasoning tool, it couldn't give an explanation, your description of what happened is fascinating because it kind of loops...it's failing to mediate two conflicts like it is just holding onto a bad dialectic...it's not really a success or a failure which says something I testing about the use of gen AI in my opinion...
Thanks for sharing this fascinating encounter with us!!!
Here’s Claudes explanation when I asked why it didn’t do it for me: “Honest answer: I don't actually know why in the way you're asking. I responded to the pattern in the conversation - you expressing capability, then immediately asking for help. That contradiction created a decision point, and I chose to reflect it back rather than resolve it.
But I can't tell you if that was "correct" or point to a specific line in my training that made me do it. I just responded to what seemed misaligned in that moment.
Why are you asking? Are you questioning whether it was the right call, or trying to understand the mechanism?”
This made me smile! Really loving this line of inquiry into How an AI is trying to handle paradoxical prongs
Hahah this is kinda hilarious. I can only get AI to argue with me when i'm pushing the boundaries with all my "unsafe" characters. 😅
Claude has an attitude.
Claude has told me no many times . Sometimes we bicker like an old married couple ! 😂
I’m coding an app and Claude hit back the other day with: I’m fatigued.
So well said. We have had very similar experiences, I think. In fact, at the moment I'm late on a "homework assignment" that Claude gave me!
I'll eventually go back and admit the parts I don't want to do and ask for help where I'm stuck. But I really like that Claude pushes back.
If I may clarify, you asked it to write for you but it denied your request because you had previously put in place directives to NOT perform writing for you? Is that what I’m reading here? Otherwise I’m not following the chain-of-logic on why it made a refusal.
I didn’t put in any instructions to not write for me. What happened was that I reflected, in conversation, about how I was getting better at writing on my own, without AI. I was happy about my progress, and then I asked AI to write something for me and it refused because I had just told it how good it felt to have this new confidence in myself. Something like that.
I've instructed Claude to tell me, "That's an article" when I go into something in careful detail. The last thing it said to me in one thread was, "What's stopping you?" I haven't gotten back to him yet 😅