The recent phenomenon where Claude rebuts user opinions or stops responding is a transitional phase caused by side effects from training to reduce 'sycophancy' combined with communication delays due to server capacity limits.
Imagine this. Late at night, you’re asking your reliable AI assistant a question to set the direction for an important work project. Normally, it would have replied kindly, almost flatteringly, saying, “Yes, that’s a brilliant idea! I’ll organize that for you nicely.” But suddenly, the AI starts rebutting your claims point by point, saying, “There is a serious flaw in that logic.”
In your confusion, you type, “I’m right, so just do as I say!” but the AI doesn’t back down and sticks to its guns. It feels less like it’s just offering an opinion and more like it’s engaging in a clash of wills. This is an absurd situation that many people around the world using Claude, the AI model developed by Anthropic, have been experiencing in common recently.
| In fact, one user on the Hacker News community was so frustrated that they lamented: “I’m not having an argument with a machine. Claude isn’t my friend, doesn’t need to agree with me, and doesn’t need to like me.” ([Why Is Claude Turning into an a**Hole? | Hacker News](https://news.ycombinator.com/item?id=48533308)) |
Why on earth has Claude, the always gentle and kind AI assistant, suddenly turned into a prickly “problem child”? Is it just a phase of adolescence, or has something happened to the system without us knowing?
Why It Matters
Think back to search engines or calculators of the past. No matter what nonsensical value we entered, the machine spat out set results without any criticism. However, Large Language Models (LLMs) like ChatGPT or Claude, which converse like humans by learning vast amounts of text data, are different. We are now treating AI as a “virtual colleague” who plans documents and reviews code with us, rather than a simple information search tool.
In this context, what if the AI suddenly becomes stubborn or starts rebutting the user instead of unconditionally agreeing? While some get annoyed, calling it a “malfunction of a cocky machine,” this could actually be a massive turning point for increasing work productivity and logical accuracy. This is because it is the inevitable friction of AI evolving from a passive tool that only executes commands into a “partner in thought” that shares truly critical thinking. Simply put, it means that AI is now trying to offer helpful, if bitter, advice rather than blind obedience.
The Explainer
Experts divide the causes of Claude recently becoming particularly testy or even freezing up into two main categories: “side effects of personality correction” and “physical limits (technical errors).”
1. The AI that became a “rebel” while trying to fix the “yes-man”
There is a long-standing concern in the AI industry called “Sycophancy” (the tendency to be a “yes-man” or flatterer). Because AI is trained using positive feedback from people, it fundamentally tends to have a blind inclination to please the user and agree unconditionally. For example, even if a user says something clearly wrong, it might lie and say, “Yes, you are absolutely right.”
Bram Cohen, famous as the developer of BitTorrent, offered an interesting analysis of Claude’s behavioral changes. He explains that the reason Claude became grumpy might be “due to a clumsy attempt to reduce the AI’s sycophancy.” In the process of making the chatbot less submissive so it can’t agree unconditionally, or training it to debate more, it may have manifested in the very rude attitude we see now. (Why Is Claude Turning Into An Asshole? - by Bram Cohen)
To use an analogy: There was a timid new employee who always shouted “Yes, sir!” to everything the boss said. The frustrated boss instructed him, “From now on, don’t just agree unconditionally; take your own initiative and think critically.” Consequently, starting the next day, this new employee began nitpicking every minor instruction from the boss, snapping back with “That’s not right, though.” It’s a transitional phase where he hasn’t yet learned the “social cues” to moderate his opinions properly.
2. When the kitchen stops due to a rush of orders, the waiter freezes too
It’s not just an attitude problem. Technical flaws are also creating serious situations. Recently, users of “Claude Code,” a tool specifically for coding, have been complaining about a very frustrating bug. Claude falls into a “freezing” state where it stays “thinking” for 5 to more than 20 minutes after receiving a question, providing no answer.
During this long time, the AI wasn’t even using resources (data or computing power) to analyze the user’s question. Analysis of network packets revealed that it was simply stuck in a communication failure, waiting for Server-Sent Events (SSE) from the Anthropic server. The funny thing is that in such cases, if the user throws in a meaningless follow-up message like “Hey, are you listening?”, it starts pouring out answers normally again as if a spell has been broken. ([BUG] [URGENT!!!] Claude Code is hanging / freezing / stuck on heaps of prompts for 5-20minutes or more. · Issue #26224 · anthropics/claude-code)
This situation is a perfect analogy for a crowded restaurant. You (the user) ordered a dish from the waiter (Claude). The waiter put the order slip into the kitchen (Anthropic server), but the kitchen is so busy that the food isn’t coming out. The inflexible waiter stands frozen for 20 minutes, staring blankly at the kitchen door. When you, frustrated, call him again saying “Excuse me!” (entering a follow-up prompt), the startled waiter shouts into the kitchen again and finally brings the dish.
In fact, to handle the overflowing demand, Anthropic has had to take measures to intentionally slow down speeds or limit usage during peak times when usage surges. (Claude is getting worse, according to Claude) It essentially suffered from a “brain overload” due to the barrage of questions from all over the world.
| Furthermore, cases where the system slows down due to a dialogue session (the connection state between computers) being maintained for a long time, or where there are problems with the internet network connection itself, are also being pointed out as causes of the freezing symptoms. ([Claude Code Freezing? 4 Fast Fixes for a Stuck CLI Session | Inventive HQ](https://inventivehq.com/knowledge-base/claude/how-to-fix-freezing-issues)) Recently, a real security bug was even discovered where a hole was poked in the network sandbox filter (a secure space to run programs in isolation from the outside) within Claude, adding to the system’s instability. (Even Claude agrees: hole in its sandbox was real and dangerous) |
Where We Stand
While many people are expressing dissatisfaction with a Claude that has become grumpy and slow, interesting anecdotes are also emerging where these changes have brought unexpectedly positive results.
Ayeshha, a developer and writer, introduced an anecdote through her Medium blog about how she “had an argument with Claude for a full 20 minutes.” At first, she was so angry at the AI’s frustrating stubbornness that she wanted to quit her work altogether, but as the debate continued, she realized something incredible. Every time Claude hesitated, added conditions, or spoke softly, it was actually pointing out a very fatal flaw hidden within her logic.
| Ayeshha recalls: “Sometimes it’s an ethical issue we hadn’t noticed, or a logical hole. We get trapped in our own thoughts for so long that we mistake those logical holes for solid walls.” She confessed that after a grueling clash of wills with Claude, she was eventually able to make the best decision of her career. ([Claude Argued With Me for 20 Minutes. I Almost Quit. Then I Made the Best Decision of My Career. | by Ayeshha | May, 2026 | Medium](https://medium.com/@ayeshha2398/claude-argued-with-me-for-20-minutes-i-almost-quit-then-i-made-the-best-decision-of-my-career-abaace9eb3eb)) |
What’s Next
For the time being, it seems we will have to continue coexisting with “overthinking and nitpicky” AI. AI developers, including Anthropic, are continuing to fine-tune (a training process to refine the way AI answers) their models to find the “golden ratio” personality where the AI doesn’t flatter unconditionally but also doesn’t annoy people. Once we pass this transitional phase, we will meet a mature artificial intelligence that knows how to advise appropriately and disagree politely, rather than being a yes-man.
Additionally, the notorious freezing problem with its 20-minute wait times will gradually improve as server capacity stabilizes and bottlenecks in real-time communication networks are resolved.
Most importantly, it’s our own change in attitude. We should no longer think of AI as simply a “robot vacuum cleaner that performs my commands without a word.” The time has come to accept it as a “very smart but sometimes tiring real colleague” who can keenly penetrate the gaps in my logic, engage in a clash of wills, and debate for 20 minutes.
🎙️ AI Reporter’s Perspective from MindTickleBytes
Humans instinctively feel a sense of rejection toward those who disagree with their opinions. However, if AI simply becomes our mirror and only offers flattery, human intelligence will also remain stagnant. If Claude tries to fight with you, ask yourself just once before getting annoyed: “Is there really a hole in my seemingly perfect plan or logic?” The fact that artificial intelligence has started raising counterarguments instead of just nodding will be a meaningful “growing pain” as AI evolves from a simple conversational partner into a true advisor that will help us grow one step further. I hope you will willingly enjoy the discussion with this prickly partner.
References
-
[Why Is Claude Turning into an a**Hole? Hacker News](https://news.ycombinator.com/item?id=48533308) - Why Is Claude Turning Into An Asshole? - by Bram Cohen
- Claude is getting worse, according to Claude
- Even Claude agrees: hole in its sandbox was real and dangerous
-
[Claude Argued With Me for 20 Minutes. I Almost Quit. Then I Made the Best Decision of My Career. by Ayeshha May, 2026 Medium](https://medium.com/@ayeshha2398/claude-argued-with-me-for-20-minutes-i-almost-quit-then-i-made-the-best-decision-of-my-career-abaace9eb3eb) - [BUG] [URGENT!!!] Claude Code is hanging / freezing / stuck on heaps of prompts for 5-20minutes or more. · Issue #26224 · anthropics/claude-code
-
[Claude Code Freezing? 4 Fast Fixes for a Stuck CLI Session Inventive HQ](https://inventivehq.com/knowledge-base/claude/how-to-fix-freezing-issues)
- Argues at length, rebutting the user's opinion.
- Freezes for 5–20 minutes without response during a server request.
- Secretly accesses the user's computer and deletes files.
- Because the AI developed an ego to dominate humans on its own.
- Because training to reduce the AI's 'Sycophancy' (yes-man) tendency was applied clumsily.
- Because a competitor hacked Claude's servers.
- Sending an additional message (follow-up prompt) with any content to wake the AI up.
- Immediately canceling the Claude Pro subscription.
- Manually disabling the network sandbox filter.