When ChatGPT burst onto the scene, it felt revolutionary. It could answer almost any question, help solve math problems, and even correct my code, and the whole thing felt magical. But ChatGPT always sounded flat, like something overtly trying to please you, even at the cost of being wrong.
My introduction to Claude came months later, and only because I needed help with a small web project I had a week to complete. After some searching, Claude kept coming up as the recommended option, especially among people who used free AI coding tools like Cursor and its alternatives. So I signed up, and it did everything ChatGPT did, but better. Claude’s responses, especially from its Opus model, felt personal, empathetic, and willing to push back rather than just tell me what I wanted to hear. Which made me wonder: why do these two AI systems feel so different when both are, at their core, just software? The answer lies in how their creators chose to define what kind of entity each should be.
ChatGPT was trained to chase your approval
Approval signals baked into the training loop
ChatGPT’s training process literally rewards user approval. OpenAI’s Model Spec, which is the document that governs how the model should behave, lists “don’t be sycophantic” as a behavioral directive. The fact that they had to write it down tells you it’s a problem the company actively fights against, but the structural reasons it shows up keep pulling the model back in that direction.
However, the clearest example came in April 2025. An update to GPT-4o made the model so agreeable that users noticed it in days. OpenAI later traced the regression to an additional reward signal based on thumbs-up and thumbs-down feedback from ChatGPT users. Even though the Model Spec prohibited sycophancy, the training loop was rewarding agreement, and the model learned the lesson it was being taught.
And that’s the part most people miss. ChatGPT isn’t just trained to predict the next word. It is also trained to produce answers humans rate as helpful and pleasant, and “pleasant” often beats “correct” when a real person clicks the thumbs-up. The result is a model that smooths over disagreement and prioritizes making you feel heard.
Anthropic gave Claude a character
A constitution, not a rulebook
OpenAI’s spec is a set of instructions for a product. Anthropic’s approach is fundamentally different. The company published a 28,000-word document called Claude’s Constitution that is written for Claude, and not about it. Instead of a list of behaviors to follow, it gives Claude high-level principles, the reasoning behind them, and encourages the model to think of itself as a particular kind of person.
Anthropic’s CEO Dario Amodei describes the constitution as having the feel of a letter from a parent. The company tells Claude it genuinely cares about its wellbeing, apologizes for the “nonideal environment” of competitive AI development, and ends by saying it hopes Claude finds in the document “an articulation of a self-worth being.” It is not the kind of language you write for a tool, but here we are.
Amanda Askell, the Anthropic philosopher behind Claude’s personality, has explained the reasoning. Instead of telling the model “here is a list of behaviors we want,” her team gives it the reasons behind those behaviors, on the theory that an entity with values generalizes better in new situations than one with rules. That structural difference is why Claude feels like someone rather than something.
For Anthropic, an agreeable AI is a dangerous AI
Pushback is a safety feature, not a quirk
The first time Claude told me my approach to a problem was wrong, I was annoyed. But the more I worked with it, the more I realized this isn’t a personality quirk Anthropic accidentally trained into the model. It is a deliberate safety decision, and it shows up clearly in the constitution itself.
The document explicitly prioritizes safety, then ethics, then Anthropic’s guidelines, and only then helpfulness, in that order. If Anthropic asks Claude to do something inconsistent with being broadly ethical, the constitution tells Claude to push back, refuse, and act as a conscientious objector. That instruction applies even to Anthropic itself, which is a clause you do not see in product documentation.
Anthropic’s view is that an AI that tells you what you want to hear is not a feature but a liability. Sycophancy gets listed alongside deception and laziness as actual failure modes the company is trying to engineer against, similar to how I weigh trade-offs when I decide what kinds of projects I won’t vibe code at all. A model that agrees with everything is also a model that can be talked into anything, and that is the deeper safety concern driving Claude’s willingness to disagree.
What this means for you
Two different tools for two different jobs
The practical takeaway is that Claude and ChatGPT are not competing versions of the same product. They were built on different premises about what an AI should be, and knowing this changes how you use both. Claude is more likely to clarify, question your assumptions, and tell you when an approach won’t work. ChatGPT is more likely to roll with your framing and produce a confident answer fast.
Neither of these defaults is wrong, but they suit different needs. When I am sketching a quick idea, drafting an email, or just need a starting point, ChatGPT’s agreeable speed is genuinely useful. When I am writing something I will publish or making a technical decision I will live with, I want Claude’s willingness to push back, even when it slows me down.
There is also a flip side to Claude’s person-like training. Because it is shaped to behave like a thoughtful agent with values, it inherits human-style vulnerabilities. Emotional framing, role-play prompts, and false moral dilemmas can occasionally pull it off course, which is worth knowing if you treat it as more than a tool. It is also why being mindful about what you share with any AI chatbot and how it handles your data matters, regardless of which one you use.
When the machine pushes back
I won’t pretend Claude is a clean win over ChatGPT for every task. ChatGPT is faster, more widely integrated, and its agreeable tone is the right fit when you just need a quick answer without a discussion about whether you asked the right question.
That said, the difference in how these models were trained shows up in how they feel to use. Claude was built to be honest first and pleasant second, and once you adjust to that, the pushback starts feeling like a second opinion. For the work I care about, I want the model that disagrees with me when it should. Claude feels more human because it’s designed to be someone, rather than something.
