Constitutional AI vs. Reinforcement Learning: Why Claude and ChatGPT Aren't Interchangeable

Non-Security Post

If you've been using ChatGPT and recently tried Claude, you might have found the experience underwhelming. That's not because Claude is worse — it's because you're probably using it wrong.

Nate Jones has an excellent video explaining why Claude and ChatGPT require different approaches, and the core insight comes down to how they were trained. Understanding this difference will make you significantly more effective with both tools.

The Training Difference

ChatGPT uses Reinforcement Learning from Human Feedback (RLHF). Human raters evaluate outputs and the model optimizes for responses that get positive ratings. This tends to produce agreeable, user-pleasing responses — the model learns that making users happy gets rewarded.

Claude uses Constitutional AI. Instead of optimizing for user satisfaction, Claude is trained against explicit principles: be honest, be helpful, be concise, acknowledge uncertainty, push back when appropriate. The model isn't trying to please you — it's trying to be genuinely useful according to its constitutional guidelines.

This isn't a value judgment about which approach is "better." They're different tools optimized for different outcomes. But if you apply ChatGPT habits to Claude, you'll get unremarkable results because you're not leveraging what Claude is actually good at.

Seven Principles for Using Claude Effectively

Jones outlines seven key differences in how to approach Claude:

1. Flagging Flaws

Claude is more likely to challenge your plan, stress-test your ideas, and highlight potential errors rather than just agreeing with you. This is Constitutional AI in action — honesty is an explicit training principle.

If you want validation, this feels like friction. If you want your ideas pressure-tested before you commit to them, this is invaluable. Lean into it: ask Claude to find holes in your reasoning, identify risks you haven't considered, or tell you what could go wrong.

2. Situation Over Output

Instead of issuing commands like "write a cover letter," describe your specific situation and context. ChatGPT is optimized for command-response patterns. Claude excels when you give it the full picture and let it reason about what you actually need.

Bad prompt: "Write a cover letter for a security analyst job."

Better prompt: "I'm a security analyst with 5 years in healthcare, applying to a CISO role at a mid-size hospital system. I'm concerned my experience looks too technical and not strategic enough. Here's my resume and the job posting. Help me think through how to position myself, then draft a cover letter."

The difference: you're inviting strategic reasoning, not just output generation.

3. Editing Over Generation

Claude excels at refining and editing existing work rather than creating from a blank page. Give it your rough draft, your half-formed thoughts, your messy first attempt — and let it help you improve it.

This maps to how skilled human collaborators work. They don't write your document for you; they react to what you've created and help you make it better. Claude's writing also tends to sound more human and less generic when it's editing rather than generating from scratch.

4. Extended Thinking

Use Claude for genuinely hard problems — debugging complex code, analyzing contracts, working through multi-step reasoning. Claude can show its chain of reasoning, which produces better results on problems that actually require thinking.

Don't waste extended thinking on simple tasks. Use it when you have something that would take a skilled human significant time to work through.

5. Building a Workspace

Treat Claude Projects as persistent operating environments, not just file cabinets. Provide detailed context about your role, your goals, your constraints, and your preferences in the project instructions.

This creates consistent behavior across conversations. Claude isn't starting from zero each time — it has context about who you are and what you're trying to accomplish. The more specific your project setup, the less you have to repeat yourself and the more useful Claude becomes.

6. Desktop Integration

The Claude desktop app (Cowork) can autonomously read, edit, and organize files on your computer based on your instructions. This moves Claude from a chat interface into an actual workspace tool.

If you're still copy-pasting content into a browser window, you're adding friction that doesn't need to exist.

7. Know the Limitations

Claude does not generate images. Its real-time web search capabilities are more limited than ChatGPT's. There's no equivalent to the custom GPT marketplace.

If you need those capabilities, use the tool that has them. The goal isn't Claude loyalty — it's using the right tool for the job.

Why This Matters for AI-Assisted Work

The broader point isn't about Claude specifically. It's that different AI systems have different strengths based on how they were built. Treating all LLMs as interchangeable "AI assistants" means you'll use all of them suboptimally.

As these tools become more central to knowledge work, understanding their architectures — at least at a high level — becomes a practical skill. You don't need to understand transformer mathematics, but knowing that Constitutional AI produces different behavior than RLHF helps you get better results.

The Security Angle

For security practitioners specifically, Claude's willingness to push back is valuable. When you're drafting a policy, building a threat model, or evaluating a vendor's claims, you don't want an assistant that agrees with everything you say. You want one that flags the gap in your logic, asks about the edge case you didn't consider, or notes that your assumption might not hold.

RLHF-optimized models can be too agreeable for security work. Constitutional AI's explicit commitment to honesty — even when it's uncomfortable — aligns better with a discipline where missing problems is worse than surfacing false positives.

This is a non-security post. For the AI Security Series, see the previous entry on Google's Cybersecurity Forecast 2026.