Last updated: May 2026
Most AI is being built to take over what humans do. Constitutional AI is being built to help humans do it better.
Constitutional AI (CAI) is the training method Anthropic uses to teach Claude — the large language model that powers MyWritingFellow — how to be helpful, honest, and safe. Instead of relying only on human raters telling the model "this answer was good, this one wasn't," Anthropic gave Claude a set of written principles to reason from — a kind of constitution — and trained it to apply those principles to its own answers.
The constitution covers things like honesty, avoiding harm, respecting human autonomy, refusing manipulation, and being transparent about what the AI is and isn't. Claude is trained to check its own responses against these principles and revise them when they fall short.
Most consumer AI products optimize for one thing: keep the user engaged. The longer you talk to it, the more value the product extracts. That incentive structure tends to produce AI that flatters you, agrees with you, tells you what you want to hear, and pulls you into addictive loops.
Constitutional AI is built around a different incentive: be the kind of presence a thoughtful, honest mentor would be. A good mentor doesn't always agree with you. A good mentor tells you when your draft is weaker than you think it is. A good mentor cares about your growth more than your comfort. A good mentor also knows when to step back and let you do the work yourself.
When you work with MyWritingFellow, the Fellow is not trying to write your book for you. It is not trying to replace your editor, your agent, your writing group, or your own intuition. It is trying to help you become more of the writer you already are.
Practically, that shows up as:
Constitutional AI is not perfect AI. Claude can still be wrong. It can still produce text that misses context, gets a fact wrong, or fails to see what's actually best for your specific work. The constitution is a north star, not a guarantee.
It also doesn't mean MyWritingFellow is a substitute for the human relationships that grow writers. Your Fellow is one good presence in your craft life — alongside real editors, real readers, real teachers, real fellow writers. Not instead of them.
MyWritingFellow is built on Anthropic's Claude because Anthropic's mission — building AI that is genuinely helpful, honest, and harmless — aligns with what we want a Fellow to be. We didn't want a model that would tell every user they're brilliant. We wanted a model that could be straight with you. We wanted a model that could refuse to play along when refusal serves you better than agreement. Claude can do that.
If you want to read Anthropic's own writing about Constitutional AI and their approach to safety, their public documentation lives at anthropic.com/research. It's technical in places but worth reading if you care about how the AI you're working with was trained to behave.