Should AI follow the design system too strictly?

I work with a design system, and lately I’ve been thinking about what happens when AI starts using it too.

Giving AI our components, tokens, and patterns sounds ideal. It makes the output much closer to something we could actually use.

But there’s a tradeoff I’m not sure about.

A design system is based on decisions we’ve already made. If AI always follows those decisions perfectly, it may also keep repeating the same solutions, even when a new problem could benefit from something different.

So I’m starting to think AI might need to understand the design system without always being constrained by it.

For teams already using AI with a design system, how strict do you want it to be? When should it follow the system, and when should it be allowed to suggest something outside of it?