As abstracted allowance functions, because this can inherently also have call-count restrictions, making it work the other way if needed, making it more open ended, potentially supporting intent-style permissions.
These are two different approaches to achieving some kind of safety. A couple advantages to this interaction pattern are:
This approach doesn’t rely on disclosing any information to the site (like your account & tx history) before granting it relevant permissions, improving privacy and reducing cherry-picking phishing attacks.
Comprehending arbitrary bytecode is an inherently hard problem, and most modern solutions rely on introducing new centralized systems for helping analyze this potentially inscrutable data. These systems are prone to censorship and downtime, are not available to offline signers, and are not able to represent all kinds of actions that an application might want to perform on a user’s behalf. By inverting the order of proposals, the user is able to understand what they are putting at risk with potentially no external dependency, while still permitting the application to perform arbitrary actions on their behalf.