The key generator can decrypt any message, of course. That is why I said threshold systems should be used, so that no single party is the key generator. In that model, the receiver would have to request that several parties jointly compute and issue the decryption key. It is also reasonable to imagine a world where there are many IBE authorities, and the sender of the message can choose which IBE authority the receiver will have to get their keys from.
After all, what tells the Key Generator that [email protected] is actually [email protected] and not Alice?
Now, someone contacts the key authority and says I am [email protected] what's my private key. Without prior communication between Bob and the key authority there is no way to do that exchange over an open channel securely. Assuming email addresses are public information and someone can get bob's email address before he communicates with the key authority.
Think of it like this in a pure one way communication you send all parts of the message, and anyone who reads it can decode it. If you instead get one of there key's (even if it's from a 3rd party) you can encrypt something for them. Or you can think of it like this, if the receiver never makes any key's there is nothing special about them that prevents others from reading there messages.