That's not technically true - you can probe for I2C devices by only sending a START, an address, waiting for the client device to ACK (or not), and then sending a STOP. I don't think the linux kernel driver allows us to do this, but I've implemented this on some microcontrollers with lower-level control over the I2C hardware. It worked on all clients I tried it on.
This is true. As stated elsewhere, sending START conditions to random I2C devices may confuse certain parts. It may also interrupt a driver that's already talking to a part on the bus. But in my lifetime working with I2C, I've never seen a part just go haywire because you scanned it this way.
