That is because I haven't commented on client side predictions, but on the general "make car driving client side".
But now we're talking about it, predictions have a chance of being incorrect.
How the cars are now, when you give input (i.e. press W for acceleration), you ask the server for permission to move the car. Server answers with permission granted, and your the car moves forward. Server tells everyone where the car is.
When the cars are client side, the client has to tell the server were the car is and the server forwards that data to other clients.
When the client is going to lose packets, it can't tell the server where the car is. As long as the client isn't giving different inputs during packet loss, client side prediction would work fine. Now, if the car is going straight, starts to lose packets and starts making a turn, the server wouldn't know what. Client side prediction would still show the car moving straight forward, even if it collided with the wall. Packet loss is over and then suddenly receives a different location of where the car is. If you were attacking that car at that moment it would disappear and appear somewhere else.
Compare it to a weather forecast, they can show you what the weather might be next week. The next day they might show a completely different or slightly altered forecast for that day. They have a week to predict and correct the forecast, and there will still be a chance for it to be wrong.
Now compare that to a online game, where it only has milliseconds to correct itself with inconsistent latency. Results will be all over the place. I don't know what the test environment was during development at RTW, but most likely local network. If that already showed issues, it would be worse over the internet.