The size of the error resistance coding you'll need will quickly dominate the size of your data. Also the argument is a probabilistic one. It's not that certain data will be recoverable, only that there may be a non-zero probability that some data may be recoverable.