As long as peers do not decide to keep the data (by pinning the hash), the old content will simply be deleted and replaced with new content. Pinning is extremely cheap, so you can imagine that you would want to auto-pin pages as you visit them. This is not on by default though, and I think that's a good thing.
> [B]ut without a central management how does IPFS associate correct file to the correct hash?
It does so using cryptographic identities. In IPFS every peer has a private/public keypair, and the PeerId is the hash of its public key. When publishing content under an IPNS name (which always has the PeerId at its root), you must prove that you 'own' the PeerID by signing the object with your private key. Other nodes will not 'mirror' your IPNS object unless the signature matches (which makes a nice consensus).
The paper calls this self-certified names: you can distribute the objects, and let them be provided by third parties without losing authenticity. See section 3.7.
Mathematics[1]. Your hash is based on the content of the file, therefore you can guarantee that the content is the same as what the original author published. Also, IPFS is not a blockchain, there is no 51% attack.
And there must indeed by some central authority giving 'ultimate trust'. This is Zooko's triangle. By default IPNS gives you 'decentralized' and 'secure', but you can also opt to tie an IPNS name to a DNS TXT record, and lose 'decentralized' (at least for the initial lookup) but gain 'human-meaningful'.
Besides, the idea is that IPFS could happily support Namecoin[1], so there's the decentralized, secure, and human-meaningful DNS service. The only 'non-secure' aspect there is the unlikely event of a 51% attack.
Namecoin is merge-mined with Bitcoin by many miners, so if you wanted to 51% it you'd need a pretty large chunk of Bitcoin's hash power. An evil pool could do this, however it would be visible on the Bitcoin blockchain that they were doing it (although not what transactions are present/not present in the attack chain).
The thing from bitcoin is 51% attack where majority of nodes collude to provide the wrong data(they might not be malicious but just serving old version of a data). This can only be prevented by time-stamping in slow scenarios.
In case of neocities, a central server is giving you a hash that the nodes would provide you the data for, but without a central management how does IPFS associate correct file to the correct hash?
If it only works for a centralised scenario then it is similar to the http://peercdn.com project where instead of a server supplying content, your previous visitor browser cache is supplying the content through p2p webrtc and you save server bandwidth.