Theoretically speaking, if you have the cooperation of 50% of the network (in terms of computational power), then you could remove 'in-appropriate' transactions. However, after you decide to remove them, their will be 2 versions of the block-chain until yours becomes longer than the original again. Obviously, this is never going to be attempted.

Unfourtuantly, there is theoretically no way to prevent people from coding arbitrary data into a transaction. Assume you do have a filtering mechanism. It seems obvious that at any point in the block chain, there is at least 2 legal, non-suspisious, transactions. Using just this, Alice can encode 1 bit of information by saying that the first legal transaction is a '0', and the second legal transaction is a '1'. Transactions can be ordered in a deterministic way (for example, look at them as a giant binary number and sort numerically). The psuedo-anonomous nature of bitcoin means that attempting to filter spamming by the source would not work, as you cannot differentiate a new account that is part of a spam attack from a new account that spent at a bad time.

Could you just set up 8 accounts then transfer bitcoins between then, with the chosen recipient indicating the value of a byte?

Obviously it wouldn't be readable to anyone who didn't know what to look for so it couldn't be used for spam/memorials/etc but it could store information.

Yes. You could also establish and publish a simple protocol for this, so that someone could run a program and extract all the hidden messages the way we can currently by looking for plain-text. Granted this would provide a means of detecting the messages (and therefore filtering them). This could be counteracted by first posting the data, then posting in the block-chain the fact that you posted the data (along with relevant meta-data). If you do this in a probabilistic way (IE, a single transaction has a 5% chance of being an advertisment of your data), then any would be filter would need to make a trade-off between false-positives, and ease at which a reader could operate.

Also, even when posted in plaintext, I wouldn't say that the messages are readable by people who are not looking for them. (How often do you look at the block chain?)

It should be pointed out that doing this wouldn't be free. Fees are mandatory when transaction amounts are very small or the funds were spent recently.

re: "is there a deterministic way[?]"

I would assume so.

Perhaps adding a word-checker to bitcoin clients that looks for recognizable words (or attempts to validate a URL found inside the most recent block).

If block contains words or URLs, disregard it. If not, accept it.

Now Namecoin on the other hand...

If I understand correctly, these people are using special software to encode data in transactions, namely the output addresses. You can only embed arbitrary data in blocks if you're able to mine a block (like Satoshi's headline in the genesis block and Dan Kaminsky's Ben Bernake ASCII art), otherwise you need to encode it in the transaction data.

Anyone sending Bitcoin can add arbitrary data in his transaction; no need to be a miner: https://en.bitcoin.it/wiki/Script#Transaction_with_a_message

Well then all of the worries about this child-porn being in the blockchain are warrantless.

You could create a link to child-porn from this very comment I just posted using the right cipher. I am sure of it.

What if A) your comment was preserved as a part of the block-chain that could not be removed without doing serious damage to bitcoin. And B) the cipher you are using is a publicly known cipher recognized as the standard way of encoding data into the block-chain.

That almost certainly cannot work. You'll have false negatives, people will sneak the words of synonyms in (just think of how many web forums has people saying "fcuk", or "f*ck"), and false positives, legitimate transactions being banned. Will anyone from scunthorpe be able to use it?