Although MD5 has known collision issues, using an MD5 collision to alter the text of the document in a specific manner is highly unlikely. This is not as trivial as simply generating some random bits of data that result in the same signature when hashed.
In the use case where you want to change the document for your own purposes without leaving evidence that you've tampered with it, you would not be searching for just a single other file which generates the same hash value. You would be looking for another file which results in the same value when hashed, has the same format (text is easy, I suppose, but it's conceivable they could switch to a binary format at some point), makes grammatical and syntactical sense, and also contains the desired content that you are looking for (or some approximate.. the 'exact' desired content would result in a different hash in most scenarios).
That said, I agree that there are better hashing algorithms out there. Multiple signatures from several common hash functions could be provided for quick confirmation purposes while still making it relatively easy to validate quickly.
(Edit: Also, since you are concerned: I am not the downvoter)
I know it is not trivial but to start off with a known broken hash would be a big mistake in my opinion. There have been some pretty impressive demos regarding this.
A collision in under a minute would seem to me to qualify as 'broken' for all practical purposes of this particular has function for the purpose of proving a document was not modified after it was signed.
That that minute is still a substantial number of cycles is not too relevant given the importance of the documents that are being modified, it counts as a 'non-expense'.
In the use case where you want to change the document for your own purposes without leaving evidence that you've tampered with it, you would not be searching for just a single other file which generates the same hash value. You would be looking for another file which results in the same value when hashed, has the same format (text is easy, I suppose, but it's conceivable they could switch to a binary format at some point), makes grammatical and syntactical sense, and also contains the desired content that you are looking for (or some approximate.. the 'exact' desired content would result in a different hash in most scenarios).
That said, I agree that there are better hashing algorithms out there. Multiple signatures from several common hash functions could be provided for quick confirmation purposes while still making it relatively easy to validate quickly.
(Edit: Also, since you are concerned: I am not the downvoter)