Yes, each witness controls their servers, which runs their Steem node. The database and software are integrated into one package called Steemd.
It requires 15 or 16 out of top 20 witnesses to co-ordinate on a hardfork. That's what it'll require essentially to remove content. However, to do this every time there's illegal content seems very messy to me.
There's no issue if a service provider removes the content and co-operates with the authorities to find who posted it. But as a Steem witness, especially as a low ranked reserve witness, I don't have the option of removing the content available to me.
Yes, I understand that from the other comments and agree: too messy. What I'm trying to propose is a stepped change of the entire software package. First a quick fix to show goodwill AND make it harder for the authorities to get proof. Then a permanent filter mechanism at the first point of data retrieval.
To my limited understanding it should be possible - is it? (Technically; never mind consensus issues for now. )
That would be possible on a centralized solution like Reddit. Not practical on Steem.