Some exit relays are considered as bad exits. They might listen in your traffic, change your traffic etc. Tor does not choose them (as long as you don't explicitly allow it).
How would TOR know that some exit nodes are "bad" and have ways to monitor or manipulate the traffic routed through them?
There's no way to know whether an exit node is "bad" as long they remain passive (only listen to traffic), at least unless someone decide to leak government confidential document.
And if they have such knowledge, why are those nodes not removed and blacklisted?
I don't know, but i guess someone know the answer.
Operators who run more than one relay should declare those relays their 'family' (There is a special option in the configuration). Tor doesn't choose more then one relay from a family.
The emphasis is on "
should". They didn't say "
must" or "
have to", but you have the option to do it or not.
It's impossible to enforce it on decentralized system, so it's pointless to use word such as "must" or "have to".