Looks like the text has changed a bit since it was last in committee ("non-availability" changed to "prevent the availability") but in this case the issue is:

"[...] service providers [...] shall [...] take measures [...] to prevent the availability [...] of works or other subject-matter identified by rightholders [...]"

This is a bit technical but "shall" here has the legal effect of "are under obligation to". This is what makes responsibility for content a matter for the service provider, whereas previously the service provider was not required to prevent availability, only to take content down when requested. The current situation is that the USER is responsible for copyright infringement, and the service provider is not, for user-provided content, but the service provider has to respond to and act on takedown requests. This is what's being changed.

And "effective content recognition technologies" is what other people are referring to as "upload filters". This "content recognition technology" (which Voss specifies to be google's content id or something very like it) is given as an example, but there are no criteria here other than appropriateness and proportionality, which are not defined anywhere (is it proportional in terms of effort for the rightholder, the user, or the service provider?) and it's unclear what the alternative to said technology could be. Practically if it's your responsibility to prevent the availability of content you need to screen it before publication, either manually or automatically, and somehow allow rightholders to prevent its publication. This could mean comparing against a list of all copyrighted works, but that would cause difficulties in compiling such a list, keeping it up to date, and deciding how close a match it is. Voss, in his post, says that publishers will provide service providers with such lists, but the comparison operation itself is computationally heavy, especially for video, and also error-prone. It's also subject to abuse as anyone who claims to be a rightholder can effectively censor anyone else's content by putting it on said list. Youtube is the only implementation I'm aware of of such a technology, and their filter is terrible in terms of recognition, both in false positives and false negatives. The law as written effectively gives no alternatives to content recognition technology (are you going to manually check that everything everyone uploads is not on a large and growing list of works?). In the end, whatever implementation is picked, it will effectively prevent the publication of a large category of legitimate content.

Putting the matter of whether the whole thing is a good idea aside, it's also bad law, in that it's an unclear formulation of requirements that leaves a lot open. It leaves the question of what is appropriate and proportional to member states, and ultimately to the courts, because it doesn't specify any criteria. It gives a specific technical solution as an example but doesn't specify what makes it appropriate or not, making it, again, a matter for the courts to decide whether specific implementations are appropriate or not. It doesn't even specify what qualifies as "large amounts" of works.