1) below a certain scale, they don't have to comply

2) if they make a good faith effort, at whatever level of expertise and resources they have, I belive they will be within the law. I dont think there is any specific legal requirement to do machine learning or any other specific implementation of content ID.

I'd start with checksums, whatever level of flagging by hand I could afford, and go from there.

As you scale you'll attract the attention of rights holders and need to spend more. But I would hope the courts would accept a good faith effort.