Plagiarism should only ever be counted for explicitly unique sentences that provide actual value.

It's actually an ongoing debate in software engineering, due to licensing, as to what you can consider "stolen code" - i.e. plagiarism.

In fact things went as far as to some companies employing AI-aided automatic cease-and-desist deliveries on GitHub, but the system was so badly configured, it detected even the most basic logic bits as license infringement. Things that are standardised in software development - like, for example, for loops, that happened to have generic parameter names (e.g if you were to create a graphic subsystem for displaying Views, whatever the primary implementation may be, you'd iterate through all views with a for loop, making it a generic call such as for(val view in views) { [do something here] }).

Well this AI aided detector was so brilliant that it detected such minute coincidences of codebases as legitimate violations (as if any company could copyright generics), and sent these spurious C&Ds to dozens of git repos. What's even worse is that the initial company's codebase used some open source libraries that were directly attacked... for being 100% copies of their own codebase.

IMO as long as the code/sentence isn't a provably unique statement, plagiarism shouldn't apply. A whole paragraph having 80%+ similarity to something unique? Now that's worrying enough to investigate.