Your first option---"someone gets [a grade] only by correct answers"---hides some problems. How does one convert a fraction of correct answers into a grade, and correct answers on what problems?
Unless you have some sort of externally given bank of problems with a built in curve, both of these choices are essentially arbitrary. (There happens to be a customary answer to the first question---90% is an A, 80% a B, and so on---but there's nothing particularly canonical about it, and it doesn't address the second question, which questions should a student get 90% of right in order to get an A.)
In some courses you may have some sort of external guide to what adequate performance is; for instance, you might be norming against an exam provided by some external source, or you might have a clear view of what should count as A work. A crucial property of this scenario is that your exam questions wouldn't change if you knew student scores; if all the students get D's this year, you'll conclude that they hadn't learned enough, give them D's, and ask comparable questions next year (possibly while reconsidering how you were teaching).
If that's not the case---if your response to a year where all the students got D's would be to adjust the questions---then implicitly you are setting the grades against an expected class performance. If you look at scores in the 60s and think "This year everyone gets a D, but next year I'll make it easier so that doesn't happen again", you're effectively applying a curve, you're just doing so inconsistently (and in a way which is unfair to some of the students).
Any time the class deviates from the expected performance, you have to ask "Is this deviation a reflection in the abilities of the students, or the quality of the exam?" The latter is actually much more variable than the former, so we often attribute most of the variation to the exam, and adjust class grades accordingly. (Depending on the system, there may be intermediate options; I generally curve exams to an expected median, but some years there are indicators that it's a stronger or weaker group---for instance, I may have particular problems I can do a direct comparison with other year's performances---in which case I adjust the curve a little.)
Phrased differently, curving a class (for instance, by saying in advance what percentage will get each grade), is often characterized as being about some kind of competition amongst students. This isn't necessarily the case. Curving a class is often using an assumed distribution of student performance to norm the exam.