"All of the above"
The MCAS itself did exactly what it was programmed to do. One of the sensors feeding it, which was not redundant (!), malfunctioned, and MCAS acted on the data - garbage in, garbage out.
By not having sufficient redundancy, MCAS was designed in a failure-prone way. This was justified because the failure would be like another kind of known failure mode (runaway trim), and pilots already had a checklist for it (the key point being to turn off electric trim).
The problems with that are:
1. The problem a malfunctioning MCAS (acting "correctly" based on bad data) introduces is intermittent/erratic, making it harder to correctly react to it. Especially since the pilots weren't aware that a system that would act like this was on board!
2. Humans make mistakes in high-stress situations, which is why accepting something that introduces additional problems and relies on humans to handle them is a bad idea.
3. Turning off the electric trim also made it hard for the pilots to correct the situation, as it disables the manually-controlled but motor-powered trim. One theory is that the forces may have been too great for the backup solution (trimming manually) to work, a situation that old (but not current) manuals addressed.
MCAS has a limit how much it will trim, but any trim input by the pilot resets it, and lets it move the trim by the limit amount again.
This is just a layman's interpretation:
From the flight data recorder, it seems like the Ethiopian ET302 pilots only partially corrected the wrong trim, then did turn trim off. After this, there was no noticeable adjustment in trim until the pilots turned electric trim back on to be able to use it. They used it to trim a little bit, and 5 seconds after their last trim input, MCAS triggered again and drove them into the ground. [1]
On the other hand, the Lion Air JT610 crew doesn't seem to have turned trim off, just corrected it each time MCAS trimmed down (only for MCAS to trim back down 5 seconds later) [2] - until they once only trimmed a little against it, so that they hadn't corrected what the previous MCAS trim had done, but unlocked it for another one. MCAS applied more trim, and that was it.
This could likely have been avoided by telling the pilots about MCAS and how it works, having a separate switch that kills MCAS but not electric trim, some sort of indication that MCAS is working, some sort of indication that the AoA sensor is broken, redundant sensors, and countless other things. The most likely reason why those weren't implemented seems (for me) that adding them would likely trigger the certification/training requirements that Boeing was trying to avoid.
[1] https://leehamnews.com/2019/04/05/bjorns-corner-et302-crash-...
[2] https://static.seattletimes.com/wp-content/uploads/2018/11/L...
