Segmentation is harder to work with, but it has benefits too. At the time, when it was all new, everyone just complained about it: they wanted the ease of use of flat, unsegmented memory, with just one memory address.
More recently, the NX bit has become desirable and promoted. The CPU can flag areas of RAM as data, and mark them as Non eXecutable. That is all: one bit, so either it is code and you can run it, or it's data and you can't.
But the original Intel 386 design had 4 protection rings: 0, 1, 2 and 3.
0: kernel mode, can do anything.
1: inner ring, privileged code, but with restrictions.
2: outer ring, some hardware privileges: can do device I/O.
3: user mode code, no hardware access privileges, has to request everything through the kernel.
OS/2 2+ used the middle rings, and Novell Netware 4.x, optionally, for NLMs (apps that ran on the server, which was rare for Netware). Nothing much else.
Microsoft only ever used 0 and 3.
But with memory segments, you could, in theory, flag each segment by ring. This is much more flexible and controllable than just one NX bit.
I've written about this:
https://www.theregister.com/Print/2011/07/11/a_brief_history...
Segmentation had benefits, but they took more work, more effort and thought and planning. You needed to design the OS for it, and the drivers, and the apps.
But remember how DOS evolved: it's a clone of CP/M. No multitasking, no permissions, only 1 app in memory at a time.
Back in the 1980s, the industry was growing by adding stuff onto a brain-dead, super-simple design from the 1970s.
They saw segments as a hindrance, an unnecessary complexity. They wanted them to go away.
In the 1990s, we got 32-bit desktop personal computers and 32-bit end-user desktop OSes. Apple's design was a single-tasking OS and it had to buy a Unix vendor to make a grown up OS.
Microsoft's engineers achieved miracles bolted on to DOS, but in the end, it had to ditch its entire OS too, and build a whole new one: Windows NT. It took headhunted talent to make it work.
IBM dropped the ball, which is the whole point of my blog post that is the root of this discussion.
Now, we have CHERI, building protection segments back in again. I've written about that too:
https://www.theregister.com/2022/07/26/cheri_computer_runs_k...
It's a huge waste of human effort: we had them in the 1980s, but they were too hard and we threw them away.
UNIX did too. Multics used them, but Multics was too hard. UNIX was Multics with the hard bits left out. And now, 50 years later, we're discovering that we needed them all along and we're painfully and with great difficulty putting them back in again.