"Why" questions don't work very well on SE, because the answer is often 'no-one knows', however…
In this case, there is little commercial drive for it.
A musician wouldn't want it because they'd be out of a job. A songwriter wouldn't want it, likewise. The consumer wouldn't care how it was made, but unless someone was going to put a whole set of fake performances behind it to fill up the YouTube etc side of the business, who would even know it existed, let alone buy it?
That's before we even get to the 'how good can it be made to sound?' issue.
Some performances - the modern pop record for example - are tuned & aligned to the nth degree. Enough to satisfy the 8 to 13-year-old the product is intended for.
They are already swayed more by the artist's video & twitter feed than they are by the music.
By the time you get to an older, more discerning audience, prejudices have already begun to form. Trying to pass off some computer-generated algorithm is likely to offend as much as amuse.
Software already exists to generate chord progressions & even melodies, but even people using these would tend to heavily edit using their ears & learned musical sensibilities. there is also software than allows the user to build from pre-made building blocks, like Lego. This also tends to really be used by the followers, not the leaders in any field or genre.*
So, there's just no real avenue for it to develop in… as yet.
I have a friend who makes
lift musak … ermm … production/library music for a living (& he makes a ruddy fortune doing so). He can turn out an album in a fortnight, so long as he doesn't have to take an orchestra into Abbey Road for the 'posh stuff'. That's probably quicker than it would be to sift through the random musings of a computer set to work on the same 20 tracks.
A real-world analogy to this. I was once told by Pete Waterman (yup, him) that if you're making a record that sounds like one in the charts, you're already three months too late. I'm not so sure that these days that is quite so true as it was in the early 80s, but there must still be some truth to it.
There seems to be some uncertainty as to whether the question is actually about whether a machine can compose & arrange a piece or the far simpler can it play a piano score.
If that's the case then the answer becomes, very simply…
They already can. It sounds like pants. No-one would ever use one.
It's like comparing the Duracell Bunny to a drummer.
Drum machines have their place. That rigidity was an element in early 80s music, however it was very often tempered by humans playing live over the top of it, softening the hard lines. Modern dance tracks, pop, EDM etc tends to re-use samples of things that have actually been moulded slightly differently. The origins of those samples have been themselves 'softened' from their drum machine ancestors. Slight temporal variations - whilst in themselves repetitive over time - have been shifted slightly out of true, to make them 'feel' better.
Once you step up from the humble drum machine to emulating how a pianist would handle a full piece… you are several generations of computer AI away from it being really possible.
Computers don't 'feel'. Early attempts at randomisation algorithms to emulate human inaccuracy turned out to just sound 'bad' not 'human'. They've largely been left by the wayside since.