Also, filtering on a waveform is much harder than text data.
I know the current infrastructure can't really deal with that, but it's because the telecoms have no real reason to implement it.
My employer seems to have gotten a lot of sequential phone numbers assigned, so a similar number used to be highly correlated with legitimacy, now it’s a mixed bag.
"I buy junk cars" is pretty easy to filter out. Nowadays they send things like "i/buy/any old/car Yconpro/autos*" (this is a real one I got 9 hours ago). But that's still easier to design a filter for than a phone call which only happens over audio.