So... their models are not AI - they are just matching patterns. If it was AI, the systems would adapt themselves to work with the new data stream. In the mid-eighties scientists could do pattern matching like this, but the hardware was not up to snuff to scale with PDP-11's with 1MB RAM or even the mighty VAX.
While some of the math has changed (more optimized), the end result is the same. It is all hardware driven. Now the problem will shift back to software since we no longer have exponential growth in CPU speed - proving yet again that our industry needs to do better than just software bloat and feature creep.
I am very fortunate that I got my degree when I did. What I learned then is still very much applicable to my work today, and that trait is sound judgement.