The paper explains that many of the performance issues of the internet (circa 2016) can be attributed to a design choice back in the 1980s to deal with TCP congestion.
Specifically, treating all packet loss as congestion
As NIC (Network Interface Controllers) and memory chips became more performant, this equivalence started to fail.
Now it is quite cheap to have bigger buffer queues, orders of magnitude larger than the buffer-delay product, which causes excess delay.
Research endeavoring to reach an optimal operating point (max bandwidth, min delay/loss) was cut short by a discovery by Jeffrey M. Jaffe which stated that no distributed algorithm could converge to such an optimal point.
However, this theorem assumes ambiguity of measurements (i.e. what is the particular cause of an RTT increase). But, if we look at a connection's measurements over time, we can disambiguate with some level of confidence.
Hence the title of the paper: instead of congestion control based on any packet loss, they want to attempt congestion control based on actual (inferred) congestion.
The best part about this paper is that their solution BBR runs purely on the sender, not requiring costly changes to the network itself - again, we see that deployability is an important feature.