No, no, I saw that in the abstract. So I looked the paper, especially the figures, and I'm not sure why the deviations they quote are so gigantic (or why they decided to quote numbers in the abstract that had such huge error estimates), but if you look at the consistency in trends and at the error bars in e.g. Figures 2 and 3, it's pretty clear that the numbers are not meaningless.

If you want to see what meaningless numbers look like, check out Figure 9.

Also, anything that comes out of the new crash test dummies is mostly going to result in changes for new cars, so if people have to buy a new car anyway to get the improvements, only the latest crop of cars matter. And I don't know where your ">2020" comes from; the 2010-2020 window has 6% disparity.

Anyway, if you look a bit more carefully at the study you can see that with regard to deaths, things really do seem very likely to have improved relatively. This is on top of large absolute-death reductions for decade after decade (see e.g. https://en.wikipedia.org/wiki/Motor_vehicle_fatality_rate_in_U.S._by_year).