It's counterintuitive - changes in perception of reality are more useful than changes in reality.
Which is precisely what is being researched and done in various areas, e.g by making the road appear narrower through various visual tricks.
And it works.
Until people learn. I don't mean by that that they necessarily know they're being tricked and consciously "undo" the trick, but they are then trained to handle a tougher constraint, and thus are able to routinely do something they could not before, out of pure training. The whole danger/constraint/learning system is actually very elastic. The net result is that not only they end up driving at the same speed as before in the modified area, they actually drive faster than before in the other non-modified areas.
These data should not be used to make the claim that attempting to increase safety is useless. Improved vehicle safety features, for example, has been compensated for by faster driving. That means the safety features didn't help lower deaths, but they did help everyone get around faster.