A neural network learns when it should not be trusted

The MIT wrtocle does not link the original article. Here is the original research paper:

https://proceedings.neurips.cc/paper/2020/file/aab085461de18...

How can we trust the uncertainty metric? Additionally humans are prone to intutively misjudging probabilities...

This isn't about turning up the honesty level on TARS, it's more like giving the neural network a third option or additional dimension with which to respond.

No doubt it will be abused in press releases and P-hacking by humans, but at least it allows developers to handle the "That doesn't seem to be a pedestrian but I really can't state anything confidently right now" case instead of just getting "is_pedestrian=false"

Is anybody else finding the recent surge in MIT PR about ML to be a bit overselling the capabilities of the underlying research?

haven't they always? especially with Media Lab