They can trivially determine if their tech is effective.

Can they? How many people real world test, and are they of all different heights and weights and face shapes too?

Besides that, when I was a kid, I used to watch a lot of old movies on late night TV. Often these movies had car chases, and cars would go careening off of cliffs for no reason. I was always flummoxed, for we had no cliffs anywhere I'd ever been, and wondered where they were, and why people were always driving on them.

When I visited California I suddenly realised "oh, they're everywhere here, just driving home".

Another poster pointed out the alarm went off, if he looked to the corner he was driving towards. People dogfooding won't notice issues with that, if the local environment doesn't have such features.

Could you test for all these things? Maybe, after realising what to test for. You'd then need a sort of regression test, too. All with people.