...the driver’s reaction time is much slower than the car, so the car reacts first, and faster, and by the time the driver had a chance to try to “correct” the car is already out of harm’s way.
I hear you, but I disagree. I have a repeatable spot on my normal route where NOA decides it wants to take an exit ramp for no reason, without indicating prior intent. I have my hand on the wheel as usual. As soon as AP attempts to turn away from my intended straight path, because I'm holding the wheel firmly, the torque sensor trips autosteer offline immediately. To the outside observer, the car barely twitched. This is one of the reasons I hold the wheel firmly enough to avoid the nag. Every once in a while, a minor crisis is averted with no drama whatsoever.
You have to make a conscious decision to apply a force to counteract the AP action and this requires time. But by the time you’ve reacted (150-250ms) you may also have processed further information like the danger the car detected before you did, and decide that fighting the AP action is not what you want to do.
On the rare occasion I do get a nag, it's usually because I'm unconsciously applying the exact same inputs as AP is commanding, resulting in no driver torque being applied to the steering wheel. So I agree, sometimes driver and AP are really on the same page.
It’s also possible the force required to “fight” an evasive maneuver by AP is dialed up a few notches relative to normal, but I don’t think that would actually be necessary.
It seems that some aspects of AP performance are linked to AP's confidence in sensor data. For example, at night, in the rain, with obscured lane lines, the nag interval is much shorter than in dry daylight conditions. It seems reasonable that if confidence in a potential collision was sufficiently high, AP may increase the rejection threshold for contrary driver inputs.
