Why Bad Human Driving Gets Into the Data

A security camera in a kitchen records great plating and also the dropped pan. Fleet cameras are not moral filters. They record tailgating, late brakes, rolling stops, and the occasional inspired save.

Teams try to keep the useful clips and down-weight the junk. That filter is itself a product. Too strict and you delete rare recoveries you needed. Too loose and the network learns to hug the bumper in front because that is common.

Fleet cameras are not moral filters.

Some “bad” driving is locally rational. A human cuts into a shoulder to let an ambulance through. A literal imitation of legal lane-keeping would block the ambulance. Some “good” driving is only good in that city. Copy it elsewhere and it becomes rude or illegal.

So data cleaning is not hygiene theater. It is choosing which teacher to trust. The long-term design is usually mixed: imitate competent humans for ordinary texture, then add extra objectives so the policy is not allowed to copy every vice that showed up on a Tuesday commute.