1) The whole airspace environment is heavily regulated (ATC, rules, etc), cooperative (think transponders or TCAS for example) and usually quite sparse (several miles between planes).
2) Pilots are regularly trained and evaluated for these kind of take over in case of emergency.
3) You have two pilots monitoring each other.
4) The development cycles for these automations are far longer, and the QA a lot stricter.
5) At least in the most critical phases (take-off, landing), the automation is not exactly end-to-end. It's more the case of grouping batch of steps together into bigger actions to reduce the load on the pilots. The pilots are still quite busy during these periods and very much in the loop.
Planes operate in a space with good visibility, any obstacle will be visible with ample lead time.
Planes operate in controlled space, with dedicated and verified spacing between traffic participants, allocated by an external traffic control. Collisions are not rare because pilots jump in within split seconds to avoid them, but because traffic control routes planes away from each other if they converge.
The planes autopilot is in fact less capable than what the teslas autopilot pretends to be. Its limitations are fairly well understood by pilots.
A planes autopilot doesn’t dump the pilot into a “you’ll crash in a second” situation. Handover from autopilot is a different process than what Tesla expects. The planes systems are designed with human psychology and limitations in mind, because of the comparatively high regulatory bar.
Autopilot on cars is a MUCH harder problem than on an airplane. Plus you're taking about trained professionals and liable companies vs average Joe and negligent developer.