The Problem is that it might not be significant safety outside of well suited situations.
This systems expect the user to shim in fastly if something goes wrong, which just can't work reliable given how human brains work.
Furthermore the situations which can make a self driving car crash are not the same which make a human driver crash.
Such crashes can be pretty arbitrary. Random things like advertisement on a bus has the potential to trick of the self driving car. Similar at least some systems are not very good at handling situations where the lanes are unexpected changed e.g. around a construction site. I had been in more than one situation where the lanes markings what completely messed up and only common sense was left to prevent a potential dangerous situation. This systems don't have common sense.
There already was at least on deadly incident where a Tesla crashed for no apparent reason (blaming the dead driver for not taking over fast enough which again isn't realistic at all, through legal maybe valid).
Tesla is supposedly the leader wrt. self driving cars, but people where easily able to trick that car to drive into the lane for the opposite direction and it has since a while one of the worst emergency brake assistants.
The problem boils down to self driving cars not being good at "unusual" situations but situations which are unusual for the self driving car are in practice pretty common given how much people drive cars and they are not necessary recognizable as such by humans.
Another problem is that traffic's in different countries have different rules and dynamics a self driving system working well for idk. the US might not at all be safe to use in idk. insert arbitrary EU country here even if trained for that country.
Humans can normally adapt to a degree in the fly, which makes them not have a crash, or less bad crashes in some situations. But the self driving car has no such common sense and might not even realize it's in a situation it's not trained for.
I see a lot of potential to make traffic more safe.(non stuff driving but assistant systems, self driving in traffic jams, slow self driving buses in car reduced inner cities, on highway self driving truck convoys, etc.).
But at least on the close future I'm not convinced "general purpose"/"everywhere any speed" self driving cars are more safe.
I mean we also shouldn't forget that a lot of deadly accidents happen because of negligence, so while self driving cars still require a driver hand over, they still all happen because the driver is as negligent (or might intentionally not use the self driving feature to get faster to work etc.).
The key metric for Tesla and other manufacturers is not necessarily duplicating human cars avoidance in any given situation. Some situations might be easier for humans, and vice versa.
The right metric is simply a much lower accident and fatality rates for self-driving cars.
This flexibility reduces the challenge for self-driving cars since they are far from understanding situations at the level people potentially can, but still results in a safety win.
This was illustrated in a chart[0] by someone on the Tesla subreddit just yesterday.
Using Autopilot is in general much safer even though it does from time to time make fatal mistakes that human would not do. (Like when it fails to recognise a truck and tries to drive through it[1].)
Of course AP is a very different thing to FSD, so this doesn't speak anything about current nor future FSD safety.
> Another problem is that traffic's in different countries have different rules and dynamics a self driving system working well for idk. the US might not at all be safe to use in idk. insert arbitrary EU country here even if trained for that country.
That is why Tesla is not hyper focused on one city and is not building a team of exclusively AI experts, rather build a tool that other people can use to configure the specifics.
Tesla already are able to read signs in many languages and know about lots of different country specific rules.
> Tesla is supposedly the leader wrt. self driving cars, but people where easily able to trick that car to drive into the lane for the opposite direction and it has since a while one of the worst emergency brake assistants.
'Tricking' the car is kind of nonsense, you can also trick people into things pretty easily.
I have seen a number of reports where Tesla AHAD got the top marks and I have seen many videos of Teslas breaking when the driver was not paying attention. What are you basing this analysis on?
Until some really big accident happens in a situation where even a drunk driver wouldn't have messed up that way.
After that self driving will be limited to situations like while you are in a traffic jam or limited to 30km/h and having it's own track or similar.