Well if you think of this closed beta as a way for Tesla to collect data about its program's shortcomings it's easy enough to see how releasing could save lives in the long run even it is takes some in the short run. Every day sooner that a 10x better self driving car comes is hundreds of lives saved.
And what if it gets stuck in a local maximum and never improves? Then you lost those extra lives for no reason, which consequently is what I believe has happened. At least 4 people are dead in preventable autopilot crashes, and the real number is probably over 10. For the number of Tesla’s on the road that number is way too high.
I don't only want to save lives, I want to maximize utility. I think driving less would raise utility, but society disagrees, so this is the next best option.
A peer with average drivers would probably be fine.
A peer with average drivers and doing completely nonsensical things every now and then is basically an average driver on their cell phone though and the verdict on that is that it's not fine.
In this case I'd prefer if "discovering error cases" was done without risking actual lives. Injuring someone would not be a valid way discover an error case.