I use AP for about 80% of my driving. Last fall a young deer ran in front of my car and AP slammed on the brakes before I saw it. We missed the deer by inches. If it wasn't for AP I would have certainly hit it. By the time I saw the deer it was less than 20 feet away. The car could not have physically stopped in time. So for me, at least, it's not all selection bias.
First, I do not deny that AP, and related Tesla safety features (like automatic braking) are beneficial from a safety point of view in situations like those you describe. There are cases, as in your anecdote, where these features can prevent an accident, or at least reduce its severity. There are, however, also cases in which AP can cause an accident. Just in the past month, I've experienced erratic lane changes and the car appearing to point itself head-on into fixed objects. I do not know whether the car would have crashed or not, since I took over from it. Those anecdotes, like yours, go into the analysis. It's only by combining all of this that you get anything resembling scientifically-admissible data, rather than mere anecdotes.
The term "selection bias" has a
specific scientific definition, relating to how subjects are selected for inclusion in the study or assignment to experimental conditions. In most cases, a scientific study in in medicine or the social sciences should ideally have subjects randomly assigned to experimental vs. control conditions. This isn't always possible, and scientists have ways to try to work around the problem when it's not, but those methods involve a lot of attention to detail, additional measurements, long-term study, etc. Selection bias is made worse if there's a reason to believe that the selection may influence the outcome of a study. For instance, if you wanted to study the effects of taking a vitamin on mental acuity, you would
not let the subjects choose which group they were in, since those choices might reflect characteristics that affect the study's measures (motivation to perform well, say), and since knowledge of whether they were taking the vitamin could affect the outcome.
Your last sentence quoted above is meaningless in light of this definition, since you're explicitly speaking about your personal (anecdotal) experience, not a scientific study. Applied to the Tesla safety report referenced earlier, the "experimental" conditions are non-Tesla cars, Teslas driving manually, and Teslas driving on AP. The last two of those conditions are defined largely by driver actions; they are not assigned randomly. As I and others (including you) have said earlier in this thread, AP is more likely to be used in easy driving conditions, which will make it seem safer than manual driving. The Tesla safety report makes no attempt to disentangle the
huge amount of selection bias at play in defining those two conditions. The non-Tesla vs. Tesla distinction is also rife with selection bias, but it's probably not as bad as the AP vs. manual driving selection bias, since it's not done on a moment-by-moment basis in response to the sort of road conditions that will themselves affect the measure under study. There are also many variables between the non-Tesla and Tesla conditions (cars' driving dynamics, prevalence of non-AP safety features like forward collision warning, demographics of people who buy non-Teslas vs. Teslas, etc.). These differences make it impossible, based on the data in the report, to do more than speculate about what might cause the differences shown in the report between the non-Tesla and Tesla vehicles.
Put another way, selection bias enables the variable under study (such as Autopilot use vs. manual driving) to be mixed up with other variables (such as road conditions, weather, speed, highway vs. non-highway driving, etc.) in such a way that it's impossible to determine which variable(s) are responsible for the observed differences in number of crashes. We cannot conclude that AP is safer than manual driving because the observed differences could be due, in whole or in part, to these confounding variables. That said, we also cannot conclude that manual driving is safer than AP. We just can't conclude anything. We simply cannot draw any conclusions about the differences observed in Tesla's safety report because of the selection bias.
Unfortunately, most people have a limited understanding of this, and may misinterpret Tesla's safety report to mean that AP is safer than manual driving. I can't say definitively that AP is
not safer than manual driving, but I can say that Tesla's safety report does not address that question. That is, AFAIK there exists inadequate data on this question. Considered as a scientific study, Tesla's report is poor at best and completely worthless at worst -- probably closer to the latter than the former, although it might have some value if combined with other data and subjected to clever statistical analyses. Considered as propaganda, Tesla's safety report exploits peoples' limited understanding of science to make them draw an unjustified conclusion and make them think it's backed up by science. Thus, it may well be effective propaganda, but it's not good science.