I think there is something funny about the Z system subject detection and have posted about it here and elsewhere several times in the past. It didn't really happen here - a testimony to the people who post here - but on other forums I took a lot of abuse from people saying I was incompetent or a "fanboy" of some other brand or a troll or whatever else for saying something isn't perfect with the system - in addition to plenty of responses from people saying they have experienced the same things. Others I've seen post about this have gotten the same kind of response.
At the outset, one potentially helpful comment: some users have reported that a factory reset has helped with this, and I found that a factory reset of my Z8 did help a bit even if it didn't alleviate the problem altogether.
Part of the problem is that the issue isn't consistent and I think it can be subtle enough that when combined with the inconsistency you can go a while without noticing it but when you notice it and start testing it becomes apparent. For reference, I have not used a Z9 but have seen this on my Z8. I've seen it with the following lenses:
70-180
70-200 S
24-120 S
85 1.8 S
40 f2
180-600
500pf
The issue is, just as you describe, that when using a "dumb" focus mode like single point the focus extremely consistently be just where you put the point. If you get the point on the right spot, then leaving a little room for the fact that nothing is perfect you will find that the focus is where that point was very consistently. This includes
both images where you review them right away and know from your own firsthand experience where the focus point was
and images that you review after in camera or in NX studio. In other words, while you will see people say that the focus point the system records in the file is not always accurate to what the system was actually doing, if you use a "dumb" focus mode it is quite consistent.
It's when you use subject detection that you will see the focus point recorded in the file being somewhere other than where the focus actually was. For several years going back to my Z7ii and then with the Z8 I always took for granted that what people were saying about this was correct: it was just a minor problem with the reporting/recording of the focus point rather than with the focus itself. After first noticing this issue and doing thousand of test shots last November and increasing that possibly to over 10,000 test shots when I again dug into this extensively in the spring of this year I think the reality is actually that the focus point recorded is in fact the focus point being used - which frankly makes a lot more sense to me from a software standpoint - but that the subject detection system doesn't always focus accurately.
I don't think it's the AF system itself. I think it's something with the subject detection component, which I have come to understand as being separate from the AF system. Think of it like this: when you use single point AF you move the little box around and tell the AF system what to focus on. The camera's AF system then focuses on whatever is under that point. The subject detection system is a software algorithm that is separate from this. It acts as a surrogate or an agent for the user by analyzing the image feed and then basically moving the AF point around
as if it were the human user, the difference of course being how quickly the algorithm can react and move the point. All the AF system, strictly speaking, is ever doing is focusing on whatever is under the point it is given. The subject detection system is a software algorithm that places the point.
I have done testing on this far more frequently with people than with animals, but what I think is happening is that the subject detection software algorithm is detecting the subject and putting the point on them in the
display very well - it is amazing how well it finds subject and stays there! - but it is sometimes telling the AF system to focus in a slightly different place. Crucially, with human beings, I have observed (and on other forums another very technical user put together a whole testing setup and found similar results) that when it misses it usually misses with focus being on
the wrong eye. For example, in this image (one of the thousands of tests I did) you can se that the "eye-detect" has claimed it is focusing on the front eye, but it is clearly missed while the rear eye is nice and sharp. The system focused on the rear eye but
told us it was focused on the front - and this is not only what is reported - I reviewed all of these one at a time immediately after taking them, including making several videos of the process so that we can go back and review what the system actually did report when taking the photo.
View attachment 91439
I wasn't smart enough to figure this out, but what the user Horshack on DPReview (known as snapsy on FM) found was that this didn't really happen with head on shots, where the subject was parallel to the plane of the lens, but when they were slightly turned to the side is when it would happen. I did a lot of testing of this as we discussed and was able to verify that it does happen mostly when the subject is turned at an angle (though I did see it a few times when head on) and it didn't matter which eye was the real target. In other words, if the real target was the front eye, it would frequently focus on the rear one,
but if the real target was on the rear eye, it would frequently focus on the front eye. This is very significant because one of the things people tend to suggest in troubleshooting things like this is that the Z AF system has closest subject priority and so perhaps it is sometimes missing when something closer comes under the AF point - but what we can see here is that the system will miss in
either direction in this way, sometimes coming to a closer subject, sometimes going to a more distant one.
This has led me to believe that what may be happening is that the system is sometimes wavering between which of its detected subjects it wants to focus on. In its computational cycles it may even be wavering more quickly than the refresh rate of the EVF, and so it might be telling the AF system to move to a different subject and then back again more quickly than the EVF or rear screen can tell the user.
This could very well be what is happening sometimes even with wildlife, as I think we have all seen the subject detection system identify a body or a beak or a wing as a full on bird or even as an eye at one point or another. If it sees these other parts of the creature as a full on subject, it could be jumping to them in this way.
Another thing I tried in all of this was to use subject detection in AF-S only, and that was an improvement. This suggests to me that my theory may be on the right track because it would mean that when the subject detection system identifies and reports its target for the
first time it is actually focusing on what it says it is, but in AF-C when it can change its target that this is where the problem arises.
I also found that this problem gets much, much worse as the light decreases. This was actually my main point of interest this spring as I was photographing a few more darkly lit events and was trying to sort out how to get more reliable performance in these situations. I ultimately concluded that in lower light the system becomes extremely unreliable. Nikon does say something to this effect in the manual, but I was surprised at just what light level this is happening in. I basically found that anything at or below around 5.5-6 EV gave extremely unreliable results with "eye-AF" but that single point would still function well. However, it still does make this sort of error even in every good light - just less frequently. Also, obviously the aperture matters a lot. In any kind of light, shooting at 1.8 was far more reliable than shooting at 2.8 or 4.
Note that I tried using small custom AF boxes to limit the subjects the system would consider, but this did not improve things much.
Another not entirely related but perhaps of interest thing I learned in all of this - and this was more or less suggested to me by Thom Hogan of all people, so there's a good deal of weight to it - is that TTL flash in AF-C mode can cause problems. Trying to photograph people walking with TTL resulted in an >80% miss rate, almost as if the AF-C system is briefly "pausing" its operation when the pre-flash is fired or something and so its a little out of focus when the main flash fires. This should all happen pretty fast, so I'm not entirely sure on the logic, but suffice it to say that those same photos in manual mode resulted in everything working well. So as I was testing and preparing and shooting with subject detect and TTL flash I was getting almost nothing in focus at all, but shifting to dynamic area mode with manual flash I got most shots in focus.
What does all of this mean in practice? For both wildlife and human beings I have started working to "relearn" the old fashioned methods from the days of DSLRs and to use single point or dynamic area modes whenever I can and to rely on subject detect only when shooting something that I just can't do on my own, or that I think I'll get more reliably from the subject detect. So, for this photo I used single point:
View attachment 91448
While for this one I used subject detect:
View attachment 91449