Modern eye and subject tracking autofocus is genuinely impressive enough that I trust it as my default for the vast majority of work now, and it still has specific, predictable failure modes I've learned to watch for rather than assuming it'll handle absolutely everything correctly without supervision.
Where it reliably struggles
Fast subjects moving directly toward or away from the camera at close range still occasionally lose tracking lock momentarily, particularly when a subject briefly turns their head enough that the system loses a clear eye to lock onto and has to re-acquire. Backlit subjects with strong rim light and a dark, underexposed face confuse tracking more often than a well-lit face does, the system seems to need at least some visible detail to confidently identify and hold an eye.
Group shots with similar-looking subjects
When multiple subjects in a frame look similar in size and position, a group of people at roughly the same distance from camera, tracking occasionally jumps between subjects unpredictably rather than holding steady on the one I actually selected. I've learned to manually confirm subject selection rather than trusting the system's automatic choice in genuinely crowded, similar-looking group scenarios, since a wrong automatic pick can ruin an otherwise perfect frame.
- Fast subjects moving directly toward camera: occasional momentary lock loss
- Strong backlight, underexposed faces: less reliable tracking confidence
- Similar-looking group subjects: manually confirm selection, don't trust auto pick
Why I still trust it as a default despite these limits
These failure modes are specific and predictable enough that I've built a mental checklist for when to double-check tracking versus when to fully trust it, and for the large majority of shooting situations outside those specific edge cases, it's more reliable than manual focus would be under the same time pressure. The technology earned my default trust. It didn't earn unconditional trust, and knowing the difference has saved a handful of otherwise-great frames from a tracking miss I caught in time to correct.