A human sees signs at 7s and 24s and the (admittedly not very clear) split of the lane at 27s as well as the yellow sign in the distance indicating the divider, while the autopilot sees a solid white line on the left and continues following it.
It probably would've been fine if there was another white line at the split, like this (marked in red here):
There is another white line at the split, following the trajectory you've drawn. At the junction of the two lines, it's badly faded, but you can see at :32 and better when he looks up after :34 that the line is there, it's just not as freshly painted.
That's not an unusual road condition at all; also: you'd think there'd be some logic reconciling the line it thinks it has to follow with the fact that an entire lane of traffic on the Ryan (the left-most I-94 lane) doesn't abruptly end.
The lane divider on the right is also quite visible; considering that its internal maps should tell it the road has a left exit, I'd expect that it should be able to guess that it should be holding to the right edge instead of the left edge of the lane.
This does bring up a theory: the lane keeping is biased towards holding itself to the left edge of lanes, on the basis that most exits are right exits and therefore the right edge will tend to diverge a ways before making a new lane. I'm curious how it would handle the split at the mixing bowl in Springfield, VA: https://www.google.com/maps/@38.7735345,-77.1816237,3a,19.8y... (functionally a left exit, although signed as if it were a right exit). The solid white line doesn't actually denote any lane exit boundaries there...
It probably would've been fine if there was another white line at the split, like this (marked in red here):
https://files.catbox.moe/10p2hh.JPG