I have this problem quite regularly (if not exclusively), and it’s always the last observation (or occasionally last 2) on the page. I’m reviewing observations individually. This happens if I select “View more unreviewed” or if I reload the page (because selecting “View more unreviewed” is only helpful if you’re on Page 1, a separate annoyance I mention here). However, I’ve noticed a workaround: if instead of selecting “View more unreviewed” or reloading the page, I close the Identify “pop-up window” and reselect the last observation, then it realizes it’s reviewed and won’t stick around when I do select “View more unreviewed” or reload. (You can also hit the right button again and it brings you back to the last observation, so when you get to the end, hitting the right button twice is a fairly quick way to deselect and reselect.) Still annoying though.
A separate issue (but related?) issue I’ve had is sometimes the observations, one by one, unreview themselves. I think this might be related to my network connection, as if the page never recognized that I reviewed them to begin with. This is much more uncommon.