In theory it should be able to work out almost exact positions.
At the moment, you're stuck with frame rate, and having to choose between two frames, guessing which is the best. Even if the computer was assuming steady motion between each frame, then it would clearly be more accurate - after all the biggest margin of error is based on a runner at full speed, and those movements are going to the most predictable and smooth. The ones where the player is moving slowly or changing direction will be more complex, but of course already have a much smaller margin of error.
However AI should be much cleverer than assuming steady motion. It's not as if we're watching 22 Ian Curtis's jerking about randomly across the pitch. We're watching people who will speed up and slow down over multiple frames, who will be making relatively infrequent and fairly predictable changes in direction (infrequent in that they won't change direction multiple times a second, and predictable to a computer in hindsight).
Every movement in your body is signalled by movements elsewhere in the body - so if the AI has data from all over the body, it can predict what's expected to happen to other parts of the body. If you're going to change direction, or set off on a run, or make a feint, then different small movements all over your body will signal that, and a computer that's watched a million players doing exactly the same motion will be able to predict what happens next pretty accurately.