From the methods section, they're counting where it was on in up to the 5 seconds before the crash; I've no idea if that's reasonable or not, and it would be interesting to know what happens to the numbers if you push that up to 20s.
> To ensure our statistics are conservative, we count any crash in which Autopilot was deactivated within 5 seconds before impact, and we count all crashes in which the incident alert indicated an airbag or other active restraint deployed. (Our crash statistics are not based on sample data sets or estimates.) In practice, this correlates to nearly any crash at about 12 mph (20 kph) or above, depending on the crash forces generated.
> To ensure our statistics are conservative, we count any crash in which Autopilot was deactivated within 5 seconds before impact, and we count all crashes in which the incident alert indicated an airbag or other active restraint deployed. (Our crash statistics are not based on sample data sets or estimates.) In practice, this correlates to nearly any crash at about 12 mph (20 kph) or above, depending on the crash forces generated.