It boils down to methodology used within the GPS receiver and processor. From general navigation perspective, GPS can provide extremely accurate velocity vectors. I'll spare the details as to why. But as an example, it can commonly be an order of magnitude more accurate than radar or the use of common ground based navaids.
However, if the GPS receiver and processor is designed to provide successive position information then the error budget goes up. The following factors come into play:
First, the position accuracy of GPS is not as accurate as the velocity information which could be available, and these errors are compounded with successive position determinations which might be used by the phone/tablet/device processor to determine an approximation of the velocity vector.
Secondly, the determination of velocity information with successive position processing may not utilize metadata which would be utilized with a dedicated navigation processor which has the goal of reliable velocity information. This means an increase in the error budget.
Thirdly, navigation devices which utilize GPS for aircraft, munitions and other applications, tend to perform continuous predictions as to the expected velocity, position, etc. in the future, so that the instantaneous readings from the device are a combination of prediction and historical data. In a tablet or smartphone, without knowing the attributes of the GPS and nav processor(s) one is likely looking at historical data, without the benefit of predictive models. This also increased the error budget.
So applied to the OP example, it is very conceivable that a reading inside the Faraday cage of a commercial aircraft, with limited sky view, multipath, and limited predictive modeling, and limited integration of sporadic readings, and without the benefit of strong predictive modeling, one may get readings that are substantially off from reality.
However, if one were using a GPS system designed for navigation, which exploited multi state Kalman filtering for velocity and position data, then I would expect a very highly accurate reading of velocity information. In the OP scenario, all of this does not exist.
So in the OP scenario, I would not expect a high level of accuracy and repeatability, and unfortunately, with the location/antenna/receiver/processor being used there is little that can be done to enhance the accuracy by calibration.
Addendum: One additional note. I don't know the current restrictions, but most GPS units have "speed restrictions" whereby they will not provide data above certain speeds (like Mach level speeds) which is done to limit their application to munitions. That may not be a factor in this case, but it is worth noting that all commercial GPS processors have this built in, if they are of US origin.