There are plenty of other ways, like analog detection of torque as you describe - and using the cabin camera to figure it out (which it could do, without being able to see the hands or the wheel). A combination of these things has a very distinct pattern.
The position of your arms provides at best a ballpark guess. Some people drive with their hands practically in their laps, which would be hard to distinguish from
actually being in their laps without knowing the length of their forearms.
They are not the same at all. As long as they can measure different amounts of torque it should be easy to distinguish (pattern match).
Amounts of torque caused by an attached mass, whether a hand or a weight, are continually affected by outside factors — bumps, turning, acceleration, deceleration, inclines, etc. — all of which would have to be somehow factored in/canceled out, because otherwise they will all cause wide variations in the amount of torque applied by that mass.
And I'm not convinced there's that much difference in the rotational torque caused by the arm having pivots compared with a free-hanging weight of similar mass.
Another easy way is to look for both hands and also look for sustained torque on the wheel when both hands are on the wheel. That is a very good indicator of a high likelihood of a defeat device. Except for those who are able to put sustained torque on the wheel with their knee (a bit tough I think, though possible, and basically unnecessary in a Tesla). Won’t see this situation with most users (seeing both hands would be pretty rare), but it is just one way to get people on a naughty list.
It took me a third reading to realize that you must have meant "when both hands are
off the wheel".
But yes, in that particular situation, assuming you have cabin cameras, seeing significant torque while your hands are visibly away from the wheel would be easily detectable, assuming that it never confuses something else for your hands, such as a passenger's hands, patterns on your shirt or pants, a chicken sandwich that your kid threw at you from the back seat, etc.
You could calculate the dominant frequency response of the steering wheel to vertical impulses imparted by the road to the car body. With a non-symmetrical steering wheel, the wheel will twist back and forth slightly with every bump in the road. The frequency spectrum of this movement can be calculated and the standard deviation of the values determined. A wheel weight should have a very dominant frequency resonance with low standard deviation.
On average, I'd expect a human to have more of a bounce-back after an impact, but it depends on how tightly you're gripping the wheel, arm weight, how tense your arms are, etc., not to mention whether you let go of the wheel at those moments or hold on. You can't count on that. If there is any bounce-back, the frequency of that bounce-back would be basically dependent on the length of the person's arms and the wheel angle, and so probably won't vary much at all. But there may not be any bounce-back if the person drives more rigidly.
A human will almost certainly have a much more varied effect, spreading out the frequency response. It would be interesting to see the data in histogram form, but I suspect that a person could easily detect the difference between the wheel weight and human.
But the question isn't whether you could detect it or not, but rather whether you could detect it without false positives, which by their nature, are trust-destroying in both directions, in the single-digit minutes before the drive ends.