This is somewhat unavoidable and how we design intelligent interfaces is still in its infancy. That is, how much transparency do you expose to the user? And how do you effectively assist them in building a useful mental model of your system without being distracting or losing the convenience of using the model in the first place? But here, I am simply referring to designing constraints within the experience to increase the chances of the model's success. A great, although slightly unrelated, example of this is household robots and the dishwasher. Despite its non-robotic characteristics, the faithful dishwasher can be considered a first generation robot for household tasks, like Rosie from Jetsons. Unlike Rosie, however, we had not been able to get the dishwasher to adapt to our environment. So, we adapted the environment for the dishwasher, that is, we encapsulated it in a box environment rather than using the existing kitchen sink we're accustomed to.

One simple approach is making the user aware of how to achieve the best results; in this example, it could be as simple as asking them to use a wall as their background. These hints can be delivered before use or delivered when there are signs of poor performance (or both). One approach for automatically detecting poor performance would be to measure the bounding box and its center of mass, comparing it with the expected center of mass, as illustrated in the following figure:

Which brings us nicely to the next strategy: embedding heuristics.