1 Truth. The real mug has no handle.
2 AI answer. The AI adds a handle because most mugs it saw during training had one, and it draws that handle exactly like the body, so nothing in the picture tells the viewer that it was made up.
3 What we want. The goal is a method that uses only the AI to mark each part of its answer. Teal marks parts where light came back, so the data supports the answer, and amber marks parts where no light came back, so the answer was made up.
Light travels from the laser to the wall, to the front of the mug, back to the wall, and back to the sensor.In view 3, teal marks parts the data supports, and amber marks parts the AI made up.Figure 1 shows a mug hidden behind a corner, where the sensor cannot see it directly. The sensor sends laser light to a visible wall, and the light that bounces off the mug and returns through the wall lets us reconstruct the front of the mug, which faces the wall. However, no light can return from the back of the mug, because the mug body blocks every path, so the measurements say nothing about it. Therefore, whatever the AI draws on the back, such as a handle, comes from what it learned rather than from the data, and we ask whether that part can be found using the AI alone.