Geometrically, linear approximation uses the gradient from a single point to create a tangent, and then approximates values from this tangent.
If the gradient of the function is increasing, then the tangent lies below the function and hence the approximation is lower than the actual value. This is certain as long as the gradient of the function continues increasing.
Similarly, if the gradient of the function is decreasing, then the tangent lies above the function and hence the approximation is greater than the actual value. This is certain as long as the gradient of the function continues decreasing.
It was best to draw a diagram for this: the function (which was

), and the tangent at x=2, then you can clearly show that the approximated value is greater than the actual value at x=2.05(from memory).