He's saying that the dominant paradigm in the "soft" sciences is that you treat your subjects as black boxes performing semi-random transformation of inputs into outputs... without ever really trying to understand (in a completely reductionist way) what's going on inside the box.

The "hard" sciences don't work that way, of course: you don't need to test a thousand different pieces of iron and copper, just to get a statistical idea of which one maybe has a bigger heat capacity, for example.

To continue the analogy, it's as if the soft sciences have no calorimeters, thermometers, and scales with which to actually measure the relevant thing, and so instead are measuring something else that only weakly correlates with the thing we want to measure.

PCT, btw, proposes that behavior -- in the sense of actions taken by organisms -- is the "weakly correlated" thing, and that perceptual variables are the thing we actually want to measure. And that, with appropriate experimental design, we can isolate and measure those variables on a per-subject basis, eliminating the need to test huge groups just to get a vague idea of what's going on.

(One psychology professor wrote how, once he began using PCT models to design experiments, his results were actually too good -- his colleagues began advising him on ways to make changes so that his results would be more vague and ambiguous... and therefore more publishable!)