Black-Box vs White-Box Assessment

This distinction is the foundation of how you'll assess any AI system you don't build yourself — and of how you'll document the ones you do. Mastering it unlocks the rest of your Fairness and Bias track: Perturbation Robustness (how to probe a black box), Trust Rating Scales (how to express what you find), and Explainable AI (how to open the box when you can). It also decides every practical question you'll face: when you can audit a model yourself, when you must trust a vendor's word, and what evidence a regulator will actually accept.

Every system you encounter sits on a spectrum of how much you can see inside it. At one end is a white box: someone hands you the model's weights, the data it was trained on, the code that produced those weights, and the log of every step of training. You can open it, examine each part, and trace exactly why it produces the outputs it does. At the other end is a black box: you only have a way to send inputs in and receive outputs out. You cannot see the weights, the data, the training process — sometimes you don't even know what kind of model it is. Most real-world AI you'll encounter is somewhere in between. A vendor's API is nearly always a black box to you: you send it a prompt, it returns text, and you have no way to inspect the billions of numbers inside. Your own trained model is a white box to you, because you wrote the training script and can load the weights. The key insight is that you can still learn a lot about a black box — often enough to decide whether to trust it — without ever looking inside. You do this by sending it carefully chosen inputs and watching how the outputs change. The choice between black-box and white-box assessment is not about which is 'better'; it's about what is possible and what is sufficient for the question you're asking.

Think of a doctor assessing a patient's heart. A white-box assessment would be an open-heart examination: the doctor can see the heart directly, measure blood flow at every point, inspect each valve. A black-box assessment is what a doctor does in a routine check-up: they take your pulse (send an input), listen to your heartbeat (receive an output), and based on how the output responds to different conditions — exercise, rest, deep breathing — they infer whether something is wrong. They never open your chest. They don't need to, to answer the basic question 'is the heart functioning well?' The black-box doctor can even discover a problem, then refer you to a specialist who does the white-box examination to find the precise cause. The analogy holds for AI: black-box assessment answers 'does this system behave correctly and fairly?' White-box assessment answers 'why does this system behave this way, and where exactly is the flaw?' The analogy breaks down in one important way: a heart has a fixed structure that is the same across all humans, but two AI systems can have completely different internal architectures. A black-box assessment of one model's outputs tells you little about another model's internals, even if they perform similarly. The doctor's pulse-reading skill transfers across all hearts; your probing strategy for one AI model may not transfer to another.

Black-box assessment evaluates an AI system purely by its observed inputs and outputs, without access to its internal weights, training data, or code; white-box assessment evaluates it with full access to those internals.

You have not met the other concepts in this track yet, so we'll place this one alone for now. When you move on, you'll see that Perturbation Robustness — measuring how much a system's output changes when inputs are varied in ways that should not matter — is the central technique of black-box assessment. Trust Rating Scales — expressing an assessment as a comparable graded score — is what you produce after a black-box assessment. And the whole Fairness and Bias track builds on this distinction: detecting systematically different treatment across groups (fairness) often has to be done black-box, because you don't have access to the model's internals, only to its decisions.