Human-Centered AI (HCAI) Mastery Course

A baseline is not a delay. It is the only thing that tells you what the model would actually be buying you.

The condition that decides: "Often a rule or heuristic-based solution will work just as well, if not better." Simpler is also easier to explain, debug and maintain.

Source: PAIR 1 — Decide if AI adds unique value

Before you tune anything, name the two people. The one your false positives hit, and the one your false negatives hit. They are never the same person.

The condition that decides: False positives and false negatives land on different people. Choosing the metric chooses the victim.

Source: PAIR 1 — Design & evaluate the reward function

Automate what people want rid of. Augment what they came to do. Accuracy tells you whether you CAN, never whether you should.

The condition that decides: "there are many activities that people want to do themselves." Capability does not decide it; desire to retain the task does.

Source: PAIR 1 — Assess automation vs. augmentation

The 18 guidelines describe an interaction that already exists, so they can never tell you not to build it. That is what PAIR chapter 1 is for.

The condition that decides: PAIR is the product arc; HAX is the interaction surface. They answer different questions and are routinely confused for competitors.

Source: PAIR 6 chapters × HAX 4 phases

A short recap of the conditions covered in this module, ahead of the scenario.

The condition that decides: When a user can act on being wrong, give them the action before the explanation.

Source: PAIR 1, 4, 5 × G9

Put capability and reliability in the same breath, before use. Keep the explanation for the moment something needs accounting for.

The condition that decides: 13 misinterpretations in the validation study, 8 between G1 and G11 alone. What it can do / how well / why it did that, after the fact.

Source: G1 · G2 · G11

Explain the benefit, not the technology — and explain it where the benefit actually appears, not on a screen before the product.

The condition that decides: "Explain the benefit, not the technology." And only introduce a feature when it is contextually relevant.

Source: PAIR 3 — Onboard in stages

If you cannot state the limit next to the voice, you have not earned the voice. Warmth is free; the first person is not.

The condition that decides: Anthropomorphism without explicit capability disclosure creates expectations the system cannot meet.

Source: PAIR 3 — human-like interaction

A short recap of the conditions covered in this module, ahead of the scenario.

The condition that decides: Reliability has to arrive where the reliance happens, and roughly once. Early it is forgotten; on every instance it becomes wallpaper.

Source: PAIR 3 × G1 · G2

Write down which guidelines do not apply, and why. Three states, not two: applied, violated, does not apply.

The condition that decides: G3 had among the highest "does not apply" ratings. For a pull interaction it is not a failure — "a very 'pull' kind of interaction."

Source: G3 — Time services based on context

They fail differently, so they are fixed differently. Misread the moment and you need a better signal; misread the person and you need more history, or a reset.

The condition that decides: Confused six times in the study. G4 is about the current task and environment; G13 is about accumulated behaviour.

Source: G4 vs G13

When the parties disagree, stop reaching for a norm and reach for a control. A default is a norm you imposed on whoever did not get asked.

The condition that decides: Also high on "does not apply" — because norms are plural, and the guideline does not say whose.

Source: G5 — Match relevant social norms

When a guideline cannot be evaluated by the method you are using, that is a finding about the method. Reach for measurement across people, not a tick.

The condition that decides: The authors' own limit: the guidelines "only begin to touch on topics of fairness."

Source: G6 — Mitigate social biases

A short recap of the conditions covered in this module, ahead of the scenario.

The condition that decides: You cannot time services based on context until you decide whose context. G3 sends you to G5; G5 sends you to a control.

Source: G3 × G5 × PAIR 5

Design trust per interaction, not per product. Ask what it costs to be wrong HERE, and let that set how confident the interface is allowed to sound.

The condition that decides: The goal is accurate trust. In one product, stakes vary per interaction, so a single explanation strategy is already wrong.

Source: PAIR 4 — Help users calibrate their trust · Account for situational stakes

Listen to the verb in the complaint. “How does it know?” is data sources. “Is it sure?” is confidence. “Why this one?” is example-based. “What would change it?” is interaction.

The condition that decides: Each answers a different user question. Picking one is picking which question you think they are asking.

Source: PAIR 4 — data sources · confidence · example-based · via interaction

Use the least granular form that still changes the decision. Usually categorical. Often nothing.

The condition that decides: "Counterintuitively, showing more granular confidence can be confusing if the impact isn't clear." If it changes no decision, it is noise.

Source: PAIR 4 — Determine if you should show confidence

Explain the decision, not the mechanism. The legitimate user needs enough to act; the attacker must not learn the rule. Those are separable.

The condition that decides: Two independent sources: explaining spam classification "could give spammers useful tips"; search engines explaining ranking would worsen gaming. G11 has a documented exception.

Source: PAIR 6 · CHI §Discussion

A short recap of the conditions covered in this module, ahead of the scenario.

The condition that decides: Where the line between decision and mechanism falls depends on who can read it and how often. A legitimate user asking "what stops this?" is asking the attacker's question.

Source: G11 · PAIR 4, 6 × G16

Before you ask how to fix it, ask whose wrong it is — and give context errors, where the model was right, an owner on purpose.

The condition that decides: They have different owners and different fixes. Calling them all "errors" guarantees the wrong team is assigned.

Source: PAIR 6 — user, system, and context errors

Before you build correction, check the prediction was actually wrong. Right but unwanted needs dismissal, and a dismissal that asks why has become a correction.

The condition that decides: Dismissal is the one teams skip, because it is the only one that assumes the feature was unwanted.

Source: G7 · G8 · G9

An empty review row is not a pass and not a fail. Give the review something to doubt, bring the model's own log, and record who can actually see it.

The condition that decides: Zero instances observed across both social networks tested — "hard to observe in a single session or without knowledge about the underlying AI algorithms." Some guidelines cannot be reviewed from outside.

Source: G10

The condition that decides: A correction only compounds if it reaches the model. Patching it silently teaches nobody; exposing the rule teaches one user; asking produces labels.

Source: G9 · G10 × PAIR 5, 6

Add a fourth column, can't see from here, and make every row in it name who can. Then send them to look.

The condition that decides: G10 and G14 produced zero outside observations, so a review staffed only by designers cannot assess them. Who is in the room is part of the method.

Source: Synthesis — CHI validation findings

Know the edge of your source. When the harm is in who the system fails, stop arguing about the row and ask who a false positive lands on.

The condition that decides: The authors' own limit — the guidelines "only begin to touch on topics of fairness and broader ethical considerations." Knowing the edge of your source is part of using it.

Source: CHI §Discussion