Glossary

Reference intermediate

This glossary covers the core vocabulary used across the AI Product Design course. Terms are listed alphabetically. Each definition focuses on what the concept means in practice for a product team — not in academic AI research. Cross-references link to the lessons where each topic is explored in depth.

A

Term	Definition
Acceptance rate	The proportion of AI-generated suggestions a user accepts without editing. A high acceptance rate signals good model calibration; a very high rate can also mean users are rubber-stamping output without reviewing it — both extremes warrant investigation.
Adaptation strategy	The rules your product follows to adjust its behavior as it learns about a user — changing tone, complexity, defaults, or recommendations based on accumulated signals. Should be transparent and reversible.
Ambient AI	AI that works silently in the background without requiring direct user interaction. Spam filters, anomaly detectors, and content ranking algorithms are ambient AI. The design challenge is surfacing its work only when users benefit from knowing.
Appropriate trust	The ideal relationship between a user and an AI feature — neither dismissing correct AI output nor blindly accepting incorrect output. Good product design calibrates trust through confidence indicators, source attribution, and explicit uncertainty language.
Augmentation vs. automation	A spectrum describing how much human judgment remains in the loop. Augmentation keeps humans in control and uses AI to surface options; automation delegates the decision to the AI. Most features start at augmentation and move toward automation only after trust is established.
Autonomous agent	An AI that performs multi-step tasks independently, checking in with the user at defined points rather than after every action. Requires clear escalation paths and easy interruption mechanisms.

B

Term	Definition
Baseline comparison	A measurement of user performance without the AI feature, used as a reference point for evaluating AI-assisted performance. Without a baseline, you cannot know whether the AI is adding value.
Bias (product context)	Systematic errors in AI output that disadvantage certain user groups — often reflecting patterns in training data. A product responsibility, not just a model responsibility; teams must audit outputs across user segments before launch.

C

Term	Definition
Calibration	How well a model's expressed confidence matches its actual accuracy. A well-calibrated model that says "I'm 80% confident" is correct approximately 80% of the time. Poorly calibrated models undermine user trust even when they are often right.
Capability disclosure	Proactively telling users what the AI can and cannot do, so they form accurate mental models. The Google PAIR Guidebook recommends doing this at the moment of first use, not buried in documentation.
Choke point	A step in a workflow where AI errors cause disproportionate downstream damage — for example, an AI that incorrectly categorizes a support ticket before it routes to a human. Identifying choke points informs where human review must remain mandatory.
Confidence indicator	A UI element that communicates the model's certainty about a specific output — a percentage, a color band, hedging language ("I think…"), or a visual flag. Allows users to apply extra scrutiny where the AI is less certain.
Consent (AI context)	Informed agreement from users to have their data used by AI systems, to receive AI-generated content, and to have their behavior used for model improvement. Consent must be specific, revocable, and not buried in terms of service.
Contextual onboarding	Introducing an AI feature's capabilities at the moment they become relevant, rather than in a separate tutorial. Reduces cognitive load and improves adoption.
Correction interface	A UI pattern that makes it easy for users to fix AI mistakes — inline editing, "regenerate," alternative suggestions, or undo. Well-designed correction interfaces double as implicit feedback collection.

D

Term	Definition
Dark pattern (AI context)	A design choice that uses AI to manipulate, deceive, or exploit users against their own interests — manufacturing urgency, suppressing negative reviews, or making it artificially hard to opt out of AI tracking. Carries regulatory and reputational risk.
Data flywheel	The virtuous cycle where more user interactions generate more training signal, which improves model quality, which attracts more users, which generates more interactions. Products that establish a data flywheel early build durable competitive advantages.
Degraded experience	The version of a feature users receive when the AI component fails or is unavailable. Every AI feature needs a designed degraded experience — defaulting to search, static recommendations, or manual input — so the product remains usable.
Deterministic fallback	A rule-based or static response the product delivers when the AI cannot produce a reliable answer. Explicitly choosing a fallback prevents users from seeing raw model errors.

E

Term	Definition
Edit distance	A measure of how much a user modifies AI-generated output before accepting it. High edit distance suggests the AI is off-target; zero edit distance may indicate users are not reviewing the output carefully.
Equitable access	Designing AI features so they work equally well for all user segments — across languages, literacy levels, device capabilities, and connectivity. Features that perform well only for dominant user groups quietly widen existing disparities.
Error correction rate	The proportion of AI outputs that a user modifies or rejects. A leading indicator of model quality and UX fit; rising error correction rates after launch signal drift or data distribution shift.
Explainability	The degree to which a product can help users understand why the AI produced a specific output. Explainability is a spectrum: "Here are three sources" is more explainable than "here is your result." Full mechanistic transparency is rarely achievable or even useful; partial explainability is the practical goal.
Explicit feedback	User-initiated feedback such as thumbs up/down ratings, star reviews, or correction submissions. High signal quality but low participation rates; best used alongside implicit signals.
Exposure risk	The potential harm created when an AI feature reaches a large user population with an undetected error — a miscategorized group, a subtly biased recommendation, or a factually wrong statement repeated at scale.

F

Term	Definition
Feature adoption	The proportion of eligible users who try an AI feature at least once. Low adoption often signals a trust or discoverability problem, not a quality problem.
Feature retention	The proportion of users who continue using an AI feature after their first session. The primary indicator that the feature is delivering real ongoing value.
Feedback loop	The system by which user behavior and explicit input flow back into model improvement. A tight feedback loop is a competitive advantage; a broken one causes AI features to degrade over time.
Filter bubble	The narrowing of a user's information environment caused by personalization systems that only surface content matching existing preferences. Designers must decide how much to optimize for engagement vs. exposure to diverse perspectives.

G

Term	Definition
Graceful degradation	The principle that a product should become progressively less capable — but never outright broken — as conditions worsen (model unavailable, low confidence, edge-case input). Contrast with catastrophic failure, where the product silently returns a wrong answer.
Guardrail metric	A secondary metric that must not be harmed while optimizing a primary metric. In AI features, a common pair is: primary metric = task completion rate; guardrail metric = user correction rate. Guardrail metrics catch improvements that come at too high a quality or trust cost.

H

Term	Definition
Hallucination	An AI output that is factually incorrect but stated with apparent confidence. Hallucinations are a model-level phenomenon; hallucination UX is the product team's responsibility for communicating uncertainty and providing verification pathways.
Hallucination UX	Design patterns that reduce user harm from model hallucinations: hedging language ("I think…", "You may want to verify…"), source links, confidence indicators, and explicit uncertainty warnings. Not a substitute for model quality improvements, but an essential safety layer.
Hedging language	Phrases that communicate uncertainty in prose — "typically," "in most cases," "I believe," "you may want to verify." Used in AI interfaces to calibrate user trust without requiring numerical confidence scores.
Human-in-the-loop	A workflow design where a human reviews or approves AI outputs before they take effect. Required wherever the cost of an AI error is high; the goal for mature features is to reduce loop frequency as trust is established.

I

Term	Definition
Implicit feedback	Behavioral signals collected without asking users to do anything extra — click-through, dwell time, scroll depth, acceptance rate, session length. Lower signal quality than explicit feedback but far more abundant.
Informed consent	See Consent (AI context).
Inline suggestion	An AI assist delivered within the user's existing editing context — autocomplete text, a suggested reply, a proposed code completion. The user's workflow is minimally disrupted; they accept, dismiss, or modify the suggestion in place.

L

Term	Definition
Leading indicator	A metric that predicts future outcomes and can be acted upon early — for example, day-7 retention as a leading indicator of long-term feature success, or suggestion acceptance rate as a leading indicator of user satisfaction.
Latency tolerance	The maximum delay a user will accept before an AI response feels unacceptably slow. Context-dependent: users tolerate multi-second waits for a thorough research summary but not for an autocomplete suggestion.

M

Term	Definition
Mental model	A user's internal representation of how a system works. AI features fail when users have inaccurate mental models — expecting perfect memory, assuming the model "knows" their history, or believing the output is always factual.
Minimal footprint	A design principle for autonomous agents: request only the permissions needed for the current task, prefer reversible actions, and ask the user before doing anything with significant consequences. Borrowed from the Microsoft Human-AI Interaction guidelines.
Mixed-initiative interaction	A conversation or task flow where both the user and the AI can proactively take actions and make suggestions. Most chat-based AI interfaces are mixed-initiative; the design challenge is making it clear who is in control at each moment.

N

Term	Definition
Non-determinism	The property of AI systems where the same input can produce different outputs across runs. Makes traditional pass/fail testing insufficient; product teams need statistical evaluation, qualitative review, and A/B testing to assess AI feature quality.
NPS (AI context)	Net Promoter Score adapted for AI features. Useful for tracking satisfaction trends but insufficient on its own — combine with task completion and correction rates to understand why satisfaction is high or low.

O

Term	Definition
Opt-out design	Providing users with clear, low-friction paths to disable or bypass an AI feature. Required for ethical AI design and increasingly mandated by regulation; also prevents trust erosion when users feel trapped.
Output review	A workflow stage where a human checks AI-generated content before it is delivered or acted upon. Essential for high-stakes domains (medical, legal, financial) and for new features before trust is established.

P

Term	Definition
PAIR Guidebook	Google's People + AI Research Guidebook — a free, practitioner-focused resource with patterns, exercises, and examples for designing AI-powered products. A standard reference for the field; covers onboarding, errors, feedback, and user control.
Personalization boundary	The point at which personalization becomes uncomfortable or counterproductive for users — revealing that the system has inferred sensitive attributes, creating a filter bubble, or making users feel surveilled. Designing to respect this boundary requires explicit user controls and transparent explanations.
Product–model contract	The implicit promise a product makes to users about what the AI can do, how reliable it is, and what happens when it fails. Violating this contract — even subtly — erodes trust faster than any single error.
Progressive disclosure	Showing users only the information they need at each step, revealing more complexity as they need it. Especially important in AI features where full uncertainty information would overwhelm most users but some users need it.
Progressive personalization	Starting with a generic experience and adapting it gradually over time, using observed behavior rather than demanding upfront configuration. Reduces onboarding friction while building toward a tailored experience.
Progressive rollout	Releasing an AI feature to a small percentage of users first, expanding as confidence grows. Limits exposure to errors, enables A/B comparison against the control experience, and provides operational learning before full scale.
Prompt prototyping	Testing a user-facing AI feature concept by quickly building a rough LLM-backed demo — often with no production engineering — to observe how users react before committing to full development.

R

Term	Definition
RLHF (product context)	Reinforcement Learning from Human Feedback — the technique behind many instruction-tuned LLMs. At the product level, RLHF-style feedback collection means gathering preference pairs ("which response was better?") from users to improve future model behavior.
Regeneration	A UX pattern allowing users to request a new AI output for the same input when they are dissatisfied with the first. Provides a low-friction correction path and implicit feedback signal.

S

Term	Definition
Safety boundary	A hard constraint on what an AI feature will or will not do, regardless of user request. Safety boundaries are product decisions, not purely model decisions; teams must define them explicitly before launch.
Source attribution	Displaying the origin of AI-generated information — a document, a URL, a database record — so users can verify it independently. One of the most effective trust-building mechanisms in information-retrieval AI features.
Structured feedback	Feedback collected through guided flows that capture preference pairs, rankings, or labeled corrections rather than free-form ratings. Produces higher-quality training signal than unstructured thumbs-up/down.

T

Term	Definition
Task completion rate	The proportion of users who successfully accomplish their goal using the AI feature. The primary measure of whether the AI is solving the problem it was designed to solve.
Time-to-value	How quickly a user gets a useful output from an AI feature after they start using it. Features with long time-to-value lose users before they experience the benefit; progressive disclosure and smart defaults reduce time-to-value.
Trust calibration	The ongoing process of helping users develop accurate beliefs about when to rely on and when to verify AI output. Achieved through confidence indicators, transparent error handling, and consistent performance.
Trust spectrum	A model for understanding user attitudes toward an AI feature — from initial skepticism ("I'll try it once") through conditional use ("I use it for X but not Y") to dependency ("I can't work without it"). Design goals differ at each stage.

U

Term	Definition
Uncertainty language	See Hedging language.
User autonomy	The principle that AI should augment rather than replace human judgment and agency. Designing for autonomy means providing overrides, making opt-out easy, surfacing reasoning, and never manipulating users toward outcomes that benefit the product at their expense.
User model	The product's internal representation of an individual user — explicit preferences they have set, behaviors observed during sessions, and attributes inferred from those behaviors. Must be transparent, correctable, and privacy-respecting.

V

Term	Definition
Vanity metric	A metric that looks good in reports but does not indicate real value — for example, high AI feature usage that coexists with high correction rates and low retention. AI product teams must pair usage metrics with quality and satisfaction metrics.

W

Term	Definition
Wizard of Oz testing	A prototyping technique where a human operator secretly simulates the AI behavior while a test user interacts with what appears to be a real AI feature. Lets teams test UX and interaction patterns before any model is built, revealing whether the concept works at the product layer.

Glossary

A

B

C

D

E

F

G

H

I

L

M

N

O

P

R

S

T

U

V

W

Further Reading