Glossary

A* search: Best-first graph search using actual path cost plus a heuristic estimate toward the goal. (Hart et al., 1968; Jones, 2008)
Absolute Trust: Also known as complete trust. The trust that is put into an AI system that it will perform as intended without any verification or supervision, and will behave as expected and make proper choices consistently. (Laplante & Voas, 2022)
Absolute Trusting: Absolute Trusting is when the user takes the computer's assertions (data, claims) as valid and true in all circumstances. (Hoffman et al., 2021)
Accountability: the giving or demanding of reasons for conduct (León et al., 2020)
ACT-R: A computational cognitive architecture (Adaptive Control of Thought–Rational) used to model human performance; in trust research, applied to simulate strategic trust decisions in interaction.
Adaptive cruise control: Cruise control that adjusts speed using sensors to maintain distance from a lead vehicle.
Agent: The idea of an agent originated with John McCarthy in the mid-1950’s, and the term was coined by Oliver G. Selfridge a few years later, when they were both at the Massachusetts Institute of Technology. They had in view a system that, when given a goal, could carry out the details of the appropriate computer operations and could ask for and receive advice, offered in human terms, when it was stuck. An agent would be a “soft robot” living and doing its business within the computer’s world. (Kay, 1984; Bradshaw et al., 2017)
AGI (Artificial General Intelligence): The original use of the term "artificial general intelligence" was in a 1997 article about military technologies by Mark Gubrud (Gubrud, 1997), which defined AGI as “AI systems that rival or surpass the human brain in complexity and speed, that can acquire, manipulate and reason with general knowledge, and that are usable in essentially any phase of industrial or military operations where a human intelligence would otherwise be needed.” (Morris et al., 2023)
Agile development: Iterative software development approach emphasizing short cycles, working product increments, and inspect-and-adapt reviews. (Beck et al., 2001)
AI alignment (values): The goal that AI or automation has goals, actions, and values aligned with human intentions—supporting human consideration without necessarily behaving like humans. (Lutkevich, 2023)
AI Bias: Because of prejudiced assumptions in the process of developing the model, the output of a machine-learning model may cause discrimination against individuals or specific groups based on gender, social class, sexual orientation or race. However, bias means a deviation from the standard and does not necessarily lead to discrimination. (Belenguer, 2022)
AI-assisted decision making: Systems in which AI supports or augments human judgment rather than replacing it outright. (Zhang et al., 2020)
Algorithm: Any well-defined computational procedure that takes some value, or set of values, as input and produces some value, or set of values, as output.
Algorithm aversion: the reluctance of humans to use algorithms, which are superior but imperfect, in favor of their own judgment (Burton et al., 2020)
Algorithmic decision-making: the process of using predefined rules, procedures, or computational algorithms to analyze data and make decisions or predictions (Kochenderfer et al., 2022)
Alignment problem (AI safety): The technical challenge of ensuring AI systems pursue objectives that match human intent and values, especially under uncertainty, misspecified rewards, or unanticipated situations. (Strickland, 2023; Lutkevich, 2023)
Altruistic punishment: [Definition pending.]
Amershi HAI design guidelines: Eighteen evidence-based guidelines for human-AI interaction in four lifecycle contexts (Amershi et al., 2019). (Amershi et al., 2019)
ANI: An single-goal oriented AI element specialized to perform a specific task such as a chess engine, facial recognition software, or an email spam filter (Kuusi & Heinonen, 2022)
Anthropomorphic design: Designing AI to behave like humans—for predictability, acceptance, or accessibility. (Deshpande et al., 2023)
Anthropomorphic expectations/attributions: Users attribute human-like motivations or properties to AI, or expect human-like behavior from systems. (Mueller, 2020)
Anthropomorphism: The belief that a non-human entity is human, often through the use of human-like features by the entity. (Nass & Moon, 2000)
Anti-trusting: [Definition pending.]
Applied Cognitive Task Analysis (ACTA): A lightweight cognitive task analysis package that includes Knowledge Audit probes; useful for eliciting trust-related difficulties with equipment and systems. (Militello & Hutton, 1998)
Appropriate trust: [Definition pending.]
APT (Argument-based probabilistic trust): [Definition pending.]
Artificial Intelligence: A machine with the capacity to adapt given that proper data/information is accessible. (Wang, 2019)
Artificial Neural Network: a mathematical model that attempts to simulate the structure and functionalities of biological neural networks (Krenker et al., 2011)
ASI (Artificial Super Intelligence): A hypothesis of artificial intelligence, ASI, matches or surpasses the human mind and possesses capabilities exceeding humans. They are even capable of experiencing emotions, relationships, and generating their own code without the need for human intervention. (Keary, 2024)
Auditability: terms to a human
Authority Hypothesis: Humans percieve computational systems as being better skilled at specific tasks than them (Mosier & Skitka, 1996)
Authority-based trusting: [Definition pending.]
Automation: The process of structuring a task/environment/system to reduce labor and decision-making time through offloading labor and decision-making to a human-created system (Goldberg, 2011)
Automation Bias: When errors occur due to the overreliance on automated cues rather than observant seeking and processing
Back Propagation: a method for calculating all derivatives of a single target quantity with respect to a large set of input quantities (Werbos, 1990)
Backtracking (search): Search that extends partial solutions and retreats when constraints or goals fail. (Jones, 2008)
Backward chaining: Symbolic inference working from goals backward to find supporting facts and rules.
Bag-of-words model: Document representation as a multiset of word tokens, ignoring order; basis for many search indices.
Bayes Network: A concise and visual representation of the depencies between variables in a joint probability distribution. (Russell & Norvig, 2021)
Bayesian inference: Updating beliefs using Bayes' rule given prior probabilities and observed evidence. (Russell & Norvig, 2021)
Behavior Tree: A way for an autonomous system to switch between tasks, which allows for multiple decisions as data is collected. (Colledanchise & Ögren, 2018)
Beneficence: The ought to assist others for their own benefit, which includes prevention of harm. (Halsband & Heinrichs, 2022)
Bias in computer systems: Systematic and unfair discrimination against certain individuals or groups of individuals in favor of others within the design, implementation, or use of computational technologies. This bias can manifest in various forms, including preexisting bias, which originates from social institutions, practices, and attitudes; technical bias, which arises from technical constraints or considerations; and emergent bias, which emerges in the context of system use. (Friedman & Nissenbaum, 1996)
Black box: [Definition pending.]
Blackboard System: An AI approach where a central knowledge base (also known as the “blackboard”) is updated continuously with new data and information and applied to new problems. (Nii, 1986)
Blind trust: [Definition pending.]
Boundary conditions (automation): The dividing line between situations where an automated or AI system performs reliably and situations where it fails or should not be used; closely related to the competence envelope. (Hoffman et al., 2013)
Boundary conditions (theory): While theories provide answers to the “what,” “how,” and “why” questions, boundary conditions refer to the “who, where, when” aspects of a theory. (Busse et al., 2015)
CAPTCHA: Completely Automated Public Turing test to tell Computers and Humans Apart; challenge tests used to distinguish humans from bots.
CASA: A framework for understanding how humans treat computers as social agents when interacting with them. (Nass et al., 1994)
Case-based reasoning (CBR): Reasoning from a library of prior cases: retrieve, reuse, revise, retain. (Schank, 1983; Watson & Marir, 1994)
Chain-of-Thought Prompting: Providing an AI model with a series of intermediate reasoning steps to improve the model's accuracy (Wei et al., 2023)
Charrette: Intensive participatory planning workshops that engage stakeholders in joint problem-solving to build ownership and reduce conflict. (Lennertz, 2003)
Chess engines: A type of ANI that specializes in analyzing chess moves and pieces based on predefined moves and strategies to determine the move that will provide the best outcome (Newborn, 2014)
Classifier or Classification System: A supervised learning technique that uses predetermined classes or groups to assign an object based on its similarity to the groups; can also be used to predict the likelihood of an object fitting into a particular group. (Vilone & Longo, 2021)
Clumsy automation: Automation that helps during low-workload routine periods but increases human burden during critical peaks. (Wiener, 1989)
Clustering: Unsupervised grouping of similar data points into clusters. (Halkidi et al., 2001)
CO-STAR Method: A prompt template developed by GovTech Singapore’s Data Science & AI team that considers (C ) Context, (O) Objective, (S) Style, (T) Tone, (A) Audience, and (R ) Response (Teo, 2024)
Coactive Design: An approach to the design of Human-Robot Interaction that uses interdependence as the central principle among people and robots who work together in a joint activity (Johnson et al., 2014)
Coffee test: Wozniak's AGI challenge: enter an average house and make coffee using common-sense knowledge of objects and affordances.
Cognition-based trust: [Definition pending.]
Cognitive anthropomorphism: The expectation that AI will perform cognitive work in the same ways humans do. (Mueller, 2020)
Cognitive architecture: AI-hybrid models that embed known human cognitive capabilities and limitations to predict and explain human performance. (Newell, 1990; Byrne, 2003)
Cognitive Decathlon: A battery of varied knowledge and motor tasks matched to human performance without requiring human-like reasoning underlying performance. (Mueller & Minnery, 2008; Mueller, 2010)
Cognitive digital twin: A model used to predict human cognitive performance; closely related to computational cognitive models.
Cognitive Load Theory: Framework for designing instruction that manages intrinsic, extraneous, and germane cognitive load. (Sweller & Chandler, 1991)
Cognitive Systems Engineering (CSE): Engineering discipline treating humans as part of automated systems rather than merely operators of them. (Woods & Roth, 1988)
Cognitive Task Analysis (CTA): Structured methods for eliciting how people perform tasks and make decisions in context. (Crandall et al., 2006)
Cognitive Tutorial: An interactive learning tool that leads users through a sequence of interactive steps or activities to help them understand complex topics or processes. (Mueller et al., 2021)
Cognitive Work Analysis (CWA): Ecological framework for analyzing work domains to guide tool and automation design. (Vicente, 1999)
Cognitve Task Analysis: The determination of the cognitive skills, strategies, and knowledge required to perform tasks using specific methods and guidance to probe the cognitive processes. (Militello & Hoffman, 2008)
Collaboration lifecycle: Stage model of human–AI work spanning task allocation, interaction, feedback, and adoption risks as tools are fielded in teams. (Ahmed et al., 2026)
Collaborative Filtering: Collaborative filtering is a type of recommender system that can provide recommendations to users based on the ratings of similar users. (Su & Khoshgoftaar, 2009)
Competence (fidelity): Whether an AI can accomplish a task at all (Mueller fidelity dimension). (Mueller & Minnery, 2008)
Competence envelope: The range of situations in which a system performs reliably; users need to know when to trust or distrust the system.
Componential trust: [Definition pending.]
Computers in-the-loop: Computers act only as helpful tools to the humans who are tasked with accomplishing some goal. The technology is not the driving factor nor an equal party to the humans in the system. (Shneiderman, 2020)
Concept map: [Definition pending.]
Constraint Satisfaction Problem (CSP): Problem with variables, domains, and constraints; goal is to find assignments satisfying all constraints. (Jones, 2008)
Contestability: [Definition pending.]
Context window: [Definition pending.]
Contextual Design: User-centered process using ethnographic field studies to understand work practices. (Beyer & Holtzblatt, 1999)
Contextual trust: [Definition pending.]
Contingent Trusting: Contingent Trusting is when the user can take some of the computer's presentations or assertions as valid and true under certain circumstances. (Hoffman et al., 2021)
Contractual Trust: when a trustor has a belief that the trustee will stick to a specific contract; rust between a user and an AI model is trust that some implicit or explicit contract will hold, and a formalization of ‘trustworthiness’ (Jacovi et al., 2021)
Controller: [Definition pending.]
Cost Function: Cost functions are a measure of undesirable a particular path is within a graph. The less desirable (due to distance, financial cost, complexity), the higher the cost. (Jones, 2008)
Counter-trusting: [Definition pending.]
Critical Decision Method: Interview method used in naturalistic decision making to elicit expert judgments about critical incidents. (Klein, 1998)
Critical periods (HCAI lifecycle): Stages in AI development when specific human-centered methods have maximal leverage and cost-benefit.
Cybernetics: Study of control and communication in animals and machines; precursor to cognitive systems engineering. (Wiener, 1948)
Dark patterns: User-hostile design patterns intended to deceive, confuse, or extract money from users; also transferred to conversational AI. (Lacey & Caudwell, 2019; Stockman & Nottingham, 2024; Traubinger et al., 2024; Dubiel et al., 2024)
Data Card: Data Cards are structured summaries of essential facts about various aspects of ML datasets needed by stakeholders across a dataset’s lifecycle for responsible AI development. These summaries provide explanations of processes and rationales that shape the data and consequently the models—such as upstream sources, data collection and annotation methods; training and evaluation methods, intended use; or decisions affecting model performance. (Pushkarna et al., 2022)
Deep learning: Neural networks with many layers trained on large datasets, often via backpropagation on GPUs. (LeCun et al., 2015)
Default trusting: [Definition pending.]
Defensive monitoring: Supervisory checking of automation when reliability is uncertain; treated as a behavioral indicator inversely related to trust. (Muir & Moray, 1996; Adams et al., 2003)
Dependability: In interpersonal and automation trust models (e.g., Rempel; Muir), a mid-stage basis for trust: confidence that the trustee will behave consistently and can be counted on, beyond short-term predictability. (Rempel et al., 1985; Muir, 1994)
Design thinking: Human-centered innovation process that cycles through empathize, define, ideate, prototype, and test. (Brown, 2008)
Desirable difficulty: Learning design that introduces manageable challenge to improve long-term retention and transfer. (Bjork & Bjork, 2020)
Desirements: Stakeholder wants and requirements for AI systems, distinct from formal specifications or regulations. (Hoffman et al., 2023)
Digital Twin: [Definition pending.]
Digressive Trusting: Digressive Trusting is when the user takes fewer of the machine’s presentations or assertions as valid and true over time and across experiences. (Hoffman et al., 2021)
Dijkstra's Algorithm: An algorithm that takes a graph as an input and finds the minimum cost between all pairs of nodes within the graph. (Dijkstra, 1959)
Dimensionality reduction: Transforming high-dimensional data into fewer dimensions while preserving essential structure. (Baruah, 2023)
Directability: In coactive design, the ability of teammates to influence or redirect one another's behavior. (Johnson et al., 2014)
Distributed Cognition: Thinking spread across people, tools, representations, and environmental structures rather than confined to individual minds. (Hutchins et al., 2000; Hutchins, 1995)
Distributional verisimilitude: Whether the distribution of AI performance matches the distribution of human performance (Mueller fidelity dimension). (Mueller & Minnery, 2008)
Distrust: [Definition pending.]
Disuse (automation): Failure to use automation that would be beneficial, often linked to distrust or undertrust relative to system capability. (Parasuraman & Riley, 1997; Lee & See, 2004)
Domination (fidelity): Whether an AI can perform a task better, faster, or cheaper than a human (Mueller fidelity dimension). (Mueller & Minnery, 2008)
Ecological Interface Design: [Definition pending.]
Effective anthropomorphism: Users initially apply human capabilities to understand an unfamiliar AI system—a useful strategy even when the system is not human-like. (Bos et al., 2019)
ELIZA: Early pattern-matching chatbot that played a Rogerian therapist, mirroring the client rather than giving direct advice. (Weizenbaum, 1966)
Emotional Connection: A social-anthropomorphic design dimension describing commonly reported emotional engagement with AI. (Gibbons et al., 2023)
Emulation (human-like AI): The design goal of creating AI that emulates human capability rather than leveraging distinctive machine strengths. (Shneiderman, 2020)
Encapsulated trust: Trust limited to known conditions within a system's competence envelope rather than blanket acceptance. (Hoffman et al., 2013)
Ethopoeia: responding to a non-human entity as if it were human while simultaneously knowing the entity does not require human-style interaction. (Nass & Moon, 2000)
Evidence-based licensure: [Definition pending.] (Tate et al., 2016)
Evolutionary algorithm: Search inspired by natural selection: evaluate fitness, select, recombine, iterate. (Russell & Norvig, 2021)
Experiential User Guide: Interactive training letting users learn AI capabilities through guided experience rather than static documentation. (Mueller et al., 2009; Mueller et al., 2021)
Expert System: A high-perfoming AI program that can generate domain-specific expert level solutions, which might aid in problemsolving in a given area. (Buchanan, 1981)
Explainability: [Definition pending.]
Explainable AI: XAI can explain their decision or predictions to human users to increase transparency, trustworthiness and accountability of system. (Saranya & Subhashini, 2023)
Faith-based trusting: [Definition pending.]
Feed-Forward Neural Network: A neural network whose connections only go in one direction; can be thought of as a directed acyclic graph with designated input and output nodes. (Russell & Norvig, 2021)
Feigenbaum Test: A Turing-style test targeting domain experts (e.g., engineering, physics), emphasizing reasoning over breadth of knowledge. (Feigenbaum, 2003)
Few-shot prompting: Providing a small number of instructions or examples to an AI model to guide it in performing a task (Guide, 2024)
Filter bubble: [Definition pending.]
Fine-tuning: There are two steps in our framework: pre-training and fine-tuning. During pre-training, the model is trained on unlabeled data over different pre-training tasks. For fine-tuning, the BERT model is first initialized with the pre-trained parameters, and all of the parameters are fine-tuned using labeled data from the downstream tasks. Each downstream task has separate fine-tuned models, even though they are initialized with the same pre-trained parameters. (Church et al., 2021; Devlin et al., 2019)
formative assessment: Lightweight early evaluation used to steer design and development rather than only to demonstrate final effectiveness.
Forward chaining: Symbolic inference applying rules to known facts to derive new conclusions (data-driven).
Functional mental models: [Definition pending.]
Functional understanding: Working knowledge of a system sufficient for effective use without deep technical mastery. (Mueller et al., 2009)
Functionality (anthropomorphism space): Capability, accuracy, or ability of AI to accomplish tasks (Gibbons anthropomorphism dimension). (Gibbons et al., 2023)
GAMs: Generalized additive models for interpretable machine learning (Chang et al., 2021)
General Problem Solver (GPS): Early AI system searching a problem space via subgoals and operators. (Ernst & Newell, 1967)
Generate and test: CSP approach that proposes candidate solutions and checks whether constraints are satisfied. (Jones, 2008)
Generative AI: A branch of artificial intelligence that receives user input (such as a prompt) and produces an output based on training data. Outputs of generative AI are vast but include text, audio, video, images, etc. (Brynjolfsson et al., 2023)
generative model: An approach to statistical classification in which the underlying patterns of the existing data are examined to create new data similar to the current data in the dataset. (Xu et al., 2015)
Genetic algorithm: Evolution-inspired search using selection, crossover, and mutation over a population of candidate solutions. (Holland, 1992)
Glass Box Model: Glass box models are transparent about how a conclusion was reached and provide an accurate and human-interpretable explanation. (Carroll & Olson, 1988)
Good regulator theorem: Every good regulator of a system must be a model of that system (Conant and Ashby). (Conant & Ashby, 1970)
Gradient descent: Iterative optimization that adjusts parameters in the direction that reduces error or cost.
Hallucination: A phenomenon that LLMs occasionally generate content that diverges from the user input, contradicts previously generated context, or misaligns with established world knowledge (Zhang et al., 2023)
heuristic: A heuristic is a rule of thumb that may help solve a given problem. Heuristics take problem knowledge into consideration to help guide the search within the domain. (Jones, 2008)
Human Readiness Level (HRL): Nine-level scale characterizing how thoroughly human-centered design and evaluation have matured for a system alongside technological readiness. (See et al., 2019; Human Factors and Ergonomics Society, 2021)
Human–Computer Trust (HCT): Madsen and Gregor’s 25-item scale covering perceived reliability, technical competence, understandability, faith, and personal attachment in human–computer trust. (Madsen & Gregor, 2000)
Human-AI collaboration: Research and design perspective treating humans and AI as cooperative partners on shared tasks. (Wang et al., 2020)
Human-AI Interaction (HAII): Interdisciplinary field focused on how people interact with AI-capable systems through interfaces, workflows, and design. (Amershi et al., 2019; Xu et al., 2021)
Human-AI teaming: Design and research framing humans and AI as teammates with complementary roles on shared goals. (Berretta et al., 2023)
Human-automation interaction: Study of how people interact with automated systems, including supervision, monitoring, and handoff. (Sheridan & Parasuraman, 2005)
Human-Autonomy Teaming: Team-focused approach treating automation or AI as a teammate with shared goals and complementary capabilities (also HAT/HMT). (Lyons et al., 2019; Berretta et al., 2023)
Human-centered AI: A method of thinking about, talking about, and designing AI that focuses on people working together, in charge of technology, and benefiting from information displays; a method of designing AI that amplifies human agency and ability (Shneiderman, 2021)
Human-in-the-loop: the AI system is able to learn from the human agent, in some capacity, to fill in holes of the knowledge base (Mosqueira-Rey et al., 2023)
Human-in-the-loop ML: Human-in-the-loop machine learning is a way to involve human operators in the learning process, such that the AI learns in some capacity from the human operator, such as from labeling data or learning from the interactions that a human has with the system (Mosqueira-Rey et al., 2023). Many researchers are attempting to use human-in-the-loop systems to deal with sparse data, as these systems are able to use the prior knowledge that humans have to supplement data (Wu et al., 2022). As AI systems are used to make more important decisions, explainable AI focuses on a human user’s ability to understand how an AI system comes to a conclusion (Dwivedi et al., 2023). (Mosqueira-Rey et al., 2023)
human-machine teaming: The collaboration between humans and intelligent systems to accomplish a shared goal. (Paleja et al., 2021)
Human-Robot Interaction (HRI): Study of how people perceive, communicate with, and work with physical robots across industrial, social, and autonomous forms. (Goodrich & Schultz, 2008)
Humanization: Treating non-human entities as human in terms of rights, expectations, or interactions. (Spatola et al., 2022)
Hybrid Intelligence: Hybrid intelligence attempts to utilize the strengths of human intelligence and AI, with the overall goal of creating a better system than either individually (Dellermann et al., 2019). This idea builds on the system 1 and system 2 theory of cognitive processes (Kahneman, 2011). System 1 thinking is fast, emotional, stereotypic, subconscious, and often is considered human intuition. This is where human intelligence is stronger. System 2 thinking is effortful, logical, and often follows strict rules. This is where AI is strongest. By pairing the human and AI together, both can focus on what they are strongest at and create better solutions. (Dellerman, 2019)
ICAP framework: Learning taxonomy: Interactive, Constructive, Active, and Passive engagement modes (Chi & Wylie). (Chi & Wylie, 2014)
IEEE P7007 Standard on Transparency: [Definition pending.]
Imitation Game: Turing's original setup: a judge distinguishes two interlocutors via text; in the intelligence version a machine replaces the deceptive participant. (Turing, 1950)
In-context learning in LLMS: [Definition pending.]
Information retrieval: Systems that index, rank, and retrieve documents or content relevant to a query.
information transparency: The disclosure of information regarding a particular AI system, especially for reasons revolving around ethics and legality. (Andrada et al., 2022)
Intelligence: The ability to learn from and adapt to an environment. (Sternberg, 2012)
Interactive Human-Centered AI (IHCAI): HCAI practice enabling users to explore, manipulate, and understand AI outputs with human benefit as the design goal. (Schmidt, 2020)
Interactivity (anthropomorphism space): Ability to interact with the user and respond to user feedback (Gibbons anthropomorphism dimension). (Gibbons et al., 2023)
Interpretability: ability to explain or to present in understandable (Doshi-Velez & Kim, 2017)
Ironies of automation: Automation can leave humans with harder residual tasks (monitoring, exceptions) and erode skills needed when manual control is required. (Bainbridge, 1983)
Jagged Technological Frontier: [Definition pending.] (Dell'Acqua et al., 2023)
Joint activity: An activity involving two or more parties are participating (Johnson et al., 2014)
Joint cognitive systems: Perspective analyzing distributed cognitive work across people, tools, and automation in a joint system. (Hollnagel & Woods, 2005)
Justifiability: The need for an AI system to provide reasoning for its decisions/recommendations. This is especially important when determining why a decision was made by a system and whether the outcome aligns with human morals. (Muralidharan et al., 2024)
Justified mistrusting: Appropriate skepticism toward a system when its limits or failure modes are understood. (Hoffman, 2017)
Justified trusting: Calibrated reliance on a system based on understanding its competencies and limits (Hoffman et al.). (Hoffman, 2017)
Knowledge elicitation bottleneck: Difficulty of extracting expert knowledge to build rule-based or expert systems. (Hayes-Roth et al., 1983)
Knowledge transfer and exchange (KTE): the process of exchanging relevant and meaningful knowledge between different agents (Kiefer et al., 2005)
leakage: in AI, the use of information from outside the training data set; most often occurs when data in the training, vaidation, and/or testing share indirect information (Bussola et al., 2021)
Lens model: [Definition pending.]
Levels of automation: Scales distinguishing how much of information acquisition, analysis, decision selection, and action implementation is automated. (Sheridan & Verplank, 1978; Parasuraman et al., 2000)
LIME: Local Interpretable Model-Agnostic Explanations. An algorithm that can explain the predictions of any classifier or regressor in a faithful way, by approximating it locally with an interpretable model. (Ribeiro et al., 2016)
LLMs: Large language models (LLMs) refer to transformer-based neural language models that contain tens to hundreds of billions of parameters, and which are pre-trained on massive text data, e.g., PaLM24, LLaMA12, and GPT-4 (Singh et al., 2024)
Logistic regression: Classification method modeling the probability of a binary outcome rather than a continuous value.
Loss Function: The loss function (cost function) is used for training a neural network or other machine learning models. Loss functions determine the performance of a deep neural networks. The same framework of deep CNNs with different loss functions may have different training results. According to the application scenario or properties of a loss function, we can design or select a loss function. Since an appropriate loss function will boost the whole performance of a neural network, the traditional machine learning losses are falling out of favor and their counterparts in deep learning are gaining traction. (Tian et al., 2022)
LSA: A natural language processing technique that involves generating a set of concepts associated with documents and terms to analyze the relationships between a set of documents and the terms they contain. (Dumais, 2004)
Machine Learning: Algorithms that can learn from data, model and make predictions/decisions independently based on these models. (Kufel et al., 2023)
Macrocognition: Cognitive processes used in natural settings (Crandall et al., 2006)
Markov model: Statistical model where the next state depends only on the current state; basis for n-gram language models. (Shannon, 1951)
material transparency: An understanding of the physical system or systems that an AI runs upon, including energy requirements, climate impact, and other socioeconomic factors. (Andrada et al., 2022)
Mayer-Hancock trust framework: [Definition pending.]
Mechamorphism: As opposed to anthropomorphism, where a human being imparts human characteristics to a non-human actor, mechanomorphism is where a human being attempts to modify their thinking to be more consistent with a non-human actor. (Glasgow et al., 2022)
Mechanistic mental models: [Definition pending.]
Mental Model: An internalized representation of a situation or individual founded on previous experiences. (Langan-Fox et al., 2004)
Mental Model Matrix: Structured evaluation of user and system capabilities and limitations (Borders et al.). (Borders et al., 2024)
Mental Model of a Dynamic System: "A mental model of a dynamic system is a relatively enduring and accessible but limited internal conceptual representation of an external system whose structure maintains the perceived structure of that system." (Doyle & Ford, 1998)
Mentalism: Attributing an agent's behavior to internal mental, emotional, or cognitive states. (Spatola et al., 2022)
Metaphor Model: [Definition pending.] (Carroll & Olson, 1988)
Microcognition: Fundamental cognitive processes underlying other, more complex forms of cognition (Crandall et al., 2006)
Microworld: A small closed-loop simulation that captures some complexity of a work domain (e.g., process control) without fully simulating one plant or operational system. (Adams et al., 2003)
Minimum necessary rigor: [Definition pending.] (Klein et al., 2023)
Mistrusting: [Definition pending.]
Misuse (automation): Inappropriate reliance on automation, often linked to overtrust—using a system beyond what its capabilities warrant. (Parasuraman & Riley, 1997; Lee & See, 2004)
Mixed Initiative Interaction/Planning: A method of planning that involves a general understanding of contributing sub-tasks between constituents to reach a shared outcome. (Novick & Sutton, 1997)
Model Card: [Definition pending.]
Monte Carlo Sampling: A computational technique used to predict probabilities of various outcomes relying on repeated random sampling (Kroese et al., 2014)
Moravec's paradox: It is easy to give machines adult-level performance on tests and games, but hard to give them the perception and mobility skills of a one-year-old. (Moravec, 1988)
Mutual Theory of Mind (MToM): [Definition pending.]
N-gram: Sequence of N consecutive tokens used in language modeling and predictive text. (Shannon, 1951)
natural language processing: Natural language processing (NLP) is a collection of computational techniques for automatic analysis and representation of human languages, motivated by theory. (Chowdhary, 2020)
Naturalistic Decision Making: Study of how experts make judgments under time pressure, uncertainty, and real operational constraints. (Klein, 1998)
Observability: The ability to understand how complex machine learning systems make decisions and interpret data by observing their external outputs. (Watts, 2023)
One-Shot Prompting: [Definition pending.]
Optimization: the systematic and iterative process of improving the accuracy of a machine learning model or algorithm’s output through testing and refinement (Surianarayanan et al., 2023)
Oracle-based testing: [Definition pending.] (Tate et al., 2016)
Organizational trust: [Definition pending.]
Overfitting: Model parameters fit training data so closely that performance degrades on new cases outside the training set.
Overtrust: [Definition pending.]
PageRank: an algorithm used to measure the "weight" of nodes within an artificial neural network; weight is based on the number of incoming relationships the node has, as well as the importance of the corresponding sources nodes (Langville & Meyer, 2006)
Participatory design: Design approach embedding users in the design team or designers in the work context.
Pathfinder networks: [Definition pending.]
PCA: Principal Component Analysis (PCA) is used for dimensionality reduction in multivariate data and is a widely applied technique in data analysis and machine learning (Baruah, 2023). Its objective is to transform high-dimensional data into a lower-dimensional representation, capturing the most important information. (Baruah, 2023)
Perceived trustworthiness: [Definition pending.]
PID controller: Proportional-Integral-Derivative controller regulating a process using three tuned feedback components. (Johnson, 2005)
Pre-training: There are two steps in our framework: pre-training and fine-tuning. During pre-training, the model is trained on unlabeled data over different pre-training tasks. For fine-tuning, the BERT model is first initialized with the pre-trained parameters, and all of the parameters are fine-tuned using labeled data from the downstream tasks. Each downstream task has separate fine-tuned models, even though they are initialized with the same pre-trained parameters. (Church et al., 2021; Devlin et al., 2019)
Predictability: In coactive design, the ability of teammates (human or machine) to anticipate one another's actions and needs. (Johnson et al., 2014)
Premortem: Planning exercise in which a team imagines that a project has already failed and generates reasons why, in order to surface risks early. (Veinott et al., 2010; Kannan, 2026)
Procedural Justice: fair resolution of disputes and application of administration (Ötting & Maier, 2018)
Production-Rule System: A set of rules for fetching data from a database. These production rules are bound to precondition and postcondition that gets checked by the database. If the input content satisfies the conditions, the rules will be successfully applied, resulting in corresponding outcomes. (Khan, 2024)
Progressive Trust: a positive evolution of trust between two subjects based on interactions. An initial trust level evolves based on interactions between the two subjects, starting with a test period, trusting each other in situations in which the risks and payoffs are relatively large, and based on the feedback the trust between parties increases or decreases. (Huerta-Canepa et al., 2011)
Progressive Trusting: Progressive Trusting is when the user takes more of the machine’s presentations or assertions as valid and true over time or across experiences. (Hoffman et al., 2021)
Prompt: An instruction from the user that the LLM follows (april2024)
Prompt Engineering: The systematic process of generating and enhancing prompts with the intentionof eliciting targeted and favorable responses from large language models LLMs (Aljanabi et al., 2023)
Propensity to trust: A trustor disposition or trait-like willingness to trust others (or systems) across situations; an antecedent in Mayer-style organizational trust and Hancock human–machine trust models. (Mayer et al., 1995; Hancock et al., 2011)
Q-learning: Model-free reinforcement learning algorithm estimating action values in states. (Li, 2023)
RAG in LLMs: [Definition pending.]
rational agent: For each possible percept sequence, a rational agent should select an action that is expected to maximize its performance measure, given the evidence provided by the percept sequence and whatever built-in knowledge the agent has. (Russell & Norvig, 2021)
Recognition-Primed Decision (RPD): [Definition pending.]
Recommender System: A system (generally machine learning or algorithm-based) that reads prior usage data of a platform or environment (Amazon, Netflix, Spotify) and compares it against demographic data to provide future predictions of behavior (vultureanualbii2021)
Recurrent Neural Network: A neural network in which cycles between nodes are allowed, allowing the network to have a form of memory where inputs from earlier in time can impact the current work being performed. (Russell & Norvig, 2021)
reflective transparency: The ability of a user to have understanding, oversight, or ability to intervene in the decision-making processes of an AI system (Andrada et al., 2022)
Reinforcement learning: Learning policies through trial-and-error interaction with an environment and reward signals. (Sutton & Barto, 2018)
Reinforcement Learning from Human Feedback (RLHF): A procedure to produce or ensure AI alignment through human feedback on outputs from the AI. (Strickland, 2023)
Reliability: Systems are based on proven methods and practices; appropriate technological processes support human responsibility, fairness, and explainability (Shneiderman, 2020)
Reliance: Making decisions and acting through AI. Over-reliance on AI can lead to a lack of professional skills or AI errors that are difficult to correct. Thus, it is important to have appropriate supervision and human intervention when relying on AI. (Schemmer et al., 2023)
Resemblance (fidelity): Whether AI behavior reproduces qualitative trends of human performance (Mueller fidelity dimension). (Mueller & Minnery, 2008)
Resilience Engineering: Discipline focused on how people and organizations monitor, adapt, and recover under surprise, stress, or failure. (Hollnagel et al., 2006)
Reversal Curse: [Definition pending.] (Marcus, n.d.)
Reverse Turing test: The user's prompts and subjective intelligence make an LLM appear more intelligent. (Sejnowski, 2023)
Rule-based system: An expert system that records data as rules to solve problems (Atanasova & Karashtranova, 2016)
SAE J3016 (Levels of Driving Automation): Six-level taxonomy from fully human-driven to fully autonomous vehicle operation. (SAE On-Road Automated Vehicle Standards Committee, 2014)
Safety: [Definition pending.]
Scalable Human Oversight: integrating human judgment and intervention in the development, deployment, and continuous monitoring of AI systems (Shneiderman, 2020)
Scrum: Agile framework characterized by short sprints, defined roles, and regular inspect-and-adapt reviews to deliver incremental improvements. (Schwaber & Sutherland, 2020)
Self-Exfiltration: [Definition pending.]
Sentiment analysis: Classification of text by overall tone or sentiment. (Medhat et al., 2014)
SHAP: An algorithm that generates values that attribute to each feature the change in the expected model prediction when conditioning on that feature. (Lundberg & Lee, 2017)
Shapley Value: A method based on game theory for assigning payouts to players depending on their contribution to the total payout (Molnar, 2023)
Shared mental models: [Definition pending.]
Shared responsibility: Division of control and accountability between human operators and automated or AI components. (Shneiderman, 2020)
Shifting bottlenecks: When automation displaces work, new constraints and workload peaks often emerge elsewhere in the system. (Sarter et al., 1997; Woods, 1996)
Simulacra: Elaborate moving sculptures depicting animals, birds, or humans; often meant to amaze rather than replace humans. (Graefe & Bischoff, 2009; Carrier, 2017)
Situated cognition: View that cognition arises from interaction among agents and their environment, not isolated information processing. (Clancey, 1997)
Situation awareness: Perception, comprehension, and projection of elements in the environment relevant to goals (Endsley). (Endsley, 1995)
Skeptical trusting: [Definition pending.]
SLAM: Simultaneous Localization and Mapping: robot builds a map while estimating its position. (Huang, 2023)
Social Actor: Any individual, artificial or physical, that can estabilsh a general understanding with humans while carrying out a conversation. (Nass et al., 1994)
Social Robotics: A field of research that seeks to improve robot abilities to mimic human behavior, for example gaze behaviors. (Ruhland et al., 2015)
Socio-technical system: Work system comprising people, technology, organization, and culture as an integrated whole. (Clancey, 1997)
Sociotechnical systems: Work settings jointly shaped by technical and social subsystems that must be designed together for automation or AI to succeed. (Baxter & Sommerville, 2011)
Soft-max confidence: Classifier output scaled to resemble probabilities; often misread as true confidence by users.
Spiritualism: Attributing a spiritual nature to non-human agents. (Spatola et al., 2022)
SQA (Social Q&A): the process of people asking questions, answering questions, and rating the content of other answers (Gazan, 2011)
Stable trusting: [Definition pending.]
Stakeholder alignment (adaptive management): Adaptive management approach aligning diverse stakeholder needs during AI development, deployment, and assessment (Wixom et al.). (Wixom et al., 2020)
Stakeholder analysis: Identifying and eliciting needs, goals, and constraints of people affected by or involved in an AI system. (Hoffman et al., 2023; Plass et al., 2022)
Stakeholder Playbook: Structured prompts and interview guide for stakeholder desirements in human-AI explainability (Hoffman et al.). (Hoffman et al., 2023)
Stochastic parrot: Metaphor for language models that produce fluent text via statistical pattern matching without deeper understanding. (Bender et al., 2021)
Strong AI: Strong AI (also known as Artificial General Intelligence or AGI) is a concept where artificial intelligence possesses a mind similar to humans and can think and reason like humans. The robust and adaptive behavior of Strong AI allows it to perform tasks in various different situations, achieving a level of comprehension or execution of complex tasks that Narrow AI cannot accomplish. (Ng & Leung, 2020)
Structural mental models: [Definition pending.]
Supervised learning: Machine learning from labeled examples to predict outputs for new cases. (El Naqa & Murphy, 2015)
Supervisory control: Human role overseeing automated operation: monitoring, troubleshooting, and intervening when needed. (Sheridan, 1992)
Surrogate Model: [Definition pending.] (Carroll & Olson, 1988)
SVD: Singular Value Decomposition (SVD) is a common technique to decompose a matrix into several component matrices, exposing many of the valuable properties of the original matrix (Chengwang L., 2010) (Chengwang, 2010)
Swift Trust: In temporary groups, people need to work on same tasks. However, they don’t have prior history of collaboration, experiences, or interactions to judge each other's. In this situation, they will develop swift trust. Traditionally, swift trust has been applied to human-human trust. Now, there is a need to understand how swift trust is developed for human-robot teams given the proliferation of robots for team tasks. (Meyerson et al., 1996; Haring et al., 2021)
Symbolic AI: AI using explicit symbolic tokens, rules, and logical pattern matching.
Technological Readiness Level (TRL): Nine-level scale of technology maturity used in R&D acquisition, from basic research through operational system use. (Mankins, 1995)
Technology Acceptance Model (TAM): Model of technology adoption in which perceived usefulness and ease of use shape attitude toward using and actual system use.
Tentative trusting: [Definition pending.]
Theory of Mind: [Definition pending.]
Thera-Turing test: Human behavioral therapists evaluate a medical chatbot's output, requiring correct reasoning and compassion. (Bunge & Desage, 2024)
Think-aloud protocol: A method in which participants verbalize thoughts concurrently or retrospectively while performing a task, used to surface trust and trustworthiness attitudes during interaction. (Ericsson & Simon, 1980)
Three R's of trust: [Definition pending.]
Toulmin model: A layout for practical arguments (data, warrant, backing, qualifier, rebuttal, claim) used in argument-based accounts of trust and probabilistic trust (APT). (Toulmin, 1958)
Training vs. deployment: Distinction between offline parameter estimation or rule creation and online use of a trained system.
Transparency: [Definition pending.]
transparency, transformational: the ability to understand and assess the broader social, cultural, and ethical implications of AI systems (Andrada et al., 2022)
transparency-in-use: When a tool or device becomes sufficiently well understood that using it requires little to no cognitive effort by the user. (Andrada et al., 2022)
Tree of knowledge: A conceptual representation that all knowledge expands from micro to macro levels, and can be separated into four key categories: matter, life, mind and, culture (Henriques, 2008)
Trust: Trust is the strong belief in someone or something. In AI, trust is the belief the user has in the system and its ability to be trustworthy and align with the characteristics of trustworthiness in AI (accuracy, reliability, resiliency, objectivity, security, explainability, safety, accountability, and privacy). (Stanton & Jensen, 2021)
Trust calibration: [Definition pending.]
Trust in Automation (TiA): Körber’s 19-item, six-subscale self-report questionnaire for trust in automation (Reliability/Competence, Understanding/Predictability, Familiarity, Intention of Developers, Propensity to Trust, Trust in Automation); released under CC BY-SA 4.0. (Körber, 2019)
Trust in XAI Context (TXAI): Hoffman et al.’s short state trust scale for a named AI tool after use in an explainable-AI setting (eight Likert items plus an open reason; one reverse-coded item). (Hoffman et al., 2021; Hoffman et al., 2023)
Trustworthiness: Regular, honest, and cooperative behavior, based on shared norms and expecations from all agents within a system; using respected independent oversight to create systems that allow users to believe in the design, operation, and maintenance of the system (Ikenberry & Fukuyama, 1996; Shneiderman, 2020)
Trustworthy AI: For AI to be considered trustworthy, it must be valid and reliable, safe to use, fair and transparent, have accountability, be explainable and interpretable, and protect the privacy of its users. (Tabassi, 2023)
Turing Test: A test designed originally to answer the question, "Can machines think?" The colloquial understanding is whether or not a computer or piece of software can actually sufficiently human so that another human cannot distinguish the artificial from the truly human. (Turing, 1950)
Universal Intelligence Measure: An information-theoretic measure of intelligence in reinforcement-learning environments capturing simplicity, adaptation, and generalization. (Legg & Hutter, 2007)
Unjustified mistrusting: [Definition pending.]
Unjustified trusting: [Definition pending.]
Unsupervised learning: Machine learning that identifies structure in unlabeled data. (El Naqa & Murphy, 2015)
Unwarranted trust: [Definition pending.]
User expectations of AI: Anticipated performance, legality, safety, transparency, and anthropomorphic behavior of AI systems. (Linja et al., 2022; Zhang et al., 2021)
Verisimilitude (fidelity): Whether AI behavior is indistinguishable from a human's (Mueller fidelity dimension). (Mueller & Minnery, 2008)
Vignette (research): A written or video scenario that asks participants to imagine themselves in a described situation so attitudes such as trust can be elicited without a live system.
Warranted trust: [Definition pending.]
Weaponized anthropomorphism: Using anthropomorphic design or interaction to manipulate users (trust, emotion, compliance) rather than support them.
Wizard of Oz (WOZ): Evaluation method simulating AI behavior with a hidden human operator to test concepts before full implementation.
Work domain analysis: Studying how work is performed in context to inform design of tools and automation. (Rasmussen et al., 1990; Vicente, 1999)
Work-Integrated Learning (WIL): An educational approach that combines traditional classroom learning with hands-on, real-world work experience (Zegwaard et al., 2023)
Zero-Shot Prompting: Asking an AI model to perform a task without giving instructions or examples (Guide, 2024)