AI Cyber Glossary

April 2026

The AI Cyber Glossary is a living reference document establishing consistent, governance-ready definitions for artificial intelligence terminology across the health sector. It was developed in direct response to a critical gap in managing healthcare AI and AI cybersecurity: the absence of shared, sector-specific language that clinical, operational, compliance, and technical stakeholders can use with confidence.

A living reference establishing consistent, governance-ready definitions for artificial intelligence terminology across the health sector.

The glossary was developed in direct response to a critical gap in managing healthcare AI and AI cybersecurity: the absence of shared, sector-specific language that clinical, operational, compliance, and technical stakeholders can use with confidence. As AI adoption accelerates across healthcare organizations of every size, inconsistent terminology creates real risk - in procurement decisions, vendor contracts, regulatory submissions, policy development, and patient safety oversight. As a living document the Glossary is designed to serve as the terminological foundation for all current and future HSCC AI Task Group guidance materials.

Term	Acronym	Audience Level	Basic Definition	Advanced Definition	Context
Adverse AI Event		General	A situation where an AI system contributes to patient harm, creates a risk of harm, or produces a near-miss that needs to be looked into. This is different from a cybersecurity event: an adverse AI event doesn't require an attack or a bad actor, it can result from something like a model making a wrong call or a breakdown in how the AI fits into the clinical workflow.	An occurrence in which an AI system contributes to patient harm, creates a foreseeable risk of harm, or produces a near-miss that warrants investigation, whether from model error, misuse, integration failure, or degraded interaction with clinical workflow. Distinguished from AI Cybersecurity by causation rather than outcome: AI Cybersecurity addresses harm arising from unauthorized access, manipulation, theft, or disruption by an adversarial actor, while Adverse AI Event encompasses non-adversarial causes, including performance degradation, workflow integration failure, and human-AI interaction breakdowns. Adverse AI Event is the broader category for incident reporting and root cause analysis purposes; a cybersecurity-caused harm may be logged as both. See also Incident Response Plan.
Agentic AI		General	AI systems that can pursue goals by planning and executing a sequence of actions, including calling external tools, querying systems, or triggering workflows, with varying levels of human oversight between steps. Unlike AI that produces a recommendation, agentic AI takes actions that have costs and real-world effects.		Think of Agentic AI like a really sharp practice manager. You tell them, 'We need to improve patient no-show rates.' They don't wait for you to tell them exactly how, they analyze schedules, tweak reminders, adjust workflows, and report back on results. Regular AI is like a scheduler who only sends a reminder when you say, 'Send this reminder now.' Agentic AI figures out the how and acts.
AI Adversarial Attack		Technical	A deliberate attempt to manipulate an AI system by feeding it carefully crafted inputs designed to cause errors, expose sensitive data, or make it behave in unintended ways.	A deliberate attempt to manipulate an AI system's behavior or outputs by introducing carefully crafted inputs. Serves as a parent category encompassing specific attack types, including Data Poisoning (corrupting training data), Prompt Injection (hijacking model behavior through crafted inputs), Model Inversion (reconstructing training data from outputs), Model Extraction (cloning a model through queries), and Jailbreaking (bypassing safety controls). See also the individual entry for each attack type.
AI Bill of Materials	AIBOM	General	A structured list of components that make up an AI system: which AI model it uses, where the training data came from, what third-party components it depends on, and who hosts it. Like a Software Bill of Materials (SBOM), but specific to AI systems.	A structured inventory documenting the components of an AI system, including foundation models used (name, version, provider); training and fine-tuning data provenance; third-party libraries, frameworks, and dependencies; API dependencies and external service calls; hosting and inference infrastructure; sub-processors; and known limitations or contraindications. Required for High and Critical risk AI systems as a contractual deliverable and used for supply chain risk assessment. See also Software Bill of Materials (SBOM).	Like a nutrition label for an AI system: it tells you exactly what ingredients went into building it, where they came from, and who supplied them.
AI Cybersecurity		General	The practice of protecting AI systems, including their data, models, and operational environments, from unauthorized access, manipulation, theft, and disruption.	The discipline of defending AI systems, their training and operational data, their model artifacts, and their runtime environments against unauthorized access, manipulation, theft, poisoning, adversarial input, and service disruption across the AI lifecycle. Distinct from general cybersecurity in its focus on AI-specific attack surfaces such as training data, model weights, and inference pipelines.
AI Governance		General	The establishment of clear policies, defined roles, and processes to ensure AI is used ethically, safely, and effectively.		The set of rules your practice creates for how to choose, use, and monitor AI tools to protect patients and staff.
AI Incident Reporting		General	The structured process for identifying, documenting, investigating, and communicating AI-related failures, performance problems, bias events, or security issues to the appropriate internal and external parties.	Structured processes by which AI-related failures, performance degradations, bias events, security compromises, and ethical concerns are identified, logged, investigated, escalated, and communicated to internal governance bodies and, where required, to regulators and external oversight entities. See also Incident Response Plan.
AI Inventory		General	A documented register of all AI systems an organization uses, whether purchased, built, or embedded in other products, that serves as the foundation for AI governance.	An enterprise register of all AI systems in use, whether developed, procured, or embedded, capturing ownership, use case, impact classification, vendor, data flows, integration points, version, and monitoring status. The AI Inventory is the foundational artifact for every subsequent AI governance control.	You can't govern what you don't know you have. Before any AI policy or oversight program can function, the organization must know which AI systems are in use.
AI Red Teaming		Technical	Structured adversarial testing of an AI system to find weaknesses, including attempts to manipulate it with crafted inputs, extract sensitive data, or bypass its safety controls. Distinct from traditional IT penetration testing.	A structured adversarial testing practice for AI systems that evaluates robustness against prompt injection, jailbreaking, data extraction, model inversion, adversarial inputs, and harmful output generation. Distinct from traditional network penetration testing in its focus on behavioral, semantic, and model-specific vulnerabilities. Required for High and Critical risk AI systems prior to initial deployment and at least annually in production.
AI Supply Chain		General	All the upstream components, vendors, models, and services that an AI system depends on, any of which could introduce security, performance, or compliance risk.	The full set of upstream dependencies, including foundation models, datasets, libraries, cloud infrastructure, APIs, fine-tuning providers, and tooling vendors, that contribute to the behavior of a deployed AI system, and on which that system's security, performance, and integrity ultimately depend.
AI Threat Analysis and Risk Assessment	AI TARA	Technical	A structured process for identifying, analyzing, and addressing security risks that are specific to AI systems throughout their lifecycle.	A structured method to identify, analyze, and treat AI-specific security risks across the lifecycle of an AI-enabled product, complementing, not replacing, standard device risk management practices such as ISO 14971 and secure software development lifecycle (SDLC) processes.
AI Washing		General	The practice of companies exaggerating or misrepresenting their AI products or services to gain a competitive advantage
Algorithm		General	Step-by-step instructions a computer follows to solve problems. Often used to refer to the mechanism by which an AI performs a task.
Algorithmic AI	AAI	General	AI that analyzes data and makes decisions or predictions based on patterns, as opposed to Generative AI, which creates new content. Examples include risk scoring models, fraud detection, and diagnostic classification tools.	AI that analyzes data using machine learning algorithms to make decisions, predictions, recommendations, or classifications based on that data. Produces structured outputs rather than novel content. Distinct from Generative AI (GAI), which generates new text, images, or other content. Examples include sepsis prediction models, clinical risk scores, and coding assistance tools.
Algorithmovigilance		General	Also known as Algorithm Vigilance. Ongoing monitoring of deployed AI algorithms for safety problems, performance changes, and unintended effects, modeled on pharmacovigilance practices used for drugs.	The scientific methods and activities relating to the evaluation, monitoring, understanding, and prevention of adverse effects or unintended consequences of algorithms, particularly those driven by artificial intelligence. See also Post-Market Surveillance.
Artificial Intelligence	AI	General	Systems or machines that mimic human intelligence (e.g., learning, decision-making) to perform tasks and can progressively improve themselves based on the information they collect.	A machine-based system that can, for a given set of human-defined objectives, make predictions, recommendations, or decisions influencing real or virtual environments. AI systems use machine- and human-based inputs to perceive real and virtual environments; abstract such perceptions into models through analysis in an automated manner; and use model inference to formulate options for information or action.	An incredibly smart and fast assistant that can process information and recognize patterns, but still needs human guidance and oversight.
Artificial Intelligence Performance Monitoring	AI Performance Monitoring	General	Regularly checking how well an AI is working and whether it needs to be adjusted.	The process of regularly collecting and analyzing data on the use of a deployed AI system to evaluate its performance in accomplishing its intended tasks in real-world settings. This monitoring aims to assess performance, detect degradation or changes (e.g., due to data drift), identify misuse, and address safety or usability concerns.
Artificial Intelligence System	AI System	General	A machine that can learn, make decisions, or act in ways that seem smart.	An engineered system that generates outputs such as content, predictions, recommendations, or decisions for a given set of human-defined objectives. AI systems are designed to operate with varying levels of autonomy.
Assistive Artificial Intelligence	Assistive AI	General	AI that helps people do their work better or faster, not replace them.	AI systems designed to assist human users in performing tasks, often by providing recommendations or augmenting human decision-making, without replacing human judgment.
Assurance Case	GSN	Technical	A structured, often visual, argument that links claims about a system's safety or security to the evidence and reasoning that support them, used to demonstrate a device is safe enough for its intended use.	A structured argument, often represented visually using Goal Structuring Notation (GSN), linking claims about device safety or security to supporting evidence and rationale. GSN diagrams show how evidence such as test results, audits, and risk analyses supports the high-level assurance claims required for regulatory approval.
Automation Bias		Technical	The tendency to over-rely on AI recommendations and accept them without question, even when other information suggests the AI may be wrong.	The tendency of human decision-makers to over-rely on AI-generated recommendations, deferring to the system's output even when disconfirming evidence is available. In healthcare, automation bias is a documented contributor to clinical error and a primary motivation for human-in-the-loop design requirements. See also Override Logs.	Like a clinician accepting an AI drug dosage recommendation without checking the patient's renal function: the AI gave an answer, so it must be right.
Autonomous Artificial Intelligence	AutonomousAI	General	AI that can make decisions and take actions without needing help from a person.	AI systems capable of performing tasks without human intervention, making decisions or taking actions independently based on their programming and data inputs. Distinguished from Agentic AI by its focus on decision authority and independence from human intervention rather than multi-step action execution or external tool use. An AI system may be autonomous without being agentic, and agentic without being fully autonomous.
Autonomy Levels		General	A tiered framework describing how much control an AI system has over decisions and actions. Ranges from Level 0 (no automation) to Level 5 (full autonomy without human input). Important for understanding risk and governance implications in clinical and operational AI deployments. See also Autonomous AI
Benchmark		General	A standardized way of testing and comparing how well different AI models perform on the same tasks, used to understand a model's strengths, weaknesses, and fit for a particular use.	The process of measuring and comparing the performance of AI models across standardized tasks to determine their strengths, weaknesses, and suitability for specific use cases. Widely used in research, product development, and deployment decisions. Examples include Stanford's HELM framework.
Bias	in AI	General	When AI results are unfair due to errors in data or design.		An AI diagnostic tool trained mostly on data from one demographic group may be less accurate for patients from other groups.
Blast Radius		Technical	In AI incident response, the scope of how far the impact of an AI failure or compromise reaches, including how many patients were affected, which clinical decisions were influenced, and over what period.	The scope of potential impact when an AI system fails, is compromised, or acts incorrectly, including the number of patients potentially affected, the clinical decisions made based on compromised AI outputs, the time window of exposure, and downstream system effects. Particularly significant for agentic AI systems that can take consequential actions before detection. Used as a structured triage concept in AI incident response. See also Incident Response Plan.
Business Associate Agreement	BAA	General	A mandatory contract under HIPAA that a vendor must sign if they handle PHI on behalf of a healthcare provider.		A non-negotiable contract you must have with any AI vendor that will touch your patient data.
Chain-of-Thought Modeling		Technical	Technical definition - click + to view	A technique in AI models, particularly large language models, that enables the model to generate intermediate reasoning steps or explanations as part of its output to improve transparency, logic validation, and safety oversight.
Clinical Oversight		General	(AI Context): The governance of clinical activities or processes that involve or are influenced by artificial intelligence. It ensures that AI supports clinical effectiveness, patient safety, and compliance with medical standards.
Clinical Skill Attrition		General	The gradual loss of clinical skills that can occur when clinicians rely heavily on AI tools for tasks they previously performed independently.	Observed degradation in clinicians' unassisted diagnostic, procedural, or reasoning capabilities following prolonged reliance on AI-supported workflows. Documented in imaging, endoscopy, and decision-support contexts, and a governance concern for long-horizon AI deployment.
Computer Vision	CV	General	A field of AI that trains computers to interpret and understand the visual world from images and videos.		Self-driving cars use computer vision to identify pedestrians, traffic lights, and other vehicles on the road.
Concept Drift		Technical	A change in what the correct AI output should be for a given input, even when the input data looks the same. Different from data drift, which is a change in the data itself.	A change in the underlying statistical relationship between input features and model outputs over time. Distinct from data drift (a change in input data distribution): concept drift means the correct answer for a given input has changed, which can degrade model performance even when input data distributions appear stable. Requires ongoing monitoring and can trigger model revalidation.	Like a clinical guideline changing: the same patient presentation that previously called for one treatment now calls for another. An AI trained before the guideline change may still give the old answer even though the input data hasn't changed.
Consent		General	Permission a person gives for AI to use their information — and in many cases, this permission is required by law before AI can be used.	A process by which individuals authorize the use of their personal or health information, with clear communication of purpose, risks, rights, and the role of AI. Increasingly, state laws are requiring explicit consent before AI can be used in certain elements of care or administrative functions (e.g., clinical decision support, transcription, or patient interaction tools). Consent ensures patients or users understand when AI is being used and agree to that use, aligning with privacy, ethical, and regulatory obligations.	Like signing a HIPAA form before your doctor can share your medical records—except now, you may also need to sign or acknowledge when AI is used to assist in care or manage records.
Continual Machine Learning		General	A type of AI that keeps learning and improving from new data over time.	A machine learning approach where models are designed to learn continuously from new data, allowing them to adapt to changes over time without forgetting previously learned information.
Continuous Monitoring		General	Ongoing observation of a deployed AI system to detect performance changes, safety signals, security issues, or bias before they cause harm. See also Artificial Intelligence Performance Monitoring.	Ongoing, instrumented observation of a deployed AI system's inputs, outputs, performance, drift, safety signals, bias indicators, and security posture, conducted at a cadence sufficient to detect meaningful change before it produces patient or operational harm.
Convolutional Neural Network	CNN	Technical	Technical definition - click + to view	A type of deep learning model particularly effective for analyzing visual data. CNNs use convolutional layers to automatically and adaptively learn spatial hierarchies of features from input images.
Corrective and Preventive Action	CAPA	General	A general quality management process, not specific to AI, for identifying what went wrong, addressing the root cause, and putting steps in place to prevent it from happening again. In AI governance, CAPA is commonly used to document and justify significant changes made to an AI system, with a clear trail back to the issue that prompted them, such as an incident, model failure, bias finding, or audit result. See also Incident Response Plan.	A structured quality management process, not unique to AI, with two components: Corrective Action, which addresses the root cause of an identified failure or nonconformance, and Preventive Action, which implements changes to reduce the likelihood of recurrence. In AI governance, CAPA is commonly applied following AI incidents, bias events, post-incident reviews, or audit findings, though it is not automatically or universally required for every such event, its use depends on the severity and nature of the finding. Common in FDA quality management programs and applied broadly across AI lifecycle governance to maintain traceability between an identified issue and the changes made to address it.
Cross-Border Data Control		General	Making sure data that moves between countries is handled safely and legally.	Policies and technical controls that govern the transmission, processing, and storage of data across international boundaries, ensuring compliance with regional laws, privacy regulations, and security standards.
Crowdworkers		General	Temporary online workers who label or process data for AI systems at scale, typically without specialized subject-matter expertise.	Temporary online workers engaged to label or process data at scale for AI training or evaluation purposes, often without domain expertise. Data quality and labeling accuracy from crowdworker sources should be accounted for in AI model validation, particularly for clinical applications.
Data Card		General	A structured summary providing essential information about a dataset, including its composition, intended use, limitations, and ethical considerations, to promote transparency and responsible use.
Data Drift		General	The change in data distribution over time, which can affect the performance of machine learning models if they are not updated to accommodate the new data patterns.
Data Labeling		Technical	The process of tagging data with information so an AI can learn from it.		Like marking X-rays to show where a tumor is.
Data Lineage		Technical		The documented chain of custody for a dataset, covering source, acquisition method, consent basis, transformations applied, quality controls, and downstream uses. Data lineage supports auditability, reproducibility, and compliance review across the AI lifecycle.
Data Loss Prevention	DLP	Technical	Security tools that detect and block sensitive data, like patient records, from being sent to unauthorized destinations. In AI, used to prevent PHI from being submitted to unapproved AI services or appearing in AI-generated outputs.	A set of security tools and controls that detect and prevent unauthorized transmission of sensitive data outside approved boundaries. In the AI governance context, DLP controls inspect prompts and inputs sent to external AI services, block or redact recognizable PHI patterns, and ensure transmission only to approved destinations. A required control for acceptable use enforcement across HSCC AI governance guidance.
Data Minimization		General	Only using the data that is needed for the AI task.		Like asking only the necessary patient questions for a diagnosis.
Data Minimization Map		General	A document that shows exactly what personal or clinical data a system collects, why each piece is needed, and how long it will be kept, so an organization only holds the data it actually needs.	Structured documentation showing which personal or clinical data elements are collected, the justification for their necessity, and applicable retention policies. Supports demonstrating compliance with data minimization principles under HIPAA and similar frameworks. See also Data Minimization.
Data Mining		General	The process of discovering patterns, trends, and insights from large datasets.		Retail companies use data mining to analyze customer purchasing habits and recommend products.
Data Poisoning		Technical	A cyberattack where an attacker intentionally feeds "bad" data into an AI's training set to make it unreliable or biased.		A sophisticated attack that could corrupt a diagnostic AI model, causing it to make mistakes.
Data Privacy		General	Protecting sensitive information, particularly patient information, used in AI systems.
Data Protection Impact Assessment	DPIA	General	A required risk review under European data protection law (GDPR) for AI or data processing activities that pose a high risk to individuals' privacy.	A GDPR-mandated risk analysis process for high-risk data processing activities, including many forms of medical AI. Identifies privacy risks, evaluates mitigations, and documents compliance with GDPR. A distinct legal requirement from a Privacy Risk Assessment (PIA) conducted under U.S. frameworks, though closely related in practice. See also Privacy Risk Assessment.
Deep Learning		General	A type of machine learning where AI uses layered networks (like digital neurons) to learn patterns in large amounts of data. It’s how AI can recognize faces, translate languages, or generate text.	A subset of machine learning involving neural networks with multiple layers (deep neural networks) that can learn complex patterns in large amounts of data.	Think of it as giving AI a brain that learns by seeing lots of examples.
Deep Research		General	An in-depth, focused approach to learning or problem-solving that goes beyond surface-level searching. In AI projects, this might mean carefully exploring complex topics, risks, or user needs before building or deciding.		It’s not just Googling — it’s diving into the full context, evidence, and implications.
Defense-in-Depth	DiD	General	A security strategy that layers multiple protections around systems and data, so that if one safeguard fails, others are still in place to prevent harm.	A cybersecurity strategy that uses multiple, layered security controls to protect systems, data, and infrastructure. The core idea is that no single safeguard is perfect, so overlapping defenses create redundancy: if one layer fails, others still provide protection.
De-identification		General	The process of removing personal identifiers from data to protect individual privacy.		A method used to prepare patient data for training AI models, though it doesn't always eliminate the risk of re-identification.
De-Identification Approach		General	The specific method an organization uses to remove or change identifying patient information so it can no longer be tied back to a person, done in a way that meets legal requirements.	A defined methodology for removing or transforming identifiable patient data to comply with HIPAA, GDPR, and ISO/IEC 20889. Includes pseudonymization, anonymization, and re-identification risk assessment. See also De-identification and Re-Identification Risk.
Differential Privacy	DP	Technical	A privacy technique that adds carefully calculated noise to data or model outputs so that no one can tell whether any specific individual's information was used, providing a mathematical guarantee of privacy.	A mathematical privacy framework that provides formal privacy guarantees by introducing carefully calibrated statistical noise into data or model outputs, ensuring that the presence or absence of any individual's data in a dataset cannot be inferred from query results or model outputs. Used in federated learning and fine-tuning contexts to train models on clinical data while protecting individual patient privacy. Prevents membership inference and model inversion attacks, where an attacker tries to recover individual patient data from a model.
Digital Health Technology	DHT	General	Technologies such as mobile health apps, wearable devices, telemedicine, and health information technology systems that use computing platforms, connectivity, software, and sensors for health care and related uses.
Digital Twin		Technical	Technical definition - click + to view	A virtual representation of a physical object or system that can be used to simulate, predict, and optimize performance in real-time, often used in healthcare to model patient-specific conditions.
Drift Metrics		Technical	The specific measurements used to track whether an AI system's performance is changing over time, covering shifts in the data it sees, the patterns it's supposed to recognize, or its overall accuracy.	Monitoring indicators used to measure model performance drift, including data distribution shift, concept drift, and label drift, in production. Essential for adaptive AI in medical devices, where performance must remain aligned with validated clinical expectations. See also Data Drift, Concept Drift, and Model Drift.
Dynamic Application Security Testing	DAST	Technical	Security testing performed on a running application to find vulnerabilities under real-world conditions, rather than by reviewing the code itself.	Testing of a running application to detect vulnerabilities in real execution conditions. Simulates external attacks, such as SQL injection or cross-site scripting (XSS), to uncover issues not visible in code review.
Embedded AI		General	AI functionality built into a larger system, like an EHR module or medical device, where the AI is not separately visible or purchased and may be activated without a distinct procurement or governance review.	AI functionality integrated within a larger system, such as an EHR module, medical device, or SaaS platform, where the AI component may not be independently visible, separately purchasable, or subject to a discrete procurement event. Requires specific discovery processes to identify and govern. Can be introduced through firmware updates, feature activations, or platform releases without a separate procurement event, creating governance gaps if not specifically addressed.	Like discovering your EHR now includes an AI-generated summary feature that was silently enabled in an update: it's AI, it touches patient data, but no one reviewed or approved it because it wasn't a separate purchase.
End Of Life	EOL	General	Life cycle stage of a product, starting when (1) the manufacturer no longer sells the product beyond its useful life (as defined by the manufacturer), and (2) the product has gone through a formal EOL process, including notification to users.
Ensemble Methods		Technical	Technical definition - click + to view	Machine learning techniques that combine predictions from multiple models to improve overall performance, robustness, and accuracy compared to individual models.
Ethical AI		General	The practice of designing and using AI in a way that aligns with moral values and societal good.		An ethical AI principle would be to ensure that an AI system used in healthcare does not perpetuate existing biases against certain patient populations.
Explainability	XAI	General	Being able to understand and explain how an AI system makes its decisions in a way that humans can follow.	The extent to which the internal mechanics of a machine learning system can be explained in human terms, enabling users to understand, trust, and effectively manage AI systems.	It answers the "why" question. Why did the AI flag this specific area on the X-ray as potentially problematic?
Explainable AI	XAI	General	A set of processes and methods that allows human users to understand and trust the results and output created by machine learning algorithms.		An XAI system in finance could provide a clear explanation for why a loan application was denied, rather than just giving a "yes" or "no" answer.
Fail-Safe Behavior		Technical		The defined, predictable response of an AI system, or of the workflow surrounding it, when the system encounters an input outside its validated operating range, an internal error, or an external dependency failure. Fail-safe design is a core patient safety expectation for High and Critical impact AI systems.
Failure Mode and Effects Analysis	FMEA	Technical	A structured method for identifying what could go wrong with a system, how bad the impact would be, and what controls are needed to reduce the risk. Used in healthcare AI to assess clinical safety risks before deployment.	A structured, systematic analysis used to identify potential failure modes in a system, assess their effects on system performance and patient safety, and prioritize risk mitigation strategies. In the healthcare AI context, applied as part of clinical safety risk assessment for AI systems in alignment with ISO 14971. Required in the HSCC Implementation Guide alongside hazard analysis and human factors review.
Feature Engineering		General	The process of selecting, modifying, or creating new input features from raw data to improve the performance of machine learning models.
Federated Learning		General	A machine learning approach where models are trained across multiple decentralized devices or servers holding local data samples, without exchanging them, enhancing data privacy.
Fine-tuning		General	Adjusting a pre-trained model with specific data to make it more accurate for a given task.
Foundation Models		Technical	Technical definition - click + to view	Large-scale machine learning models trained on vast amounts of data that can be adapted to a wide range of tasks, often serving as a base for more specialized models.
Fourth-Party Risk		Technical	The risk a healthcare organization inherits through its vendors' own vendors and dependencies: parties the organization has no direct relationship with, but whose failures can still affect AI systems in use.	The risk a healthcare organization inherits through its vendors' own dependencies, including subcontractors, upstream model providers, open-source components, and cloud services, which are typically neither disclosed nor directly controlled by the healthcare customer but can drive AI system compromise, outage, or regulatory exposure.	If your AI vendor relies on a foundation model from a third party, and that foundation model provider has a breach or service outage, your organization is affected even though you have no direct relationship with that provider.
Generative Adversarial Network	GAN	General	A class of machine learning frameworks where two neural networks (a generator and a discriminator) contest with each other to produce data indistinguishable from real data.
Generative AI		General	AI that creates new content such as text, images, and video based on input and learned patterns. See also Algorithmic AI (AAI).		Using a tool like DALL-E to generate an image of "an astronaut riding a horse on the moon" by providing a text prompt.
Generative Artificial Intelligence	Generative AI	General	AI that creates new content such as text, images, and video based on input and learned patterns. See also Algorithmic AI (AAI).	AI systems capable of generating new content, such as text, images, or audio, that resembles human-created data, often using models like GANs or transformer-based architectures.	It’s like a chef who has learned thousands of recipes. When you ask for a dish, they don’t just repeat one recipe—they combine what they know to create something new that fits what you asked for.
Governance, Risk, and Compliance	GRC	General	An integrated approach to managing an organization's oversight, risk assessment, and compliance obligations together as a coordinated program rather than in silos.	A structured framework that aligns an organization's governance structure, enterprise risk management, and regulatory compliance activities. In the AI context, GRC specifically refers to the processes and assessment frameworks used to evaluate AI vendor practices, classify AI system risks, and verify adherence to applicable regulations and organizational policies. The AI-specific GRC assessment is a primary tool in the vendor due diligence process.
Guardrail Specs		Technical	The exact technical rules built into an AI system that define what it is and isn't allowed to do, such as blocking certain outputs or refusing unsafe requests.	Formal specifications of the safety and security boundaries embedded into an AI model or device software. Examples include restricting recommended dosing ranges, blocking unsafe parameter inputs, or filtering out-of-scope AI outputs. See also Guardrails.
Guardrails		General	The policies, procedures, and technical controls put in place to ensure that AI systems operate within ethical and legal boundaries and do not cause harm.		A guardrail for a generative AI model could be a filter to prevent it from generating hateful or violent content.
Hallucination		General	When AI generates false or misleading information.		Hallucination in AI is like a patient giving you made-up medical history with total confidence — sounds convincing, but none of it checks out.
HHS 405d HICP	HICP	General	Health Industry Cybersecurity Practices. A set of voluntary, consensus-based cybersecurity guidelines from HHS to help healthcare organizations.		The government's recommended "playbook" for cybersecurity that is highly relevant for protecting AI tools.
HIC-SCRiM	HIC-SCRiM	General	The Health Industry Cybersecurity Supply Chain Risk Management Guide (HICSCRiM) – is primarily written for leadership in small to medium sized organizations. It is intended to provide actionable guidance and practical tools to enable those organizations to manage the cybersecurity risks they face through their dependencies within the health system supply chain.
HIPAA		General	Health Insurance Portability and Accountability Act. The U.S. federal law setting national standards for protecting patient health information.		The foundational privacy and security law that governs how all patient data, including data used by AI, must be handled.
Human Factors Engineering Files	HFE Files	Technical	Documentation showing that a device's design and workflows have been tested to minimize the chance that a user makes a harmful mistake while operating it.	Documentation of usability engineering activities performed in accordance with IEC 62366-1, demonstrating that device user interface and workflows minimize use errors that could cause harm. For AI-enabled devices, this includes explainability, override workflows, and user alerts.
Human-in-the-Loop	HITL	General	An approach to AI oversight where a human must actively review and approve an AI system's output before it is acted upon. See also Autonomy Levels.	A governance model in which a human reviews, approves, modifies, or intervenes in an AI system's outputs or actions before they are finalized or carried out, corresponding to non-autonomous AI operation. May be applied during model development, validation, or operational use.	Like a clinician reviewing and approving an AI-generated treatment recommendation before it is entered into the patient's record, the AI assists, but the human remains responsible for the final decision.
Human-on-the-Loop	HOTL	General	An approach to AI oversight where the AI system acts on its own, but a human continuously monitors its outputs and can step in or override it if needed. See also Autonomy Levels.	A governance model in which an AI system executes actions automatically while a human continuously supervises system outputs and retains the ability to intervene or override, corresponding to supervised or conditionally autonomous AI operation.	Like a pharmacist monitoring an automated medication dispensing system: the system runs on its own, but the pharmacist is watching in real time and can stop it if something looks wrong.
Human-Out-of-the-Loop	HOOTL	General	An approach to AI oversight where an AI system makes and acts on decisions with no human able to intervene. See also Autonomy Levels.	A condition in which an AI system operates and makes decisions without any mechanism for human intervention, corresponding to fully autonomous AI operation.	An AI system managing supply chain reordering with no human review step: orders are placed automatically with no opportunity for a person to catch or stop an error before it happens.
Impact Classification		General	Assigning an AI system to a risk tier, such as Low, Moderate, High, or Critical, based on what could happen if the system fails, which determines which governance controls apply.	Assignment of an AI use case to a risk tier (commonly Low, Moderate, High, or Critical) based on the consequences of model failure, the reversibility of affected decisions, the populations affected, and the degree to which clinical or operational workflow depends on model output. Impact classification drives the level of validation, monitoring, vendor contracting, and oversight required.
Incident Response Plan	IR Plan	General	The documented, end-to-end procedure an organization follows to detect, report, contain, investigate, and recover from a cybersecurity or safety incident involving an AI system. See also AI Incident Reporting, Root Cause Analysis (RCA), Corrective and Preventive Action (CAPA), and Blast Radius.	Documented procedure for detecting, reporting, mitigating, and recovering from cybersecurity or safety incidents in medical devices. Includes escalation steps, communication plans, containment, forensic analysis, and recovery actions. Serves as the umbrella procedure connecting AI Incident Reporting (identification and communication), Root Cause Analysis (investigation), Corrective and Preventive Action (remediation), and Blast Radius (impact scoping).	The master playbook for what happens when something goes wrong. AI Incident Reporting, RCA, CAPA, and Blast Radius are each one chapter of that playbook.
Infrastructure-as-Code Scanning	IaC Scans	Technical	Automated checks of the configuration files used to set up cloud and server infrastructure, done to catch security misconfigurations before systems go live.	Automated review of configuration scripts, such as Terraform, CloudFormation, or Kubernetes YAML files, for misconfigurations or policy violations. Detects issues like open storage buckets, exposed ports, or missing encryption controls before deployment.
Internet of Things Device	IoT	General	Physical devices embedded with sensors, software, and connectivity to collect and exchange data over the internet, often used in healthcare for monitoring and diagnostics.
Interoperability		General	The ability of different information systems, devices, or applications to connect, exchange, and use data cohesively and effectively.
Interpretability		General	The degree to which a human can understand the cause of a decision made by a machine learning model, often essential for trust and regulatory compliance.
Jailbreaking		Technical	Crafting inputs designed to trick an AI model into ignoring its safety rules and producing responses it was designed to refuse, such as dangerous medical advice or harmful content.	The practice of crafting prompts or inputs designed to circumvent an AI model's built-in safety guardrails, causing it to produce outputs it was programmed to refuse. Distinct from prompt injection in that jailbreaking targets the model's own safety controls rather than manipulating the context of a query. In healthcare, jailbreaking a clinical AI assistant could cause it to provide dangerous medical advice it was designed to refuse.
Kill Switch		General	A tool that can quickly shut down an AI system to stop harm or prevent risk.	A fail-safe mechanism designed to immediately disable or isolate an AI system in the event of malfunction, attack, or unintended behavior, helping to prevent further harm or data compromise.
Large Language Model	LLM	General	AI trained on vast amounts of text to understand and generate language.	A type of AI model trained on extensive text data capable of understanding and generating human-like language, used in applications like chatbots and language translation.	OpenAI's GPT-4 and Google's Gemini are examples of LLMs that power various AI applications.
Lifecycle Management		General	Governing an AI system through every stage of its existence, from initial selection and deployment through ongoing monitoring to eventual retirement, with defined responsibilities and controls at each stage.	Disciplined governance of an AI system across every stage of its existence (ideation, data sourcing, development, validation, procurement, integration, deployment, ongoing monitoring, revalidation, and secure decommissioning) with defined owners, controls, and documentation at each stage.
Locked Model		Technical	Technical definition - click + to view	A machine learning model that, once trained and validated, is fixed and does not change its parameters or behavior over time, ensuring consistent performance.
Machine Learning	ML	General	AI that learns patterns from data instead of being directly programmed.	A subset of AI involving algorithms that improve automatically through experience by learning patterns from data without being explicitly programmed.	Like a medical student who learns by studying many patient cases (data) to recognize symptoms and predict outcomes for new patients.
Machine Learning Algorithm	ML Algorithm	General	A specific procedure or formula for solving a problem, used in ML to learn from data and make predictions or decisions.
Machine Learning Algorithmic Bias	ML Algorithmic Bias	General	Systematic errors in ML models that result in unfair outcomes, often due to biased training data or flawed model assumptions.
Machine Learning Model	ML Model	Technical	Technical definition - click + to view	The output of a machine learning algorithm trained on data, used to make predictions or decisions without being explicitly programmed for the task.
Membership Inference		Technical	An attack where a malicious actor tries to determine whether a specific individual's data was part of a model's training set. This is a privacy threat.		An attacker queries a hospital's diagnostic AI with a specific patient's data to infer if that person's records were used in training, thus revealing they were a patient.
MITRE ATLAS		Technical	A knowledge base of attack techniques specifically targeting AI systems: the AI equivalent of the MITRE ATT&CK framework used for traditional cybersecurity threats.	Adversarial Threat Landscape for Artificial-Intelligence Systems: a knowledge base of adversarial machine learning attack techniques against AI systems, analogous to MITRE ATT&CK for traditional cybersecurity. Provides a structured taxonomy of AI-specific threat techniques used for AI threat modeling and red team exercises. Referenced in the HSCC Implementation Guide's AI Threat Model Template (Appendix M).
Mobile Application Security Testing	MAST	Technical	Security testing focused specifically on mobile apps, checking for risks unique to how mobile devices store data and connect to other systems.	Security testing specific to mobile applications, combining static application security testing (SAST) and dynamic application security testing (DAST) with platform-specific checks. Includes mobile-specific risks such as insecure data storage, weak encryption, or unsafe API integrations.
Model		General	A trained mathematical representation used by an AI system to make decisions or predictions.