IAPP Artificial Intelligence Training 2025 - Module 1-7

Flashcards covering foundational concepts, technical mechanisms, governance frameworks, and legal regulations (including NIST and the EU AI Act) from the AI Training 2025 transcript.

Artificial Intelligence (AI)

A machine performing tasks that normally require human intelligence.

Alan Turing Test

A test to determine whether or not a machine is intelligent by checking if it can fool an interviewer into thinking he’s human.

Socio-technical system

A system where AI influences society and vice versa, requiring consideration of all relevant stakeholders.

Computing

The use of computers to process, store, and manage information, involving calculations, running software, and handling data.

Cloud Computing

The delivery of computing services (storage, processing power, software) over the internet using remote servers hosted by providers.

Internet of Things (IoT)

A network of physical devices, vehicles, and appliances connected to the internet that can collect, exchange, and share data.

Privacy-enhancing technologies (PETs)

Tools and techniques designed to protect personal data and ensure privacy while still enabling data processing and analysis.

Encryption

A PET that protects data by converting it into a secure format that can only be accessed by authorized users.

Anonymization

A PET that removes or alters personal identifiers from data so individuals cannot be identified.

Artificial Narrow Intelligence

AI systems designed to perform specific narrowly defined tasks within a specialized domain but lacking the ability to transfer knowledge to other areas.

Artificial General Intelligence (AGI)

Strong AI with human-level intelligence capable of reasoning, learning, and solving problems across a wide range of domains; it does not exist yet.

Artificial Super Intelligence (ASI)

Hypothetical AI systems that surpass human intelligence in every domain.

Big Data

Extremely large sets of data too complex, fast-moving, or massive for traditional tools, characterized by Volume, Velocity, Variety, Veracity, and Value.

Development (AI)

The process of designing, coding, and building software or applications, including testing and debugging.

Deployment (AI)

The process of releasing and installing finished software for use by end users, making it live or accessible on platforms.

Ethics

A set of moral principles or rules that guide human behavior and help determine what is right or wrong in different situations.

Guardrails

Protective measures or guidelines designed to prevent mistakes, risks, or unwanted outcomes.

Outputs

The direct, measurable results or products produced by a process (the 'what').

Outcomes

The longer-term effects, impacts, or changes brought about by outputs (the 'what happens because' of the products).

Deterministic outputs

Outputs that are predictable and fixed based on the input.

Probabilistic outputs

Outputs that are uncertain or expressed in terms of probabilities.

Supervised Learning

A machine learning model involving training on labeled data where input is mapped to a known target.

Unsupervised Learning

A machine learning model that uses unlabeled data to find patterns or relationships without predefined targets.

Reinforcement Learning

A model where an agent learns to make decisions by interacting with an environment and receiving rewards or penalties.

Classification Models

A subcategory of supervised learning that produces outputs in specific categories or discrete labels.

Regression Models

A subcategory of supervised learning used to predict continuous numerical values.

Clustering

An unsupervised learning category that automatically groups data points sharing similar attributes.

Association Rule Learning

An unsupervised learning category that identifies relationships and associations between data points.

Deep Learning

A subset of machine learning using multi-layered neural networks to simulate the human brain for recognizing complex patterns.

Generative AI

Deep learning models that use training data to create new output, such as text, images, or video, that resembles the original data.

Large Language Models (LLMs)

AI using deep learning algorithms (often semi-supervised) trained on massive text datasets containing billions to trillions of parameters.

Natural Language Processing (NLP)

Technology enabling computers to recognize, understand, and generate text and speech by combining linguistics and machine learning.

Transformer Models

A neural network architecture that captures long-range dependencies in sequential data, preserving context for coherent text generation.

Expert Systems

AI intended to mimic human decision-making abilities in a specific field using a knowledge base, inference engine, and user interface.

CPU (Central Processing Unit)

The 'brain' of the computer that handles general tasks one by one with high precision.

GPU (Graphics Processing Unit)

A specialized processor designed to perform many calculations simultaneously, critical for AI training and graphics.

Hyperparameters

Parameters adjusted for an AI model to tune it toward desired outcomes, usually via trial and error.

Learning Rate

A hyperparameter controlling how fast a model learns or how much information it uses for decisions.

Epoch

One full cycle through the entire training dataset during the learning process.

Momentum (AI)

A hyperparameter representing the amount of history included in the model's equation.

Data Drift

A decline in model performance occurring when the relationship between input data and output predictions changes over time.

Harms Taxonomy

A list or ontological map of negative consequences that could befall a data subject or organization if information is leaked or misused.

Ontological map

A structured representation of fundamental concepts and their relations within a specific domain.

Implicit Bias

Unconscious discrimination or prejudice toward a particular group or individual.

Hallucinations

Instances where a generative AI model creates factually incorrect or contradictory output that appears to be factual.

Filter Bubbles

The confinement in an information bubble personalized by algorithms that limit a user's exposure to diverse opinions.

Liability

The legal or financial obligation for consequences, such as damages or penalties.

Accountability

The ethical and organizational responsibility to be answerable and transparent about conduct and decisions.

Trustworthy AI

AI that is human-centric, accountable, transparent, explainable, and privacy-enhanced.

NIST AI Risk Management Framework

A guide for managing AI risks through four core functions: govern, map, measure, and manage.

ARIA (Assessing Risks and Impacts of AI)

A NIST program designed to assess LLMs based on predefined scenarios, testing methods, and metrics.

Benchmarking

The process of comparing a system's performance against a reference standard or known alternative processes.

GIGO (Garbage In, Garbage Out)

The concept that the quality of a system's output is determined by the quality of the data fed into it.

Structured Data

Data organized in fixed fields, such as spreadsheet rows and columns.

Unstructured Data

Data that lacks a specific structure and does not fit into typical database fields, such as images or social media posts.

Data Wrangling

The process of converting raw data into valuable, usable information through reformatting and cleaning.

Differential Privacy (DP)

A mathematical technique that adds random noise to datasets to ensure individual information cannot be identified.

Federated Data

Decentralized data stored in separate locations that can be used collaboratively for model training without being moved.

Feature Engineering

The transformation of raw data into useful representations or features to improve model performance.

Three Lines of Defense (3LOD)

A governance model for risk management involving process owners (1st line), risk teams (2nd line), and internal audit (3rd line).

Brittleness

An AI characteristic where a system performs successfully in one instance but fails in another.

Model Cards

Standardized documents providing information on a model's function, output, version, and the dataset used.

Containerization

A packaging method that places an AI model and all its dependencies into a self-contained unit for deployment.

REST APIs

A set of rules for software communication using standard web methods like GET, POST, PUT, and DELETE.

Retrieval-augmented generation (RAG)

A process that optimizes LLM output by referencing a knowledge base separate from its original training sources.

Frontier Models

The most advanced, large-scale general-purpose AI models at the cutting edge of development, such as GPT-4o.

EU AI Act

A landmark risk-based regulation aimed at ensuring AI development and deployment in the EU is safe, transparent, and respectful of rights.

Prohibited Risk (EU AI Act)

AI systems with unacceptable risks, such as social credit scoring or untargeted facial scraping, which are banned from the market.

High Risk (EU AI Act)

AI systems used in critical areas (e.g., healthcare, education, law enforcement) that must meet strict requirements for risk management and data governance.

Conformity Assessment (CA)

The process of demonstrating compliance with regulatory requirements for high-risk AI systems before they are placed on the market.

Fundamental Rights Impact Assessment (FRIA)

An assessment deployers must conduct to evaluate a high-risk AI system's impact on privacy, non-discrimination, and freedom of expression.

General-purpose AI (GPAI)

An AI model, such as an LLM, that displays significant generality and can perform a wide range of distinct tasks.

Systemic Risk (GPAI)

Risk associated with very large models trained with substantial resources, typically exceeding $10^{25}$ floating point operations (FLOPs).

ISO 22989:2022

An international standard establishing terminology and describing core concepts in the field of artificial intelligence.

ISO 42001:2023

An international standard providing guidance for a management system to use AI responsibly and effectively.

HUDERIA

The Council of Europe's methodology for assessing AI impact on Human Rights, Democracy, and the Rule of Law.

GDPR Article 22

The provision generally prohibiting decisions based solely on automated processing if they produce legal or similarly significant effects on individuals.

DPIA (Data Protection Impact Assessment)

A GDPR-required process to identify and minimize risks arising from the processing of personal data, particularly using new technologies.

Homomorphic Encryption (HE)

An advanced encryption type allowing computations to be performed on encrypted data without ever decrypting it.

Strict Liability

A liability regime where a claimant only needs to prove a product was defective and caused harm, regardless of manufacturer negligence.