AI Glossary (Artificial Intelligence FAQ)
Complete 105-question executive glossary (FAQ) covering Artificial Intelligence—from foundational Prompts and LLMs to autonomous Agentic Workflows, RAG, and Neural Networks.
Glossary Questions and Answers
A mathematical function inside artificial neurons that determines whether a neuron should be activated and pass a signal to the next layer, introducing non-linearity into neural networks.
A specialized machine learning methodology in which the algorithm interactively queries a human expert to label the most ambiguous or informative data points, reducing annotation costs.
The deliberate introduction of subtle, often imperceptible perturbations into an AI model’s input designed to deceive the system into producing an incorrect output or classification.
An autonomous artificial intelligence system that does not merely reply to prompts, but perceives its environment, formulates multi-step plans, invokes external tools (APIs, databases, browsers), and executes actions to achieve a defined goal.
The technical and philosophical discipline focused on steering artificial intelligence systems so that their objectives, behaviors, and decisions remain strictly aligned with human values and intent.
The multidisciplinary framework of moral principles and governance guidelines ensuring that artificial intelligence systems are designed and deployed fairly, transparently, and without causing societal harm.
The overarching framework of internal corporate policies, security guardrails, audit trails, and regulatory controls that govern how an organization builds, procures, and uses artificial intelligence.
A finite, well-defined sequence of logical or mathematical instructions executed step by step by a computer to solve a problem, analyze data, or make a decision.
Systematic and repeatable errors in an AI system that create unfair outcomes or discrimination because of skewed, incomplete, or historically prejudiced training data.
The process of tagging raw data (images, text, audio, or video) with descriptive metadata so that supervised machine learning models can learn to recognize patterns accurately.
The human tendency to attribute genuine emotions, consciousness, or human intent to artificial intelligence systems that are actually operating through statistical token prediction.
A theoretical stage of artificial intelligence capable of understanding, learning, and performing any intellectual task that a human being can do across any domain.
The branch of computer science dedicated to building software systems and machines capable of performing tasks that typically require human intelligence—such as reasoning, learning, perception, and language understanding.
Also known as Weak AI, it refers to every AI system in existence today—engineered to excel at specific, bounded tasks (such as language generation, image recognition, or chess) without general human consciousness.
A hypothetical future form of AI that surpasses the brightest human minds across virtually every field, including scientific creativity, strategic wisdom, and social skills.
The foundational neural network technique that allows a model to dynamically weigh the relative importance of different words or tokens in an input sequence, regardless of their distance from one another.
An unsupervised neural network architecture trained to compress input data into a compact latent representation and then reconstruct the original input as accurately as possible.
Hardware or software systems—such as self-driving vehicles, drones, or enterprise software agents—capable of operating and making real-time decisions in dynamic environments without human intervention.
The core mathematical algorithm used to train neural networks by calculating the gradient of the loss function and propagating error signals backward through the network to update its weights.
A hyperparameter defining the number of training examples processed together in a single forward and backward pass before the model’s internal weights are updated.
A probabilistic graphical model that represents a set of variables and their conditional dependencies via a directed acyclic graph (DAG).
Standardized test suites and datasets (such as MMLU, SWE-bench, GPQA, and Vals Index) used to objectively measure and compare the reasoning, coding, and accuracy of different AI models.
Bidirectional Encoder Representations from Transformers—a landmark Google language model that reads text bidirectionally to understand deep contextual relationships between words.
Extremely large and complex datasets—characterized by high Volume, Velocity, and Variety—that cannot be processed by traditional databases and serve as the fuel for modern AI training.
A complex deep learning system whose inputs and outputs are observable, but whose internal decision-making path across billions of parameters cannot be easily interpreted by humans.
A prompting and reasoning paradigm that guides a Large Language Model to break a complex problem down into explicit intermediate logical steps before arriving at its final answer.
A software application designed to simulate human conversation via text or voice—ranging from simple rule-based scripts to advanced LLM-powered assistants.
A supervised learning task in which the model learns to assign input data into predefined categories (for example, classifying an incoming email as "Spam" or "Legitimate").
An unsupervised machine learning technique that automatically groups data points with similar characteristics into clusters without pre-existing labels (e.g., behavioral customer segmentation).
Technological platforms that combine machine learning, natural language processing, and reasoning to mimic the way the human brain synthesizes context and supports complex decisions.
The field of artificial intelligence that enables computers to inspect, interpret, and extract meaningful information from digital images, medical scans, and live video streams.
A performance evaluation table for classification models that compares actual target values against model predictions across True Positives, False Positives, True Negatives, and False Negatives.
A training methodology (pioneered by Anthropic) in which an AI model critiques and refines its own outputs against an explicit constitution of ethical and safety principles.
The maximum amount of information (measured in tokens) that an LLM can hold in its active working memory and process in a single conversation—ranging from a few pages to over 1,000,000 tokens.
The suite of technologies enabling natural, context-aware human-machine dialogue across text and speech interfaces.
A deep learning architecture specifically designed for processing grid-like data such as images, using convolutional filters to detect edges, shapes, and complex visual objects.
An embedded AI assistant integrated directly into daily software tools (office suites, IDEs, CRMs) that collaborates alongside the human user to draft content, analyze data, and write code.
A massive, structured collection of written or spoken texts in one or more languages used to train and evaluate Natural Language Processing (NLP) and Large Language Models.
A statistical evaluation technique that partitions data into multiple subsets to verify that a machine learning model generalizes reliably to unseen data.
A technique used to artificially expand the size and diversity of a training dataset by creating modified versions of existing data (e.g., rotating images or paraphrasing text).
The analytical process of exploring massive datasets using statistical and machine learning algorithms to uncover hidden correlations, anomalies, and commercial trends.
An interdisciplinary field combining statistics, computer science, machine learning, and domain expertise to extract actionable knowledge from structured and unstructured data.
A supervised machine learning model structured like a flowchart, where each internal node tests an attribute, each branch represents a decision rule, and each leaf node represents an outcome.
An advanced subfield of machine learning based on multi-layered artificial neural networks capable of learning hierarchical representations directly from raw text, images, and audio.
Synthetic video, audio, or imagery generated by deep learning models that convincingly replaces a person’s likeness or voice, creating significant cybersecurity and fraud risks.
The generative AI architecture behind modern image and video synthesis tools that learns to generate high-resolution visuals by iteratively removing statistical noise.
The process of reducing the number of random variables in a dataset while preserving its essential structure and predictive signal (e.g., via PCA or t-SNE).
A compression technique in which a compact, fast "student" model is trained to reproduce the reasoning and behavior of a massive "teacher" model at a fraction of the compute cost.
Running artificial intelligence inference locally on physical hardware devices (smartphones, IoT sensors, vehicles) rather than remote cloud servers, ensuring ultra-low latency and data privacy.
Dense numerical arrays (vectors) that represent words, sentences, or images in a high-dimensional mathematical space where items with similar meaning are positioned close together.
A machine learning paradigm that combines the predictions of multiple distinct models to achieve higher accuracy and robustness than any single model could deliver alone.
One complete pass of the entire training dataset forward and backward through the neural network during the learning process.
The world’s first comprehensive legal framework for artificial intelligence, classifying AI systems by risk level and imposing binding transparency, governance, and compliance obligations on providers and corporate deployers.
A set of methods and design principles that allow human auditors and users to understand, trace, and trust why an AI model arrived at a specific prediction or decision.
The process of using domain knowledge to select, transform, and construct the most informative input variables (features) from raw data to improve model accuracy.
A decentralized machine learning approach that trains an algorithm across multiple edge devices or servers holding local data samples, without ever exchanging the raw data itself.
Providing an LLM with a small number of concrete input-output examples (typically 2 to 5) directly inside the prompt so it immediately adopts the exact desired structure and tone.
Taking a pre-trained foundation model and continuing its training on a smaller, domain-specific dataset so it masters specialized terminology, formatting, or tasks.
A massive AI model trained on vast, broad-spectrum data at scale (such as Gemini, GPT, or Claude) that can be adapted to a wide range of downstream tasks.
A deep learning architecture in which two neural networks—a Generator and a Discriminator—compete against each other, driving the creation of highly realistic synthetic data.
The class of artificial intelligence models capable of creating original content—including text, code, images, video, audio, and synthetic data—in response to user prompts.
Connecting an LLM’s generation process to verifiable, real-time external sources of truth (such as Google Search, corporate databases, or verified documents) to eliminate hallucinations.
Programmable safety filters, validation rules, and policy boundaries placed around an AI system to prevent it from leaking sensitive data, violating brand policies, or generating harmful outputs.
The phenomenon in which an LLM generates a response that sounds fluent, confident, and grammatical, but is factually false or includes fabricated citations.
Practical problem-solving rules or shortcuts used by algorithms to find satisfactory solutions quickly when an exhaustive search is computationally impractical.
An operational workflow design in which critical, ambiguous, or high-stakes AI decisions require explicit review, validation, or approval by a human expert before execution.
External configuration variables (such as learning rate, batch size, or network depth) set by engineers prior to training that govern how a model learns.
A computer vision capability that identifies and classifies specific objects, people, text, or places within digital photographs or video frames.
The live operational phase in which a trained AI model processes new user inputs (prompts or data) and computes predictions or responses—driving your ongoing API token costs.
A specialized fine-tuning process that trains a base language model to follow complex human instructions accurately and helpfully across diverse tasks.
A structured network of real-world entities (people, companies, products, concepts) and the explicit semantic relationships connecting them, enabling machines to reason with factual accuracy.
A deep learning model with billions or trillions of parameters trained on massive text corpora to understand, summarize, translate, reason, and generate human-like language and code.
An abstract, compressed multi-dimensional mathematical space in which a neural network maps internal representations of complex data features.
A crucial training hyperparameter that controls how aggressively a neural network adjusts its internal weights in response to the estimated error at each step.
A mathematical function that quantifies the exact discrepancy between a model’s predicted output and the true target value, guiding the optimization process.
A core branch of AI in which algorithms learn patterns from historical data to make predictions or decisions without being explicitly programmed with rigid rules.
The automated translation of text or speech from one natural language to another using neural networks that preserve context and idiomatic meaning.
An efficient neural architecture (used in models like Gemini 4 Argon) that divides a massive model into specialized "expert" sub-networks, activating only the relevant experts for each token.
An open standard that acts as a universal connector ("USB-C for AI"), enabling LLMs and AI Agents to securely read data from and execute actions across corporate tools, databases, and APIs.
Artificial intelligence models natively engineered to process, cross-reference, and generate multiple data types simultaneously—including text, images, audio, video, and code.
An NLP task that automatically locates and classifies key entities within unstructured text—such as person names, organizations, locations, dates, and monetary values.
The sub-discipline of NLP focused on transforming structured data or internal representations into fluent, coherent human-readable text.
The field of AI that gives computers the ability to read, parse, interpret, and manipulate human language.
The branch of NLP focused on comprehending the deeper intent, sentiment, and semantic meaning behind human sentences despite typos or ambiguity.
A computational model inspired by the biological neural networks of the human brain, consisting of interconnected layers of nodes that process information and learn patterns.
Technology that converts scanned paper documents, PDFs, or photographs of text into machine-readable, searchable, and editable digital text.
A modeling error that occurs when an algorithm memorizes the noise and specific quirks of its training data so closely that it fails to generalize accurately to new, unseen data.
The internal numerical variables (weights and biases) learned by a neural network during training that encode its knowledge and determine how inputs are transformed into outputs.
The simplest foundational unit of an artificial neural network—a binary linear classifier that laid the historical groundwork for modern deep learning.
The initial, compute-intensive phase of building a foundation model, in which it learns grammar, world knowledge, and reasoning structures from trillions of tokens of unlabeled data.
The use of historical data, statistical modeling, and machine learning to forecast future business outcomes, customer churn, or demand trends.
The input instruction, question, context, or data provided by a user or software system to guide a Generative AI model toward producing a specific output.
The discipline of structuring instructions, system roles, constraints, and examples so that Large Language Models deliver accurate, reliable, and business-ready outputs.
A cybersecurity vulnerability in which an attacker crafts malicious input text that tricks an LLM into ignoring its system instructions, leaking sensitive data, or executing unauthorized tool calls.
An optimization technique that reduces the numerical precision of a model’s weights (e.g., from 16-bit floating point to 8-bit or 4-bit integers) to shrink memory usage and accelerate inference.
Next-generation LLMs (such as Gemini 4 Argon and Gemini 3.8 Thinking) that allocate dedicated internal compute to plan, verify hypotheses, and self-correct before outputting their final answer.
A class of neural networks designed for sequential data (such as time series or speech) where connections form directed cycles, allowing information to persist across steps.
A machine learning paradigm in which an agent learns to make optimal sequential decisions by interacting with an environment and receiving numerical rewards or penalties.
The alignment technique used to refine LLMs by training a reward model based on human rankings of model responses, ensuring helpfulness and safety.
An architectural pattern that connects an LLM to your private corporate knowledge base (PDFs, ERP, CRM), retrieving verified relevant excerpts before generating an answer.
An NLP technique that evaluates text (such as customer reviews, survey responses, or social mentions) to determine whether the underlying emotional tone is positive, negative, or neutral.
The unauthorized use of consumer AI tools or personal AI accounts by employees to process corporate data without IT oversight or a signed Data Processing Agreement (DPA).
A compact, highly optimized language model (typically 1B to 10B parameters) designed to run at low cost and high speed on local hardware or edge devices.
AI audio technologies that transcribe spoken human voice into written text (STT) or synthesize natural-sounding human speech from written text (TTS).
The most common machine learning paradigm, in which a model is trained on labeled input-output pairs so it can accurately predict outcomes for new inputs.
Data artificially generated by computer simulations or AI models that mirrors the statistical properties of real-world data without containing actual personal or confidential records.
The foundational, persistent instruction set given to an LLM by the developer that defines its persona, operational rules, tone, and security boundaries across an entire session.
A generation parameter (typically 0.0 to 1.0+) that controls the randomness of token selection: low temperature yields deterministic, factual answers, while higher temperature increases creative variety.
The fundamental multi-dimensional mathematical array used to represent data, inputs, and weights inside deep learning frameworks such as TensorFlow and PyTorch.
The basic unit of text or media processed by an LLM (roughly 3/4 of an English word), which serves as the unit of measurement for context windows, speed, and API billing.
A machine learning method in which knowledge gained while training a model on one large general task is reused as the starting point for a specialized task.
The revolutionary deep learning architecture introduced by Google Research in 2017 ("Attention Is All You Need") that processes sequences in parallel and powers all modern LLMs.
The classic test proposed by Alan Turing in 1950 to evaluate whether a machine can exhibit conversational behavior indistinguishable from that of a human.
A modeling failure in which an algorithm is too simple to capture the underlying structure of the data, resulting in poor accuracy on both training and test sets.
A machine learning paradigm in which algorithms analyze unlabeled data to discover hidden structures, anomalies, or natural groupings without human supervision.
A specialized database engineered to store, index, and query high-dimensional vector embeddings at millisecond speeds—serving as the memory backbone of enterprise RAG systems.
Asking an AI model to perform a task directly using only instructions, without providing any prior input-output examples in the prompt.
Explore the Other FAQ Glossaries
Found this useful? Share it with your network
Help other executives and teams make better technology & marketing decisions.
Want to turn terminology into practical strategy?
Let’s discuss how AI, Marketing, and Google Workspace can work together as one unified system inside your organization.