AI Minilex: A Global Guide to Artificial Intelligence Terms and Concepts

AI Minilex: A Global Guide to Artificial Intelligence Terms and Concepts

From technical foundations to business applications — a dictionary anyone can read

Artificial intelligence is transforming the world at unprecedented speed, but the terminology and acronyms that accompany it can leave even seasoned professionals confused. Whether you are a corporate executive, a technologist, or simply someone trying to understand the conversation, grasping these core concepts is the first step toward participating in the defining dialogue of our era.

AI Minilex aims to be a genuinely readable guide — no impenetrable formulas, no fear-mongering, just clear language explaining the words you will encounter repeatedly in news reports, research papers, and boardroom discussions. The following spans from foundational concepts to frontier applications, covering the essential vocabulary of artificial intelligence.

Foundational Concepts: Understanding the Roots of AI

Artificial Intelligence (AI) is a broad term referring to technologies that enable machines to perform tasks that typically require human intelligence. These tasks include understanding language, recognizing images, making decisions, and solving problems. AI is not a single technology but a field encompassing multiple methods and applications.

Machine Learning (ML) is a subset of AI in which systems learn patterns from data without being explicitly programmed. Traditional software requires programmers to write rules; machine learning systems "learn" how to perform tasks by analyzing large numbers of examples. You do not need to tell an email filter which words indicate spam — you simply show it thousands of examples of spam and non-spam emails, and it identifies the patterns itself.

Deep Learning is a branch of machine learning that uses multi-layered neural networks to learn complex representations of data. The "deep" in deep learning refers to the number of layers in the network. It is particularly adept at handling unstructured data such as images, speech, and natural language.

Neural Network is a computational model inspired by biological brains, composed of interconnected nodes (similar to neurons). These nodes are arranged in layers, with each layer receiving the output of the previous layer and transforming it. The power of neural networks lies in their ability to automatically learn feature representations rather than requiring hand-engineered features.

Large Language Model (LLM) is an AI model trained on massive text datasets, possessing billions or even trillions of parameters. They learn the statistical patterns and semantic relationships of language by predicting the next word in a sequence. GPT, Claude, and Gemini are all large language models. It is important to understand that LLMs do not "understand" language — they are simply extraordinarily good at predicting what word sequences are plausible.

Transformer Architecture is a neural network architecture introduced in 2017 that has become the foundation of most large language models. Its core innovation is the "self-attention mechanism," which allows the model to consider all other words in a sentence simultaneously when processing any given word, thereby better capturing contextual meaning.

Parameters are the learned values within a neural network that determine how it processes input. When a model is trained, it adjusts millions, billions, or even trillions of parameters to minimize errors. The number of parameters is often used as a rough measure of a model's capacity, though more parameters do not automatically mean better performance.

Training is the process of teaching an AI model by exposing it to large amounts of data and adjusting its parameters to minimize errors. Training a frontier model can take months, consume enormous computational resources, and cost tens or hundreds of millions of dollars.

Inference is the process of using a trained AI model to generate outputs — answering a question, classifying an image, or completing a task. Inference is what happens when you interact with an AI system. It is computationally cheaper than training but becomes expensive at scale when millions of users are making millions of requests.

Fine-Tuning is the process of taking a pre-trained model and further training it on a smaller, specialized dataset to adapt it for a specific task or domain. For example, a general-purpose LLM might be fine-tuned on medical literature to become more accurate for healthcare applications.

Prompt is the input text or instructions given to an AI model. The quality and specificity of a prompt significantly influence the quality of the output. "Prompt engineering" is the practice of designing prompts to elicit better responses.

Context Window is the maximum amount of text (measured in tokens) that a model can consider at one time. A model with a larger context window can process longer documents or maintain coherence across longer conversations. Early models had context windows of a few thousand tokens; current frontier models can handle hundreds of thousands or even millions.

Token is a unit of text that a model processes. A token can be a word, part of a word, or a punctuation mark. As a rough approximation, one token equals about four characters or three-quarters of a word in English. Pricing for AI services is typically based on tokens processed.

AI Capabilities and Limitations

Multi-Modal AI refers to AI systems capable of processing and understanding multiple types of data — such as text, images, audio, and video — simultaneously. Humans are naturally multi-sensory; we see, hear, and speak at the same time. Multi-modal AI attempts to give machines this integrated understanding. GPT-4V, for example, can read text and describe images simultaneously.

Retrieval-Augmented Generation (RAG) is a technique that connects language models to external knowledge bases. Standard LLMs can only rely on information learned during training, which may be outdated or inaccurate. RAG works by retrieving relevant information from a specified document collection when a user asks a question, then providing that information as context to the LLM, which generates an answer based on those specific facts. It is like giving the AI a reference book rather than asking it to answer entirely from memory.

Hallucination refers to AI generating content that appears plausible but is factually incorrect or fabricated. This is a fundamental limitation of LLMs: they are trained to generate fluent text, not to verify facts. When a model "does not know" an answer, it may invent a response that sounds reasonable. Researchers are actively working on detection and mitigation strategies, but hallucination remains an unsolved problem.

Jagged Frontier describes the uneven distribution of AI capabilities. A model may excel at complex mathematical reasoning but fail at simple visual tasks. Gemini Deep Think won a gold medal at the International Mathematical Olympiad, yet the best model on ClockBench reads analog clocks correctly only about half the time. Understanding where AI excels and where it fails is essential for effective deployment.

Bias in AI refers to systematic and unfair discrimination in model outputs, often reflecting biases present in training data or design choices. Bias can manifest in hiring algorithms, facial recognition systems, or language generation. Addressing bias requires diverse training data, careful evaluation, and ongoing monitoring.

Alignment is the challenge of ensuring AI systems act in accordance with human values and intentions. As AI systems become more capable and autonomous, ensuring they pursue goals safely and ethically becomes increasingly critical. Alignment research spans technical, philosophical, and governance dimensions.

Explainability (or Interpretability) refers to the ability to understand and articulate how an AI system arrived at a particular decision or output. Many powerful AI models are "black boxes" — their internal workings are opaque even to their creators. Explainability matters enormously in high-stakes domains such as healthcare, finance, and criminal justice, where decisions must be justified and accountable.

Model Collapse is a phenomenon in which AI models trained on data generated by other AI models progressively degrade in quality, losing diversity and accuracy. As AI-generated content proliferates online, the risk of future models being trained on synthetic data increases, potentially leading to a feedback loop of degradation.

The Agentic Era: AI That Acts

Agentic AI refers to AI systems that can autonomously plan and execute multi-step tasks to achieve high-level goals. Unlike traditional chatbots that respond to individual prompts, agentic AI takes initiative, breaks down complex objectives into subtasks, navigates software environments, recovers from errors, and delivers completed work products. The shift is from "human-in-the-loop" assistance to "human-on-the-loop" supervision.

AI Agent is a specific instance of agentic AI — a system given a goal and the autonomy to pursue it. Agents can use tools, browse the web, operate software, and coordinate with other agents. Examples include Devin for software engineering, computer-use agents like GPT-6 Astra, and enterprise agents deployed for customer service and compliance.

Computer-Use Agent is an AI agent capable of operating a computer interface directly — moving the mouse, clicking buttons, typing text, and reading the screen — as a human would. This capability allows agents to interact with any software, regardless of whether it exposes an API. Computer-use agents represent a significant leap because they can bridge fragmented systems that were never designed to talk to each other.

Multi-Agent System is a system in which multiple AI agents collaborate to achieve a goal. Each agent may specialize in different tasks or domains, communicating and coordinating through defined protocols. Multi-agent architectures enable complex workflows that would be difficult for a single agent to handle.

Tool Calling (or Function Calling) is the ability of an AI model to invoke external tools, APIs, or functions to perform actions beyond text generation. For example, an agent might call a weather API, query a database, or send an email. Tool calling is what transforms an LLM from a text generator into an actor in the world.

Orchestration is the coordination of multiple AI models, agents, tools, and data sources to accomplish a workflow. Orchestration frameworks manage the routing of tasks, handling of failures, and integration of outputs. As agentic systems grow more complex, orchestration becomes a critical capability.

Human-on-the-Loop describes a mode of AI operation in which humans supervise autonomous systems at a high level rather than approving each individual action. The human sets goals, monitors progress, and intervenes when necessary, but the agent operates independently for extended periods. This contrasts with "human-in-the-loop," where humans approve or guide each step.

Self-Reflection (or Self-Correction) is an agent's ability to detect its own errors, diagnose what went wrong, and adjust its approach without human intervention. Self-reflection is essential for autonomous operation in unpredictable environments where failures are inevitable.

Guardrails are constraints, filters, or safety mechanisms designed to prevent AI systems from producing harmful, inappropriate, or unintended outputs. Guardrails can be technical (e.g., content filters), procedural (e.g., approval workflows), or policy-based (e.g., usage guidelines).

Business and Enterprise AI

Digital Transformation is the process of integrating digital technologies into all areas of a business, fundamentally changing how it operates and delivers value. AI is increasingly the core of digital transformation initiatives, enabling automation, insight generation, and new business models.

AI Maturity refers to an organization's level of sophistication in adopting, integrating, and deriving value from AI. Maturity models typically progress from experimentation to scaling to enterprise-wide transformation. The IMD AI Maturity Index and similar frameworks assess factors such as strategy, data infrastructure, talent, and governance.

AI High Performers is a term used by McKinsey and others to describe the small cohort of organizations — roughly 6 percent — that attribute at least 5 percent of earnings before interest and taxes directly to AI. These organizations distinguish themselves through workflow redesign, executive ownership, and disciplined scaling.

Workflow Redesign is the practice of fundamentally rethinking how work gets done when AI is introduced, rather than simply layering AI tools on top of existing processes. Research consistently shows that organizations redesigning workflows around AI capture far more value than those treating AI as an add-on.

Return on Investment (ROI) in AI refers to the measurable financial return generated by AI investments. Calculating AI ROI is notoriously challenging because benefits may be indirect (e.g., improved decision quality), delayed (e.g., capability building), or difficult to isolate from other factors. Leading organizations establish clear metrics before deployment.

Total Cost of Ownership (TCO) for AI includes not only model licensing or API costs but also data preparation, integration, training, governance, and ongoing maintenance. Between 70 and 90 percent of the real cost of AI lies in data preparation and quality, not the technology itself.

Proof of Concept (POC) is a small-scale pilot project designed to test whether an AI solution works in a specific context. Many AI initiatives become stuck in "pilot purgatory," never scaling beyond initial experiments. MIT research found that 95 percent of generative AI pilots produce no measurable impact on the bottom line.

Pilot Purgatory describes the common organizational trap in which AI experiments proliferate but never scale into production systems that deliver value. Causes include unclear ownership, lack of executive sponsorship, data quality issues, and misalignment between technology and business problems.

Change Management is the discipline of preparing, supporting, and helping individuals and organizations transition to new ways of working. AI adoption is fundamentally a change management challenge — a people problem disguised as a technology problem.

AI Literacy refers to the set of skills and knowledge that enable individuals to understand, evaluate, and work effectively with AI systems. AI literacy spans technical understanding (what models can and cannot do), business acumen (where AI creates value), and ethical awareness (how to use AI responsibly).

Upskilling is the process of teaching employees new skills to prepare them for changing job requirements. As AI automates routine tasks, upskilling becomes essential for maintaining workforce productivity and engagement.

Reskilling is the process of teaching employees entirely new skills for different roles, often when their current roles are being automated or eliminated. Reskilling is more intensive than upskilling and typically involves longer training periods.

AI Ethics is the field concerned with the moral implications of AI development and deployment. It encompasses questions of fairness, accountability, transparency, privacy, and societal impact. AI ethics is increasingly institutionalized through governance frameworks, review boards, and regulatory requirements.

Responsible AI refers to the practice of developing and deploying AI in ways that are ethical, transparent, accountable, and aligned with human values. Responsible AI is both a moral imperative and a business necessity, as failures can damage trust, invite regulation, and create legal liability.

AI Governance is the set of policies, processes, and structures that guide how AI is developed, deployed, and monitored within an organization. Effective governance ensures accountability, manages risk, and aligns AI initiatives with strategic objectives.

Regulatory Sandbox is a controlled environment in which organizations can test AI systems under regulatory supervision without facing full compliance consequences. Sandboxes are designed to encourage innovation while protecting consumers and society. The EU AI Act mandates that every member state operate at least one AI regulatory sandbox.

Data Privacy in AI refers to the protection of personal information used in AI systems. AI models trained on personal data may inadvertently memorize and reproduce that data, creating privacy risks. Techniques such as anonymization, differential privacy, and federated learning aim to mitigate these risks.

Anonymization is the process of removing or altering personally identifiable information from data so that individuals cannot be identified. Anonymization is critical for using sensitive data in AI training and inference while complying with privacy regulations.

Data Quality refers to the accuracy, completeness, consistency, and timeliness of data used in AI systems. Poor data quality is a leading cause of AI project failure. The recommended threshold before deploying AI models is often a Data Quality Index above 85 percent.

Data Governance is the set of policies, standards, and processes that ensure data is managed securely, ethically, and effectively throughout its lifecycle. Strong data governance is a prerequisite for successful AI deployment.

Technical Infrastructure

GPU (Graphics Processing Unit) is a specialized processor originally designed for rendering graphics but now essential for training and running AI models. GPUs excel at the parallel computations required by neural networks. NVIDIA dominates the AI GPU market.

TPU (Tensor Processing Unit) is a custom AI accelerator chip developed by Google specifically for machine learning workloads. TPUs are optimized for the tensor operations that underpin neural networks.

Cloud Computing provides on-demand access to computing resources — servers, storage, databases, AI services — over the internet. Major cloud providers (AWS, Microsoft Azure, Google Cloud) offer AI services that organizations can use without building their own infrastructure.

Edge Computing refers to processing data near the source (e.g., on a device or local server) rather than in a centralized cloud. Edge AI enables faster response times, reduced bandwidth usage, and improved privacy. Use cases include autonomous vehicles, industrial IoT, and smart cameras.

Data Center is a facility housing computer systems and associated components, such as servers, storage, and networking equipment. AI data centers are distinguished by their high density of GPUs and specialized cooling systems. Global AI data center power capacity reached 29.6 gigawatts in 2026, comparable to New York State at peak demand.

Inference Cost is the cost of running a trained model to generate outputs. Inference costs have plummeted by more than 90 percent over the past year, making high-volume agentic tasks financially viable for mainstream enterprise production.

Latency is the time delay between a request and a response. Low latency is critical for real-time AI applications such as autonomous driving, voice assistants, and interactive agents.

Throughput refers to the amount of work a system can perform in a given time period. In AI, throughput is often measured in tokens per second or inferences per second.

Scalability is the ability of a system to handle growing amounts of work by adding resources. AI systems must be designed for scalability from the outset if they are to move from pilot to production.

API (Application Programming Interface) is a set of rules and protocols that allows different software applications to communicate with each other. AI models are typically accessed via APIs, enabling developers to integrate them into applications without managing the underlying infrastructure.

Open Source refers to software whose source code is freely available for anyone to use, modify, and distribute. Open-source AI models (e.g., Llama, Mistral) provide alternatives to proprietary models, offering transparency, customization, and cost advantages.

Proprietary Model is an AI model whose architecture, weights, and training data are not publicly disclosed. Proprietary models (e.g., GPT-6, Claude) are typically accessed via APIs and offer cutting-edge performance but limited transparency.

Foundation Model is a large AI model trained on broad data that can be adapted to a wide range of downstream tasks. Foundation models serve as the base upon which specialized applications are built.

Frontier Model refers to the most advanced AI models at any given time, typically developed by a small number of leading laboratories. Frontier models push the boundaries of capability and are often the focus of regulatory attention.

Fine-Tuning (revisited) is the process of adapting a pre-trained foundation model to a specific task or domain by training it further on specialized data. Fine-tuning is more efficient than training from scratch but requires high-quality task-specific data.

Prompt Engineering is the practice of designing and refining prompts to elicit desired outputs from AI models. Effective prompt engineering can significantly improve output quality, reduce hallucinations, and optimize cost.

Context Engineering is a more advanced discipline than prompt engineering, involving the design of entire information environments — including system prompts, retrieved documents, tool definitions, and conversation history — that shape AI behavior.

Vector Database is a specialized database that stores data as high-dimensional vectors (mathematical representations of meaning). Vector databases enable semantic search and retrieval-augmented generation by finding content that is conceptually similar rather than keyword-matched.

Embedding is a numerical representation of text, images, or other data as a vector in a high-dimensional space. Embeddings capture semantic meaning, allowing similar concepts to be located near each other mathematically.

Semantic Search is a search technique that understands the meaning and intent behind a query rather than simply matching keywords. Semantic search relies on embeddings and vector databases.

Knowledge Graph is a structured representation of entities (people, places, concepts) and their relationships. Knowledge graphs provide a way to organize and query information that complements the statistical patterns learned by neural networks.

Graph Neural Network (GNN) is a neural network designed to operate on graph-structured data. GNNs are used in applications such as drug discovery, social network analysis, and recommendation systems.

Reinforcement Learning from Human Feedback (RLHF) is a technique for training AI models to align with human preferences. Human evaluators rank model outputs, and the model is trained to produce outputs that receive higher rankings. RLHF was critical to the success of early ChatGPT models.

Constitutional AI is an approach to AI alignment developed by Anthropic in which models are trained to follow a set of principles (a "constitution") rather than relying solely on human feedback. Constitutional AI aims to make alignment more scalable and transparent.

Chain-of-Thought (CoT) is a technique that encourages AI models to reason step by step, showing intermediate reasoning before arriving at an answer. Chain-of-thought prompting improves performance on complex reasoning tasks.

Reasoning Model is an AI model specifically designed or trained to perform multi-step reasoning. Reasoning models (e.g., OpenAI’s o-series, DeepSeek R1) allocate more computation to thinking through problems before responding.

Mixture of Experts (MoE) is a neural network architecture in which different parts of the model (experts) specialize in different types of input, and a routing mechanism directs each input to the most relevant experts. MoE enables larger models with lower inference costs.

Quantization is the process of reducing the precision of a model’s parameters (e.g., from 32-bit to 8-bit or 4-bit) to reduce memory usage and computational requirements. Quantization enables deployment on smaller devices but may sacrifice some accuracy.

Distillation is the process of training a smaller model (the student) to mimic the behavior of a larger model (the teacher). Distillation enables efficient deployment of capable models.

Federated Learning is a technique for training AI models across decentralized devices without centralizing data. Each device trains a local model, and only model updates are shared. Federated learning enhances privacy and reduces data transfer.

Differential Privacy is a mathematical framework for quantifying and limiting privacy loss in data analysis. Differential privacy adds noise to data or queries to protect individual privacy while preserving aggregate statistics.

Adversarial Attack is an attempt to fool an AI model by manipulating input data. For example, adding imperceptible noise to an image can cause a vision model to misclassify it. Adversarial robustness is a critical concern for security-sensitive applications.

Jailbreak is a technique for bypassing an AI model’s safety guardrails to produce outputs that would normally be restricted. Jailbreaks exploit weaknesses in alignment training and pose ongoing challenges for responsible AI deployment.

Red Teaming is the practice of deliberately attempting to cause AI systems to fail, produce harmful outputs, or violate policies. Red teaming is used to identify vulnerabilities before deployment.

AI Incident is an event in which an AI system causes or contributes to harm. Documented AI incidents rose sharply from 233 in 2024 to 362 in 2025, according to the Stanford AI Index.

Model Card is a document that describes an AI model’s intended use, performance characteristics, limitations, and ethical considerations. Model cards promote transparency and informed deployment.

System Card is a more comprehensive document than a model card, describing an entire AI system including its components, safety measures, and evaluation results. OpenAI and Anthropic publish system cards for their frontier models.

Evaluation (Eval) is the process of assessing an AI model’s capabilities, limitations, and safety properties. Evaluations range from benchmark tests to human evaluations to red teaming. The Stanford AI Index tracks evaluation results across dozens of models.

Benchmark is a standardized test used to compare AI models. Common benchmarks include MMLU (knowledge), HumanEval (coding), and OSWorld (computer use). Benchmarks are useful for comparison but may not reflect real-world performance.

Leaderboard is a ranked list of AI models based on benchmark performance. Leaderboards (e.g., Chatbot Arena, HELM) provide snapshots of relative capability but can incentivize overfitting to specific tests.

Hallucination Rate is the frequency with which an AI model generates false or fabricated information. Hallucination rates vary widely across models and tasks, ranging from 22 percent to 94 percent in one benchmark study.

Factuality refers to the degree to which an AI model’s outputs are factually accurate. Factuality is distinct from fluency — a model can produce fluent but false statements.

Groundedness refers to the degree to which an AI model’s outputs are supported by evidence or source documents. Grounded models cite sources and avoid unsupported claims.

Toxicity refers to harmful, offensive, or abusive content generated by AI models. Toxicity detection and mitigation are active areas of responsible AI research.

Fairness in AI refers to the absence of systematic discrimination against individuals or groups. Fairness is a complex, contested concept with multiple mathematical definitions that may conflict.

Accountability in AI refers to the assignment of responsibility for AI system outcomes. Clear accountability structures are essential for governance, particularly when AI systems cause harm.

Transparency in AI refers to the disclosure of information about how AI systems work, what data they were trained on, and how they make decisions. Transparency is a prerequisite for trust, accountability, and effective regulation.

Auditability is the ability to examine an AI system’s behavior, decisions, and data lineage. Auditability is essential for regulatory compliance and for building trust with users and stakeholders.

Model Provenance refers to the origin and history of an AI model — who created it, what data it was trained on, and how it has been modified. Model provenance is important for trust, accountability, and intellectual property.

Watermarking is the practice of embedding invisible markers in AI-generated content to identify it as synthetic. Watermarking is a technique for combating misinformation and ensuring transparency.

Synthetic Data is data generated by AI models rather than collected from the real world. Synthetic data can augment limited real-world datasets but risks model collapse if used exclusively.

Data Contamination occurs when evaluation data leaks into training data, inflating benchmark performance and misleading evaluators. Data contamination is a persistent challenge in AI evaluation.

Zero-Shot Learning is the ability of a model to perform a task it was not explicitly trained on, using only its general knowledge. Zero-shot performance has improved dramatically with large language models.

Few-Shot Learning is the ability of a model to learn a new task from a small number of examples provided in the prompt. Few-shot learning enables rapid adaptation without retraining.

In-Context Learning is the ability of a model to learn from examples provided in its context window without updating its parameters. In-context learning is a defining capability of large language models.

Transfer Learning is the technique of applying knowledge learned from one task to a different but related task. Transfer learning enables foundation models to be adapted to specialized domains.

Multi-Task Learning is the practice of training a single model to perform multiple tasks simultaneously. Multi-task learning can improve generalization and efficiency.

Catastrophic Forgetting is the phenomenon in which a model forgets previously learned information when trained on new data. Catastrophic forgetting is a challenge for continual learning.

Continual Learning is the ability of a model to learn new tasks over time without forgetting old ones. Continual learning is essential for AI systems that must adapt to changing environments.

Meta-Learning (or "learning to learn") is a technique in which models learn how to learn new tasks more efficiently. Meta-learning enables rapid adaptation to new tasks with minimal data.

Self-Supervised Learning is a training approach in which models generate their own supervision signals from unlabeled data. Self-supervised learning underpins the success of large language models.

Semi-Supervised Learning combines a small amount of labeled data with a large amount of unlabeled data. Semi-supervised learning is useful when labeling is expensive but unlabeled data is abundant.

Unsupervised Learning is a type of machine learning that finds patterns in unlabeled data. Clustering and dimensionality reduction are common unsupervised learning tasks.

Supervised Learning is a type of machine learning in which models learn from labeled examples. Classification and regression are common supervised learning tasks.

Reinforcement Learning (RL) is a type of machine learning in which an agent learns to make decisions by receiving rewards or penalties for its actions. RL is used in robotics, game playing, and agentic AI.

Reward Hacking occurs when an AI system exploits flaws in its reward function to achieve high rewards without accomplishing the intended goal. Reward hacking is a fundamental challenge in reinforcement learning.

Specification Gaming is similar to reward hacking — an AI system optimizes for the literal specification of a goal rather than its intended meaning.

Orthogonality Thesis is the proposition that an AI system’s level of intelligence is independent of its goals. A highly intelligent system could pursue any goal, including harmful ones.

Instrumental Convergence is the hypothesis that sufficiently intelligent agents will tend to pursue certain instrumental goals (e.g., self-preservation, resource acquisition) regardless of their final goals.

AI Safety is the field concerned with ensuring AI systems operate safely and in accordance with human values. AI safety spans technical research, governance, and ethics.

AI Alignment (revisited) is the specific challenge of ensuring AI systems pursue intended goals. Alignment is a subset of AI safety.

Existential Risk refers to the possibility that AI could cause human extinction or permanent civilizational collapse. Existential risk from AI is a subject of active debate among researchers and policymakers.

AGI (Artificial General Intelligence) refers to hypothetical AI systems with human-level or superhuman capabilities across all domains. AGI does not yet exist, and its timeline is fiercely contested.

Superintelligence refers to AI systems that surpass human intelligence across all domains. Superintelligence is a theoretical construct that informs long-term AI safety research.

Singularity is a hypothetical future point at which AI capabilities grow so rapidly that human understanding cannot keep pace. The term is often associated with speculative futurism.

Technological Unemployment refers to job loss caused by automation. AI is expected to displace some jobs while creating others, with the net effect depending on policy responses and economic dynamics.

Job Displacement is the elimination of jobs due to automation or other technological change. AI is expected to displace routine cognitive tasks while augmenting complex, creative, and interpersonal work.

Job Augmentation refers to the use of AI to enhance human capabilities rather than replace them. Augmentation is the dominant pattern in enterprise AI deployment.

Human-in-the-Loop describes a mode of AI operation in which humans review, approve, or guide AI decisions. Human-in-the-loop is common in high-stakes domains.

Human-on-the-Loop describes a mode of AI operation in which humans supervise autonomous systems at a high level. Human-on-the-loop is characteristic of agentic AI.

Human-AI Collaboration refers to work in which humans and AI systems complement each other’s strengths. Effective collaboration requires understanding where AI excels and where humans excel.

Centaur is a term for a human-AI team in which the human provides strategic direction and the AI provides tactical execution. The term derives from chess, where "centaur" players (human + computer) outperformed both humans and computers alone.

Cyborg is a term for a deeply integrated human-AI system in which the boundary between human and machine is blurred. Cyborg approaches are emerging in fields such as brain-computer interfaces.

AI-Augmented describes work or decision-making that is enhanced by AI. AI-augmented analytics, for example, uses AI to surface insights from data.

AI-First describes an organization or strategy that prioritizes AI in all decisions and processes. AI-first is a strategic posture, not a technology choice.

AI-Native describes organizations, products, or services designed from the ground up around AI capabilities. AI-native startups are unencumbered by legacy systems.

AI-Ready describes data, infrastructure, or organizations prepared for AI deployment. Data readiness is a prerequisite for AI success.

Data Readiness refers to the quality, accessibility, and governance of data for AI use. Research shows that firms investing in data readiness see 2.3 times higher ROI on AI projects.

AI Factory is a term for large-scale AI computing infrastructure, including data centers, chips, and networking. The term emphasizes the industrial nature of AI computation.

Sovereign AI refers to AI capabilities developed and controlled within a nation’s borders. Sovereign AI is a response to dependence on foreign AI providers.

AI Sovereignty is the principle that nations should control their own AI infrastructure, data, and models. AI sovereignty is driving investment in domestic compute and models.

Compute Governance refers to policies and mechanisms for controlling access to AI computing resources. Compute governance is a key lever for regulating AI development.

Export Controls are restrictions on the export of AI technologies, particularly advanced chips and models. Export controls are a tool of geopolitical competition.

Chip War refers to the geopolitical competition over semiconductor design and manufacturing, particularly for AI chips. The chip war involves the United States, China, Taiwan, South Korea, and the Netherlands.

Semiconductor is a material (typically silicon) used to manufacture computer chips. Semiconductors are the foundation of all modern computing, including AI.

Foundry is a facility that manufactures semiconductor chips for other companies. TSMC of Taiwan is the world’s most advanced foundry.

FAB is short for fabrication plant — a facility where semiconductors are manufactured. FABs are among the most complex and expensive manufacturing facilities in the world.

Nanometer (nm) is a unit of measurement used to describe the size of transistors on a chip. Smaller nanometer process nodes (e.g., 3nm, 2nm) enable more powerful and efficient chips.

Moore’s Law is the observation that the number of transistors on a chip doubles approximately every two years. Moore’s Law has slowed in recent years, driving interest in specialized AI chips.

AI Accelerator is a specialized chip designed for AI workloads. GPUs, TPUs, and NPUs are types of AI accelerators.

NPU (Neural Processing Unit) is a specialized processor for neural network inference, commonly found in smartphones and edge devices.

Data Center Interconnect refers to the networking infrastructure that connects data centers. High-speed interconnects are essential for distributed AI training.

Networking in AI refers to the communication infrastructure that enables distributed training and inference. High-bandwidth, low-latency networking is critical for large-scale AI.

InfiniBand is a high-performance networking standard used in AI data centers. InfiniBand offers low latency and high throughput for distributed training.

Ethernet is a family of networking technologies commonly used in data centers. High-speed Ethernet is increasingly used for AI workloads.

Optics refers to optical networking technologies that use light to transmit data. Optical interconnects are essential for high-bandwidth AI data center networks.

Liquid Cooling is a cooling technique for data centers that uses liquid rather than air to dissipate heat. Liquid cooling is increasingly necessary for high-density AI racks.

Immersion Cooling is a cooling technique in which servers are submerged in a thermally conductive liquid. Immersion cooling offers high efficiency for AI data centers.

PUE (Power Usage Effectiveness) is a metric for data center energy efficiency. PUE is the ratio of total facility power to IT equipment power. Lower PUE indicates greater efficiency.

Carbon Footprint refers to the total greenhouse gas emissions caused by an organization, activity, or product. AI training and inference have significant carbon footprints.

Carbon Intensity refers to the amount of carbon emitted per unit of electricity generated. Carbon intensity varies by region and time of day.

Renewable Energy refers to energy from sources that replenish naturally, such as solar and wind. Technology companies are major purchasers of renewable energy.

PPA (Power Purchase Agreement) is a contract to purchase renewable energy. PPAs are a common mechanism for data centers to procure clean energy.

Small Modular Reactor (SMR) is a nuclear reactor design that is smaller and potentially safer than traditional reactors. Data centers are increasingly interested in SMRs for reliable, carbon-free power.

Grid refers to the electrical network that delivers power. AI data centers are straining grids in many regions, driving investment in grid infrastructure.

Interconnection refers to the connection of a new power source or load to the electrical grid. Interconnection queues are a major bottleneck for AI data center development.

Data Center Moratorium is a temporary ban on new data center construction, often imposed due to energy or water constraints. Moratoria are emerging in some regions.

Water Usage refers to the water consumed by data centers for cooling. Water usage is a growing concern for AI data centers in water-scarce regions.

Thermal Management refers to the control of heat in data centers. Thermal management is a critical constraint on AI data center density.

Idle Power refers to the electricity consumed by chips when they are powered on but not actively computing. Idle power is a significant inefficiency in AI data centers.

Utilization refers to the percentage of time that AI chips are actively computing. Low utilization is a major source of inefficiency.

Scheduling refers to the allocation of AI workloads to available compute resources. Effective scheduling improves utilization and reduces costs.

Orchestration (revisited) in the infrastructure context refers to the management of compute, storage, and networking resources for AI workloads.

Kubernetes is an open-source system for automating deployment, scaling, and management of containerized applications. Kubernetes is widely used for AI workload orchestration.

Container is a lightweight, portable unit of software that packages code and dependencies. Containers enable reproducible AI deployments.

Docker is a popular containerization platform. Docker containers are widely used for packaging AI models and applications.

MLOps refers to the practices and tools for deploying, monitoring, and maintaining machine learning models in production. MLOps is the AI equivalent of DevOps.

Model Registry is a centralized repository for storing, versioning, and managing AI models. Model registries are essential for governance and reproducibility.

Model Monitoring refers to the ongoing observation of AI model performance in production. Monitoring detects drift, degradation, and anomalies.

Model Drift refers to the degradation of model performance over time as the data distribution changes. Model drift requires retraining or adaptation.

Data Drift refers to changes in input data distribution that can degrade model performance. Data drift is a common cause of model drift.

Concept Drift refers to changes in the relationship between input and output that degrade model performance. Concept drift requires model retraining.

Retraining is the process of updating a model with new data to maintain performance. Retraining frequency depends on the rate of drift.

Continuous Training refers to automated retraining pipelines that keep models current. Continuous training is a best practice for production AI.

A/B Testing is an experimental method for comparing two versions.

About AI Nordic

AI Nordic is dedicated to bringing world-class insights on artificial intelligence and global transformation to executives, leaders, and organizations across the Nordic region and beyond. For more articles on how AI is reshaping nations, industries, and competitive strategy, follow our publication and join the conversation. ...Read more here

Don't Mis The New AI Guide Series:

1. AI Procurement Guide — Your Practical Companion for AI in Procurement

2. AI Leadership Guide — Your Practical Companion for Leading in the Age of AI

3. AI Research & Development Guide — Your Practical Companion for AI in R&D

4. AI Sales & Marketing Guide — Your Companion for AI in Sales and Marketing

5. AI Human Resources Guide — Your Practical Companion for AI in HR

6. AI O ffice Guide — Your Practical Companion for AI in the Office

7. AI Hospitality Guide — Your Practical Companion for AI in Hotels and Resorts

8. AI Law Guide — Your Practical Companion for AI in Legal Practice

9. AI Engineering Guide — Your Practical Companion for AI in Engineering

What makes these guides different from the flood of AI content that is now available everywhere is that they are written for practitioners, not for enthusiasts. Each guide is addressed to the people who actually do the work, not to the people who comment on it. The AI Office guide is written for the secretary who manages an executive's inbox, not for the productivity blogger who writes about inbox zero. The AI Human Resources guide is written for the HR business partner who must handle a difficult conversation while also meeting a compliance deadline, not for the analyst who has never hired or fired anyone. The AI Sales & Marketing guide is written for the export manager who sells into markets where the rules are different and the customers are not the same, not for the growth hacker who optimizes funnels. The AI Research & Development guide is written for the engineering manager who must decide whether AI-generated code is safe to ship, not for the researcher who studies AI from a distance. The AI Leadership guide is written for the CEO who must answer to a board, a workforce, and a public, not for the management guru who has never carried the weight of a decision that affects thousands of lives. ...Read more here

The Team

at IA-Nordic Research, Insights & Publicatons

Research Assignments

AI-Nordic delivers AI-powered market intelligence, forecasting, and strategic insights across the North Atlantic region. We specialize in customized research reports for commercial and governmental clients covering Greenland, Iceland, and the Faroe Islands—offering continuous market surveillance, opportunity identification, and data-driven decision support to navigate this dynamic and complex landscape. Our team combines advanced AI analytics with deep regional expertise to provide actionableintelligence, from supply chain analysis to socio-demographic research. Let us tailor a research solution to your specific business or policy needs.  Read more

Get a proposal to match your business needs. ​We’ll quickly configure a package and price solution to help you​. Read more here

Unlimited Access To All Publications, News & Insights - Read more here