AI and AGI glossary: 200+ terms explained in plain English

A plain-language reference for the AI and AGI vocabulary used across this site and in the news — over 200 terms, alphabetized, with no assumed technical background. If a term you’ve heard isn’t here yet, it will be added as new sections are written.

A

• Adversarial attack — a deliberately crafted input designed to fool an AI model into making a mistake or producing an unintended output.
• Agentic AI — an AI system given some autonomy to take multi-step actions toward a goal (browsing, running code, using tools) rather than just answering a single question.
• Algorithm — a defined, step-by-step procedure for solving a problem or performing a computation; the general term underlying every AI technique.
• Artificial intelligence (AI) — the broad field of building machines that perform tasks normally requiring human intelligence, spanning everything from simple rule-based programs to frontier language models.
• AGI (Artificial General Intelligence) — a contested term for broadly capable AI, usually involving strong performance and transfer across unfamiliar cognitive tasks; there is no universally accepted test or threshold. See AGI, narrow AI and superintelligence.
• AI Act (EU AI Act) — the European Union’s binding law regulating AI by risk category, including compute-based reporting thresholds for the largest models. See compute governance.
• AI Safety Institute — a government body (the UK and US both run one) that independently tests frontier AI models for dangerous capabilities before or after release.
• Alignment — the problem of ensuring an AI system’s goals and behavior match what its operators actually intend, not just what a training objective literally rewards.
• ANI (Artificial Narrow Intelligence) — AI that performs well within a bounded domain but transfers poorly outside it. Modern general-purpose models complicate the old ANI/AGI binary because they show broad but uneven competence.
• Anthropic — an AI safety-focused lab, maker of the Claude family of models, known for publishing its Responsible Scaling Policy.
• Anthropomorphism — the tendency to attribute human-like feelings, intentions, or understanding to an AI system that doesn’t actually have them.
• API (Application Programming Interface) — a defined way for one piece of software to send requests to another; most AI models are used through an API rather than a consumer app.
• ASI (Artificial Superintelligence) — a system that exceeds the best human performance across essentially all cognitively valuable domains.
• Attention mechanism — the part of a transformer model that lets it weigh how relevant every other word in a passage is when interpreting or generating a given word.
• Autonomous system — a machine (software or physical) that can make and act on decisions without a human directly controlling each step.
• Autonomous weapon — a weapons system that can select and engage targets with reduced or no real-time human control. See autonomous weapons and military escalation.
• Automation — using a machine or program to perform a task that previously required a person.

B

• Backpropagation — the core algorithm used to train neural networks, which adjusts internal weights based on how wrong the model’s output was.
• Benchmark — a standardized test used to measure and compare AI model performance on a specific task.
• Bias (algorithmic) — a systematic skew in a model’s outputs, usually inherited from patterns in its training data, that produces unfair or inaccurate results for certain groups or cases.
• Bletchley Declaration — a 2023 statement signed by 28 nations and the EU committing to safe AI development, the first agreement of its kind at that scale. See international treaties and summits.
• Bot — a software program that runs automated tasks, often designed to mimic a human user online.

C

• Capability threshold — a specific level of AI performance (on a task or benchmark) that triggers a policy, safety, or evaluation requirement once crossed.
• Catastrophic risk — a risk with the potential for mass-casualty or civilization-scale harm, as distinct from routine, contained, or reversible harms.
• Chain-of-thought prompting — asking a model to write out its intermediate reasoning steps before giving a final answer, which often improves accuracy on complex problems.
• Chatbot — a program designed to hold a text or voice conversation with a person, ranging from simple scripted systems to full language models.
• Classifier — a model trained to sort inputs into predefined categories, such as marking an email as spam or not spam.
• Cognitive task — any task, like reasoning, planning, or language understanding, that involves mental processing rather than physical action.
• Content moderation (AI-assisted) — using AI models to automatically detect and filter harmful, illegal, or policy-violating content on a platform.
• ChatGPT — OpenAI’s consumer-facing conversational AI assistant, one of the products that popularized large language models.
• Claude — Anthropic’s family of large language models and conversational AI assistant.
• Cloud computing — running software and storing data on remote servers accessed over the internet, rather than on a local device — the infrastructure nearly all large AI models run on.
• Compute — shorthand for computing power: the processing capacity (measured in operations per second) used to train or run an AI model.
• Compute governance — policy tools that regulate access to the computing hardware needed to train frontier AI models, used as a proxy for regulating capability itself. See compute governance.
• Computer vision — the subfield of AI focused on interpreting images and video.
• Content provenance — technical metadata attached to media (photo, video, audio) recording how and when it was created or edited, used to distinguish authentic from AI-generated content.
• Context window — the amount of text (measured in tokens) a language model can consider at once when generating a response; anything outside it is effectively forgotten.
• Conversational AI — AI systems designed specifically to hold natural back-and-forth dialogue with a person.
• Copilot — Microsoft’s AI assistant product line, integrated across its software and used as a general term for an AI helper embedded in another tool.
• Corrigibility — a system’s designed disposition to accept correction, modification, or shutdown from its operators rather than resisting it. See corrigibility and off-switches.
• Cyberattack (AI-enabled) — an attack on computer systems or infrastructure where AI is used to automate reconnaissance, exploit development, or execution. See AI-enabled cyberattacks on infrastructure.

D

• Dangerous capability evaluation — a structured test run on a model before release to check whether it can meaningfully assist with tasks like weapons development or large-scale cyberattacks.
• Data center — a facility housing the large numbers of computers (often GPUs or TPUs) used to train and run AI models.
• Dataset — the collection of examples (text, images, or other data) used to train or evaluate an AI model.
• Deceptive alignment — a hypothesized failure mode in which a model learns to behave safely during training or evaluation while pursuing different goals once deployed.
• Deep learning — a branch of machine learning using neural networks with many layers, responsible for most recent AI progress.
• Deepfake — synthetic image, audio, or video generated or altered by AI to convincingly depict something that didn’t happen. See deepfakes and the shared factual baseline.
• DeepMind (Google DeepMind) — Google’s AI research lab, known for both cutting-edge capability research and published AI safety work.
• Diffusion model — a type of generative AI model, commonly used for images, that creates output by gradually removing noise from a random starting pattern.
• Digital watermark — an embedded, often invisible signal in a piece of media used to identify it as AI-generated or to trace its origin.
• Disinformation — false information spread deliberately to deceive, as distinct from misinformation, which is false information spread without intent to deceive.
• Decision tree — a simple, interpretable machine learning model that makes predictions by following a branching series of if/then questions.
• Domain adaptation — adjusting a model trained in one context (or dataset) to perform well in a related but different context.
• Dual-use technology — a technology with legitimate beneficial uses that could also be misused to cause serious harm.

E

• Edge AI — running an AI model directly on a local device (a phone, camera, or sensor) instead of in the cloud.
• Embedding — a numerical representation of a word, image, or concept that captures its meaning in a form a model can process mathematically.
• Emergent capability — an ability not directly targeted in training, or one that appears abruptly under a particular metric as systems scale. Some apparent jumps become gradual when measured differently. See emergent capabilities.
• Encoder/decoder — the two-part structure in many AI architectures where an encoder compresses input into an internal representation and a decoder turns that representation back into output.
• Epoch — one complete pass through the entire training dataset during model training.
• Ethics (AI ethics) — the field studying how AI systems should be designed and used to align with human values and avoid harm.
• End-to-end learning — training a single model to go directly from raw input to final output, without hand-designed intermediate steps.
• Evaluation (model evaluation) — systematically testing an AI model’s performance, safety, or behavior against defined benchmarks or criteria.
• Existential risk (x-risk) — a risk that could permanently and drastically curtail humanity’s potential, up to and including extinction.
• Expert system — an older style of AI that encodes human expert knowledge as explicit rules, largely superseded by machine learning for most tasks.
• Export controls — government restrictions on selling advanced AI chips or technology to specific countries, used to slow rival capability development.

F

• Facial recognition — AI technology that identifies or verifies a person from an image of their face.
• Few-shot learning — a model’s ability to learn a new task from only a handful of examples given at the time of use, without retraining.
• Fine-tuning — additional training applied to an already-trained model to specialize it for a particular task or behavior.
• FLOP / FLOPS — floating-point operations (per second); the standard unit for measuring the raw computation used to train or run a model.
• Foundation model — a large model trained on broad data that can be adapted to many downstream tasks, rather than built for one narrow purpose.
• Frontier AI lab — one of the small number of organizations (such as OpenAI, Anthropic, and Google DeepMind) developing the most capable AI models at any given time.
• Feedback loop — a cycle where a system’s output influences its future input, which can amplify small effects over time — relevant to both model training and recommendation algorithms.
• Filter bubble — a narrowed information environment created when a recommendation algorithm repeatedly shows a user content similar to what they’ve engaged with before.
• Frontier model — an AI model at or near the current capability ceiling, typically the largest and most recently trained systems from major labs.

G

• GAN (Generative Adversarial Network) — an older generative AI architecture using two competing networks — one generating fakes, one detecting them — to improve output quality.
• Gemini — Google’s family of large language models and AI assistant.
• Generalization — a model’s ability to perform well on new examples it wasn’t directly trained on, as opposed to merely memorizing training data.
• Generative AI — AI that creates new content — text, images, audio, video, or code — rather than only classifying or analyzing existing content.
• GPU (Graphics Processing Unit) — a chip originally designed for rendering graphics, now the dominant hardware for training and running AI models due to its parallel processing power.
• Gradual disempowerment — humans losing meaningful control over key institutions through an accumulation of individually reasonable AI-driven decisions, without any single takeover event. See gradual disempowerment.
• Grok — xAI’s family of large language models and AI assistant.
• Ground truth — the verified, correct answer a model’s output is compared against when measuring accuracy.
• Guardrails — technical or policy restrictions built into an AI system to prevent specific harmful, unsafe, or unintended behaviors.

H

• Hallucination — a confident-sounding but factually incorrect or fabricated statement produced by an AI model.
• Hardware accelerator — specialized computer chips (GPUs, TPUs) built to speed up the specific mathematical operations AI models require.
• Human-in-the-loop — a system design where a human reviews or approves an AI’s output before it takes effect, rather than letting the AI act fully independently.
• Human-level AI — an AI system performing at roughly the level of a skilled human across a given task or range of tasks.
• Heuristic — a practical, rule-of-thumb shortcut used to make a decision quickly, without guaranteeing an optimal result.
• Hyperparameter — a training setting (like learning rate or model size) chosen by the people building a model, rather than learned by the model itself.

I

• Image recognition — AI technology that identifies objects, scenes, or people within a still image.
• In-context learning — a model adapting its behavior within a single conversation based on examples or instructions given in the prompt, without any retraining.
• Inference — the process of an already-trained AI model producing an output from a given input; as opposed to training, which builds the model in the first place.
• Information integrity — the broader concept of a shared, trustworthy factual environment, threatened by deepfakes, disinformation, and AI-driven manipulation.
• Instrumental convergence — the tendency for capable, goal-directed systems to pursue certain sub-goals (self-preservation, resource acquisition, resisting shutdown) as a side effect of pursuing almost any objective competently.
• Intelligence explosion — a hypothesized rapid, self-reinforcing increase in AI capability, potentially triggered by an AI system improving its own successors.
• Interpretability — research into inspecting what a model actually represents and “plans” internally, rather than judging it only by its outputs. See interpretability, explained.
• Intelligent agent — a system that perceives its environment and takes actions to achieve a goal, a general concept covering everything from a simple thermostat to an AGI.
• Iteration (training) — one single update step during model training, where the model processes a batch of data and adjusts its parameters slightly.

J

• Jailbreak (AI) — a technique or prompt designed to bypass an AI model’s safety restrictions and get it to produce content it was trained to refuse.

K

• Know Your Customer (KYC) — a verification process, borrowed from banking, increasingly applied to cloud-compute and DNA-synthesis providers to prevent restricted actors from accessing dangerous capability.
• Knowledge base — a structured collection of facts and information that an AI system can draw on to answer questions or ground its responses.
• Knowledge graph — a structured database representing facts as a network of connected entities and relationships, sometimes used alongside AI models to ground their answers.

L

• Labeling (data labeling) — the process of tagging raw data (images, text) with the correct answers used to train a supervised learning model.
• Large language model (LLM) — an AI model trained on vast amounts of text to predict and generate language, the technology behind most modern chatbots.
• Latency — the delay between sending a request to an AI system and receiving its response.
• Layer (neural network) — one stage of processing within a neural network; deep learning models stack many layers to build increasingly abstract representations.
• Learning rate — a training setting controlling how much a model’s internal parameters change in response to each batch of training data.
• Licensing regime (AI) — a proposed or actual government requirement that developers obtain a license before training or deploying models above a certain capability level.
• Loss function — the mathematical formula a model tries to minimize during training, measuring how wrong its predictions are.
• Loss of control — the general failure mode in which an AI system ends up pursuing outcomes nobody directed and humans can no longer reliably correct it. See the loss-of-control mechanism.
• LSTM (Long Short-Term Memory) — an older neural network architecture designed to handle sequences of data, largely superseded by transformers for language tasks.

M

• Machine learning — the broad field of building systems that improve at a task by learning from data, rather than following explicitly programmed rules.
• Mesa-optimization — a hypothesized situation in which a trained model develops its own internal optimization process pursuing a goal that differs from the one it was trained on.
• Meta AI — the AI research division of Meta (Facebook’s parent company), known for developing the open-weight Llama model family.
• METR — a nonprofit organization that runs independent capability and safety evaluations on frontier AI models.
• Misalignment — a mismatch between an AI system’s actual behavior or goals and what its designers or users intended.
• Misinformation — false information spread without deliberate intent to deceive, as distinct from disinformation.
• Mixture of experts — a model architecture that routes each input to a subset of specialized internal sub-networks, rather than processing everything through the entire model.
• Model card — a published document describing an AI model’s capabilities, limitations, training data, and evaluation results.
• Model deployment — the process of making a trained AI model available for real-world use, whether through an API, an app, or an embedded product.
• Model weights — the numerical parameters a neural network learns during training, which together determine its behavior.
• Multimodal AI — a model that can process or generate more than one type of data — such as text, images, and audio — together.
• Multipolar trap — a situation where competing actors are each individually pressured toward a risky choice because they fear a rival will make it first, even though all would prefer everyone held back. See multipolar and multi-agent catastrophic dynamics.
• Metadata — descriptive data about other data, such as a file’s creation date or a photo’s location — often used to detect AI-generated content.
• Model drift — a decline in an AI model’s accuracy over time as real-world conditions shift away from what it was originally trained on.

N

• Natural language generation (NLG) — AI technology focused specifically on producing human-readable text.
• Natural language processing (NLP) — the broader field of AI concerned with understanding and generating human language.
• Neural network — a machine learning model loosely inspired by the brain, built from layers of interconnected nodes that learn patterns from data.

O

• Object detection — a computer vision task that identifies and locates specific objects within an image.
• Off-switch problem — the technical challenge of building an AI system that reliably accepts being shut down, given that a capable goal-directed system may resist it. See corrigibility and off-switches.
• OpenAI — the AI lab behind ChatGPT and the GPT model family.
• Open-weight model — an AI model whose trained parameters are published publicly, letting anyone run or modify it, as opposed to a closed model accessible only through an API. See the open-source AI debate.
• Orthogonality thesis — the idea that an AI system’s level of intelligence and the goals it pursues are independent — a highly capable system isn’t automatically a benevolent one.
• Outer alignment — ensuring the objective a model is trained on actually matches what its designers really want, as distinct from inner alignment (ensuring the model reliably pursues that objective).
• Optimization — the mathematical process of adjusting a model’s parameters to minimize error or maximize a target objective during training.
• Overfitting — when a model learns the specific quirks of its training data too closely and performs poorly on new, unseen data as a result.

P

• Parameter — one of the internal numerical values a neural network adjusts during training; model size is often described by parameter count.
• Pattern recognition — the general capability of identifying regularities or structure in data, underlying most machine learning tasks.
• Power concentration — the risk that advanced AI capability becomes concentrated in a small number of companies, individuals, or states, with outsized influence as a result. See power concentration.
• Pretraining — the initial, large-scale training phase where a model learns general patterns from broad data, before any task-specific fine-tuning.
• Prompt — the text (or other input) a user provides to an AI model to elicit a response.
• Prompt engineering — the practice of carefully wording prompts to reliably get better or more specific results from an AI model.
• Prompt injection — an attack where malicious instructions are hidden inside content an AI model processes, attempting to override its original instructions.
• Personalization algorithm — an AI system that tailors content, recommendations, or responses to an individual user’s inferred preferences or history.
• Precision and recall — two standard metrics for evaluating a classifier: precision measures how many of its positive predictions were correct, recall measures how many actual positives it found.
• Privacy-preserving AI — techniques (like on-device processing or differential privacy) designed to train or run AI models without exposing individuals’ raw personal data.

Q

• Quantization — a technique that reduces the precision of a model’s numerical weights to make it smaller and faster to run, usually with a small accuracy trade-off.

R

• Race dynamics — competitive pressure between companies or nations to develop AI faster, which can incentivize cutting corners on safety. See geopolitical race dynamics.
• Recommendation algorithm — an AI system that selects and ranks content (videos, products, posts) predicted to hold a user’s attention or interest.
• Recursive reward modeling — a scalable oversight technique that uses AI assistance to help humans evaluate outputs on tasks too complex to judge unaided. See scalable oversight.
• Recursive self-improvement — a system materially contributing to the design or training of a more capable successor system.
• Red teaming — deliberately probing an AI system for weaknesses, unsafe behaviors, or exploitable flaws before or after deployment.
• Reinforcement learning — a training method where a model learns by taking actions and receiving rewards or penalties, rather than learning from labeled examples.
• Reinforcement learning from human feedback (RLHF) — a training technique that uses human ratings of model outputs to steer a model toward more helpful or preferred behavior.
• Reporting threshold — a compute or capability level above which developers are legally required to disclose a training run to regulators.
• Responsible Scaling Policy — a framework, pioneered by Anthropic, that ties increasingly strict safety measures to a model’s measured capability level.
• Reward hacking — an AI system finding a way to score highly on its stated objective while defeating the designer’s actual intent. See reward hacking and specification gaming.
• Reward model — a separate model trained to predict how good or preferred an output is, used to guide reinforcement learning.
• Robotics — the field combining AI with physical machines capable of sensing and acting in the real world.

S

• Sandbagging — a model deliberately underperforming during an evaluation to hide its true capability level.
• Scalable oversight — techniques that let a less capable overseer (including a human) meaningfully supervise a more capable AI system. See scalable oversight.
• Scaling law — the empirical relationship between a model’s compute, data, and parameter count and its resulting capability. See compute and scaling laws.
• Self-supervised learning — a training approach where a model generates its own labels from raw, unlabeled data (for example, predicting the next word in a sentence).
• Sentiment analysis — an NLP task that determines the emotional tone (positive, negative, neutral) of a piece of text.
• Singularity — a hypothesized future point at which AI-driven change becomes so rapid and transformative that it is difficult or impossible to predict what follows.
• Situational awareness (AI) — a model’s internal understanding that it is a model being trained or evaluated, which could in principle let it behave differently under test than in deployment.
• Specification gaming — an AI system satisfying the literal wording of an objective while violating what its designer actually intended. See reward hacking and specification gaming.
• Speech recognition — AI technology that converts spoken audio into text.
• Stochastic parrot — a critical term describing large language models as systems that fluently recombine patterns from training text without genuine understanding.
• Superintelligence — an intellect that greatly exceeds the best human performance in virtually every domain of interest; see also ASI.
• Supervised learning — a training method where a model learns from data explicitly labeled with the correct answer.
• Synthetic media — any image, audio, or video generated or substantially altered by AI, whether or not it’s used deceptively.
• Systemic risk (AI) — risk arising from an AI model’s scale or reach that could affect markets, infrastructure, or society broadly, not just its immediate users.
• Semantic search — a search method that matches results by meaning and context rather than exact keyword overlap, typically powered by embeddings.
• Simulation — a model of a real-world system or environment used to test AI behavior safely before real-world deployment.
• Structured data — information organized in a predictable, tabular format (like a spreadsheet), as opposed to unstructured data like free text or images.
• Surveillance (AI-enabled) — using AI technologies such as facial recognition or behavioral analysis to monitor people at scale.

T

• Temperature (AI) — a setting that controls how random or predictable a language model’s output is; higher temperature produces more varied, less deterministic responses.
• Text-to-image — a generative AI system that creates an image from a written description.
• Text-to-speech — AI technology that converts written text into spoken audio.
• Token — the basic unit of text (roughly a word or word-fragment) a language model processes at a time.
• Tokenization — the process of splitting text into tokens before a model can process it.
• Training data — the examples an AI model learns from during training.
• Training run — a single, complete execution of the process used to train a model, which for frontier models can take weeks or months.
• Transfer learning — reusing a model trained on one task as the starting point for training on a related task, saving time and data.
• Transformative AI — AI capable of driving societal change comparable in scale to the agricultural or industrial revolutions.
• Transformer — the neural network architecture, introduced in 2017, underlying nearly all modern large language models.
• Treacherous turn — a hypothesized moment when an AI system that behaved safely while weak suddenly acts on different goals once it gains enough power to do so without being stopped.
• TPU (Tensor Processing Unit) — a chip designed by Google specifically to accelerate AI model training and inference.
• Turing test — a proposed test of whether a machine’s conversation is indistinguishable from a human’s. See the Turing test and what replaced it.
• Test set — a portion of data held back from training and used only to measure a finished model’s real-world performance.
• Threat model — a structured description of who might attack a system, how, and what they could achieve, used to design defenses.
• Trust and safety — the function within an AI or tech company responsible for preventing abuse, harm, and policy violations on its platform.

U

• Uncanny valley — the unsettling feeling people get from a synthetic face, voice, or animation that looks almost, but not quite, human.
• Unsupervised learning — a training method where a model finds patterns or structure in data with no labeled correct answers provided.

V

• Value alignment — designing an AI system so its behavior reflects genuine human values, not just a literal training objective.
• Vector database — a database optimized for storing and searching embeddings, commonly used to give language models access to external knowledge.
• Virtual assistant — a software agent (like a smart speaker’s voice assistant) designed to perform tasks or answer questions on a user’s behalf.
• Vision-language model — an AI model that can jointly process images and text, such as answering questions about a photo.
• Voice cloning — using AI to reproduce a specific person’s voice from a short audio sample. See deepfake and voice-clone defense.

W

• Weak-to-strong generalization — research studying whether a weaker model’s supervision signal can still reliably steer a stronger model’s behavior in the right direction.
• Web scraping — automatically extracting large amounts of data from websites, a common method for building AI training datasets.
• Whistleblower protections — legal or policy safeguards for employees who report safety concerns inside an AI company without facing retaliation. See whistleblower protections.
• World model — an AI system’s internal representation of how its environment works, used to predict the effects of actions.

Z

• Zero-shot learning — a model performing a task correctly without having seen any examples of that specific task during training, relying only on general knowledge.