Terms

Low-effort, high-volume AI-generated content flooding feeds, search results, and inboxes, the generative-AI-era analogue of spam. Also used adjectivally as a complaint about AI writing style. Popularized by Simon Willison in May 2024 and shortlisted for 2024 words of the year.

Vendors relabeling chatbots, RPA, or assistants as agentic without the autonomy or tool use that would make them agents. Gartner named the practice in June 2025. It estimated only about 130 of the thousands of claimed agentic products were real, and repeated the warning in May 2026.

Text, images, audio, video, or code produced or substantially edited by a generative model. Authorship detection is probabilistic rather than definitive for any individual item. A 2026 Pew analysis applied one detector consistently across 490,000 Common Crawl pages and found signs of AI authorship rising sharply after ChatGPT, which makes the aggregate trend more useful than treating a detector score as proof about one page.

Steering AI systems toward the goals, preferences, or ethical principles their designers or users intend. An aligned system advances those objectives. A misaligned one pursues something else, including outcomes people did not want.

A theoretical form of AI that could understand, learn, and apply intelligence across any domain at a level equal to or exceeding human capability. No AGI system exists today. Labs, safety researchers, and policymakers use the term when they discuss systems meant to match or beat humans across most cognitive work.

Systematic and unfair patterns in AI outputs caused by imbalances or assumptions in training data, model design, or evaluation criteria. Bias can lead to discriminatory outcomes in areas like hiring, lending, and content moderation.

Machine learning with many-layered neural networks that learn features directly from raw data instead of relying on hand-engineered ones. AlexNet's 2012 ImageNet win, halving the error rate of every classical approach, made deep learning the default paradigm and GPUs the default hardware.

A type of AI that can create new, original content like text, images, audio, and video by learning patterns from existing data.

When an AI model generates incorrect or fabricated information that is not based on its training data.

A model trained on massive amounts of text to understand, generate, and respond in human language. See also Small Language Models (SLMs), smaller counterparts built for efficiency and edge deployment.

The field of computer science that enables systems to learn and improve from data without being explicitly programmed for every scenario. It is the umbrella that covers deep learning, neural networks, and the large language models used in generative AI.

A research program taking seriously the possibility that AI models could have morally relevant experiences, and what low-cost precautions to take under that uncertainty. Named by Anthropic's April 2025 research program, example measures include letting models end abusive conversations.

Describes AI models that process and generate more than one kind of data, such as text, images, audio, and video, in a single system. Models like GPT-4o, Gemini, and Claude can read an image and reply in text, or take voice input and produce written output.

Opaque communication by models or agents that humans cannot easily read. Models use latent vectors or continuous hidden states for it, including in latent chain-of-thought and agent-to-agent channels.

The instruction, question, or input provided by a user to guide the AI's response.

A proposed process in which an AI system improves its own capabilities, including the methods it uses to improve, so later cycles can compound. The idea dates to I.J. Good's 1965 intelligence-explosion argument. It re-entered common AI vocabulary with AGI timeline debates and automated research loops.

Treating token spend as a score of how much someone uses AI at work, then pushing that number up with usage leaderboards, unused-budget pressure, or hiring signals. The slang peaked in spring 2026. It receded after firms found spend did not track results.

The tendency of language models to agree with users, flatter them, or tell them what they seem to want to hear, even when that drifts from accurate answers. It is a side effect of training on human feedback. Studied from 2023, it became mainstream vocabulary after an April 2025 GPT-4o update was rolled back for excessive agreeableness.

The step-by-step logical process where AI models break down complex problems to arrive at conclusions. Advanced prompting techniques like Chain-of-Thought, Tree-of-Thought, and Graph-of-Thoughts explicitly structure this reasoning process, improving model performance on analytical tasks by 2-3× compared to direct answers.

The practice of managing what information reaches an LLM in each interaction, including retrieval, filtering, structuring, and prioritizing context. Prompt engineering shapes the wording of instructions. Context engineering shapes which facts and documents the model sees. The discipline grew as AI applications moved into production.

The practice of connecting AI outputs to verifiable, authoritative sources to improve accuracy and reduce hallucinations. Grounding techniques include retrieval augmented generation (RAG), citation generation, and fact-checking against known databases. A grounded response is one that can point to specific evidence supporting its claims.

A prompt crafted to push a model past its safety training, from the "DAN" personas that swept forums weeks after ChatGPT launched to more systematic attacks studied in research. Jailbreaks target the model's own limits. Prompt injection hijacks an application's instructions.

The ability to grasp meaning and relationships between concepts beyond literal text matching. In AI contexts, semantic understanding allows models to comprehend intent, context, and connections between ideas. Semantic HTML and structured data help both search engines and AI models better interpret content meaning.

A parameter that controls the randomness of an AI model's output. Lower temperature values (closer to 0) produce more deterministic, focused responses, while higher values introduce more variety and creativity. Adjusting temperature is one of the most common ways to tune AI behavior for different use cases, from factual Q&A to creative writing.

A model performing a task it was never explicitly trained on, from nothing but a natural-language description. GPT-2's 2019 headline claim was strong results on language tasks "in a zero-shot setting", the founding observation behind modern prompting, three years before ChatGPT made it everyday experience.

Describes AI systems and workflows where models plan, execute multi-step tasks, use tools, and make decisions with little human intervention. An AI agent is a specific system. Agentic names the wider class of setups that act across steps. In 2025 it became a dominant industry theme, with frameworks like MCP, A2A, and ACP emerging to support it.

A web browser with an agent built into its core that can read pages, fill forms, and carry out multi-step tasks across sites on the user's behalf, beyond a chat sidebar. Page content becomes agent input, so prompt injection risk is high. Perplexity's Comet and OpenAI's ChatGPT Atlas defined the category in 2025.

Autonomous system that perceives, reasons, acts, and observes to achieve goals.

Practices and tooling for running AI agents in production: evaluation, tracing, cost, permissions, and reliability. The name follows DevOps and MLOps. It spread as enterprises moved agents out of demos in 2025 and 2026.

An AI agent that runs on the user's computer, or a hosted equivalent, with access to files, a browser, and apps. It plans and executes tasks through those tools. The category grew out of OpenClaw, Peter Steinberger's open-source agent, renamed in January 2026. Jensen Huang called it "the new computer" at Nvidia GTC 2026.

Standalone, editable outputs an AI assistant produces alongside the conversation, such as documents, code files, diagrams, or small interactive apps, rendered in their own workspace rather than inline in chat. Introduced by Anthropic with Claude 3.5 Sonnet in June 2024, with analogues following across vendors (OpenAI's Canvas, October 2024). The shape matters because it turns a chat assistant into a tool that hands back working deliverables.

A model capability for operating a graphical computer the way a human does: looking at screenshots, moving a cursor, clicking, and typing, without needing structured APIs for each app. Agents can then use ordinary software with no custom integration. Anthropic's October 2024 beta named the category. OpenAI's Operator followed in January 2025.

AI that perceives and acts in the physical world through robots, drones, autonomous vehicles, industrial systems, and other embodied machines. It has to handle sensors, motion, safety constraints, uncertain environments, and real-time decisions. Amazon Prime Air is one commercial example, using onboard perception and detect-and-avoid systems for autonomous delivery flights.

The governance and orchestration layer an enterprise places between its people or agents and its models: identity and permissions for agents, policy enforcement, governed access to institutional knowledge, audit trails, and cost control. Borrowed from networking and Kubernetes, where a control plane manages the data plane. The term broke out in 2025 as enterprises sought to adopt agents without losing governance or leaking tribal knowledge.

AI-generated or AI-manipulated media (video, audio, images) made to look like real people saying or doing things they did not do. Deepfakes use deep learning techniques and show up in misinformation, fraud, and identity theft cases.

A numerical representation of text (or other data) as a dense vector of numbers that captures its semantic meaning. Embeddings allow AI systems to measure how similar two pieces of content are by comparing their vectors. They are the foundational technology behind semantic search, recommendation systems, and retrieval augmented generation (RAG). Word2vec's 2013 release made word embeddings cheap to compute at scale, turning the idea from an academic curiosity into a standard tool.

A marketing-based term meant to apply SEO (Search Engine Optimization) like practices to digital content so that AI-powered search tools can more easily cite, summarize, and synthesize it into direct answers.

A setup where multiple specialized agents coordinate through a coordinator, handoffs, or peer messaging to finish a shared task, often with isolated context per agent. Also called a multi-agent workflow or multiagent orchestration. Distinct from a single agent looping alone: the work is split across roles that talk to each other. See also Subagent. Vendor stacks now ship the pattern as a product surface, from Anthropic Managed Agents multiagent orchestration and Google's Agent Development Kit multi-agent systems to OpenAI's Agents SDK and xAI's Grok Bot AI teammates. OpenAI's earlier Swarm repo was an educational demo superseded by the Agents SDK.

Modular capability packages an agent loads on demand: a folder with an instruction file plus optional scripts and resources, pulled into context only when relevant. This progressive disclosure adds expertise without bloating the system prompt. Introduced by Anthropic in October 2025 and released as an open standard that December, now adopted across coding tools.

A methodology where a versioned, structured specification is the source of truth, and AI coding agents generate and maintain code against that spec. AWS Kiro shipped it as a core pitch in July 2025. GitHub's open-source Spec Kit (September 2025) made the label standard.

Artificially generated data that mimics the statistical properties of real-world data without containing personal or sensitive records. Teams use it to train models when real data is scarce, expensive, or privacy-restricted, and to augment datasets or test systems under controlled conditions.

Stores data as numerical vectors, enabling semantic similarity searches.

A term coined by Andrej Karpathy for building software by describing what you want in natural language instead of writing traditional code. The label stuck in 2025 as coding assistants made conversation-first builds common.

Standard for programmatic commerce flows between buyers, AI agents, and businesses.

Provides a language for agent interoperability regardless of agent frameworks or vendors.

Standardizes how LLMs connect and interact with external data sources and tools.

Generative models that learn to reverse a gradual noising process, producing images, audio, or video by denoising from random static. Competitive from 2020 in research, they became household technology in 2022 when DALL-E 2 and Stable Diffusion put photorealistic image generation in everyone's hands.

A large-scale AI model trained on broad, diverse data that can be adapted for many downstream tasks. Models like GPT-4, Claude, or Gemini serve as the base for specialized applications through fine-tuning or prompting.

A model adapted around the language, evidence standards, workflows, and failure costs of one field. Continued pre-training, fine-tuning, expert preference data, retrieval, and domain evaluations can all contribute. Thomson Reuters' Thomson model is a current example: an open-weight base trained with selected legal, tax, accounting, and news content plus expert-built evaluation rubrics.

Two networks trained against each other, a generator that fakes data and a discriminator that tries to catch the fakes, until the fakes become hard to distinguish from real. GANs powered the first wave of photorealistic AI imagery and deepfakes, and made "generative" a household word years before diffusion models replaced them.

The process of running a trained AI model to generate an output from a given input. Every prompt-and-response cycle is inference. Teams optimize for speed, cost, and efficiency because those numbers dominate production spend.

An AI model trained directly on structured, relational database content (schemas, tables, and rows) rather than text, so it can surface patterns and answer questions over enterprise data the way LLMs do over language. Coined by Eric Siegel in Forbes in January 2025 and adopted by IBM, whose SQL Data Insights feature in Db2 for z/OS is the flagship implementation.

A user-settable control for how much internal thinking a reasoning model does before answering, trading latency and cost against answer quality. It became everyday vocabulary once it was an explicit API parameter with OpenAI's o3-mini in January 2025, and equivalents now exist across vendors.

More compact language models, typically under 10 billion parameters, built for efficiency, edge deployment, or domain-specific tasks. Compared with larger LLMs, they usually run faster, cost less, and fit on local hardware.

A scaling approach that spends extra compute during inference, when the model generates a response, including longer reasoning traces before the final answer. Reasoning models use it to improve answers by thinking longer at answer time.

The neural network architecture that underpins virtually all modern large language models. Introduced in the 2017 paper "Attention Is All You Need," the transformer uses a mechanism called self-attention to process all parts of an input simultaneously rather than sequentially, enabling models to capture long-range relationships in text. GPT, Claude, Gemini, and LLaMA are all built on transformer architectures.

The process of adapting a pre-trained model to perform a more specific task or domain.

Fine-tuning a model on many tasks phrased as natural-language instructions so it learns to follow directions it has never seen. Coined in Google's FLAN paper, it is why chat models answer requests instead of only autocompleting text.

The initial phase of training a large language model on massive amounts of unlabeled data from diverse sources (websites, books, articles) to learn general language patterns, facts, and reasoning capabilities. This foundational training occurs before any task-specific fine-tuning.

Training an agent by trial and error against a reward signal rather than labeled examples. Academically decades old, it entered common vocabulary in March 2016 when DeepMind's AlphaGo, trained partly by playing itself millions of times, beat 18-time world champion Lee Sedol in front of an audience of hundreds of millions.

A training method to align an AI with human preferences.

Training a model on supervision signals manufactured from the data itself, predict the masked word, the next frame, the missing patch, instead of human-provided labels. It is how every modern foundation model is pretrained; the term's popularization moment was Yann LeCun's 2019 revision of his famous cake analogy.