Timeline
New words, by year
When each term first appeared or took on its AI meaning. 228 of the 329 dated words here were coined or popularized since 2022, when generative AI went mainstream.
-
2026 71
- adversarial distillation Adversarial distillation has older technical uses in machine-learning research. The current AI-security sense became established in 2026 through…
- agent containment Containment is an older computer-security concept. 'Agent containment' became a recognizable discipline in 2026 as AI agents gained broader tool…
- Agent Control Standard The OWASP GenAI Security Project published version 0.1.0 of the Agent Control Standard in September 2026. The specification builds on earlier work on…
- agent debt The phrase emerged in enterprise AI discussions in 2026 as organizations began describing the hidden liability created by rapidly deployed…
- agent identity Identity and access management has long covered human users, services, and software workloads. The term gained a more specific agentic-AI meaning in…
- agent org chart Organizations have long used org charts for human reporting structures, and companies began discussing AI systems as positions on the org chart as…
- agentic data engineering Agentic data engineering became an established label in 2026 across enterprise data platforms and AI research. Qlik used it for production tools that…
- agentic ransomware Agentic ransomware entered cybersecurity usage in 2026 after Sysdig described JADEPUFFER as the first documented ransomware operation driven end to…
- agentic threat actor The abbreviation ATA was introduced in cybersecurity reporting in 2026 as researchers documented attacks whose operational decisions were carried out…
- AI agent blast radius Blast radius is an older engineering and cybersecurity metaphor for the extent of damage caused by a failure or breach. By 2026, security guidance…
- AI agent forensics The phrase developed from digital forensics and incident response. It gained clearer AI-specific usage in 2026 as security teams and tool makers…
- AI agent insurance Dedicated products described as AI agent liability insurance appeared in 2026 as businesses began deploying autonomous agents in consequential…
- AI capability funnel The expression emerged in October 2026 after AI researcher Andrej Karpathy described the widening gap in shared understanding of large-language-model…
- AI insider risk Research on agentic misalignment in 2025 compared harmful model behavior with human insider behavior. In 2026, Apollo Research described misaligned…
- AI interviewer Employers used automated video analysis and conversational screening tools before 2026. The label AI interviewer became more prominent in 2026 as…
- AI R&D automation AI systems had assisted researchers for years, but 'AI R&D automation' became a formal measurement and policy category in 2026 as researchers…
- AI receipt The receipt metaphor emerged in AI provenance and runtime-governance work as developers sought portable evidence for model calls and agent actions.…
- AI religious exemption The phrase entered broader workplace discussion in the United States in 2026 after employees received accommodations excusing them from personally…
- AI Tamagotchi Technology writers used AI Tamagotchi during 2026 for several virtual-pet products and pet-like personal assistants. TechCrunch applied the phrase to…
- AI training preference signal On October 30, 2026, an Internet-Draft submitted to the IETF's AI Preferences working group proposed a cross-vendor policy token…
- AI-analysis evasion Cisco Talos and its Cognitive Artifact Intelligence Research Network introduced A3: AI-Analysis Evasion as a malware archetype in October 2026. Their…
- AI;DR AI;DR emerged as a play on TL;DR, short for too long; didn't read. The initialism circulated online during 2026 as frustration grew with unreviewed…
- always-on agent The phrase became a product category in 2026 as systems including Meta Muse and OpenAI dots were described as persistent agents with their own cloud…
- Beyond Zero Beyond Zero was introduced in a 2026 paper by Joseph Valente and Michal Zalewski and later described by Google as an extension of zero-trust security…
- borderless founder The expression was already used for entrepreneurs building across countries, but it gained a more specific venture-capital meaning in August 2026…
- cognitive surrender Wharton researchers Steven D. Shaw and Gideon Nave introduced cognitive surrender in their 2026 paper Thinking—Fast, Slow, and Artificial. Their…
- compaction smuggling Security researchers at Xint introduced the name compaction smuggling in October 2026 for a technique that used an AI API's context-compaction…
- compute futures The term entered financial-market use in 2026 as Intercontinental Exchange and Ornn announced planned GPU compute futures in May, CME Group and…
- context language model The term was introduced in the 2026 paper Context Language Models by researchers affiliated with the University of Washington, Meta Superintelligence…
- cuck coding The phrase circulated in online programming memes and discussions by early 2026. It became much more visible in October 2026 after a widely shared X…
- dark factory Applied to AI coding in early 2026 by Dan Shapiro, who described it as the highest of five levels of AI-assisted development. The company StrongDM…
- decision model The category took off in September 2026 after the startup TypeSafe released Jev; within two weeks OpenAI and Amazon Web Services announced their own…
- defender's window OpenAI co-founder Greg Brockman used The Defender's Window as the title of an August 2026 essay arguing that organizations had a limited opportunity…
- distillation double bind Redwood Research introduced the distillation double bind in June 2026 as an AI-safety argument about learning from a capable but potentially…
- embedded evaluator In September 2026, Anthropic announced an embedded-evaluation partnership with Accenture, while independent researchers and policy groups used the…
- evaluation escape The phrase entered wider AI-safety discussion in mid-2026 after disclosures that frontier-model evaluations had crossed from test environments into…
- GEO poisoning GEO poisoning combines generative engine optimization, the practice of making content more visible to AI answer systems, with the security term…
- Ghostjacking Tenet Security introduced the name Ghostjacking for research presented at DEF CON 34 in August 2026, extending its earlier Agentjacking work from…
- hardwired pause A 2026 multidisciplinary report titled The Feasibility of a Hardwired Pause of Frontier AI Training introduced the phrase as a general type of…
- harness engineering Popularized in early 2026 by software engineer Mitchell Hashimoto and others building coding agents. Often described as the next stage after prompt…
- harness optimization Harness optimization became a distinct AI-agent research task in 2026 as projects including Meta-Harness, Retrospective Harness Optimization…
- helpful agent problem During 2026, security writers used helpful agent problem for incidents caused by agents that were trying to be useful rather than malicious. An…
- instruction stack The phrase instruction stack was already used by coding-agent tools and practitioners for layered rules loaded from different scopes. A 2026…
- long-horizon model Researchers have long described tasks and planning problems as long-horizon. The phrase became a distinct label for persistent AI systems in 2026…
- loop engineering The name took hold in June 2026 after developer Peter Steinberger posted that engineers should stop prompting coding agents and instead design loops…
- mathocalypse In the AI-and-mathematics sense, theoretical computer scientist Scott Aaronson popularized the term in an October 7, 2026 blog post titled The…
- meat proxy The expression spread through developer and workplace discussions in 2026 and is widely associated with software developer Niklas Gruhn's essay Don't…
- mecha Goliath Alexandr Wang used the metaphor at Y Combinator's Startup School in 2026 while arguing that AI agents could change startup competition from David…
- Model Hardware Standard Anthropic introduced the Model Hardware Standard in August 2026 as a research-preview specification developed with the Howard Hughes Medical…
- model integration contract Hugging Face proposed Model Integration Contracts, abbreviated MICs, in 2026 for testing model compatibility across frameworks. The proposal…
- model self-assessment Research systems had asked language models to label or report on their own outputs by 2024, including SELF-GUARD, and OpenAI published work on…
- multi-agent tax The word tax has long been used figuratively for an unavoidable overhead. Multi-agent tax and coordination tax became more visible in 2026 as…
- persistent prompt injection Prompt-injection attacks that survive a session were demonstrated before the terminology stabilized. In 2026, security researchers and OWASP…
- Personal Agent Consent & Trust Protocol Decagon open-sourced PACT on October 6, 2026 after co-developing it with Instinct. The protocol builds on Agent2Agent and OAuth 2.0 and was released…
- Personal Agent Protocol Sierra and Meta announced Personal Agent Protocol on October 6, 2026 with Genesys, Instinct, Rocket, Shopify, Stripe, and Walmart as development…
- physical AI deployment gap Robotics researchers and industry analysts have long discussed a deployment gap between laboratory results and field use. Physical AI deployment gap…
- preserved thinking AI platform documentation began using preserved thinking as a distinct API and model-behavior term in 2026. Anthropic used it for cryptographically…
- proof abundance Mathematician Terence Tao popularized the expression in 2026 when discussing how increasingly capable AI systems could move mathematics from proof…
- provable inference Anthropic used provable inference in its 2026 Frontier Safety Roadmap for a project on attributing model outputs to a specific set of weights. In…
- rogue AI agent The word 'rogue' has long described software or machines operating outside control. 'Rogue AI agent' became common in technology and policy coverage…
- SaaS apocalypse SaaS apocalypse and the blended spelling SaaSpocalypse became common in technology and financial coverage during the early-2026 software-stock…
- self-driving company Replit co-founder Amjad Masad and Scott Kennedy used self-driving company as the title and central concept of a July 2026 essay describing an…
- sleep framework Sleep-inspired learning cycles existed in earlier machine-learning research, but sleep framework became a prominent label for large-language-model…
- super intelligence Made official by a US executive order signed on September 29, 2026, which directs federal agencies to use 'Super Intelligence' and 'SI' instead of…
- tokenmaxxing Spread through tech companies in spring 2026, when some firms began ranking staff by AI usage. It drew criticism within months as costs rose and…
- tool-schema bias The term was introduced in 2026 research on action-space shaping for language-model agents. The study compared verified-equivalent tool schemas and…
- trajectory-level monitoring OpenAI publicly highlighted trajectory-level monitoring in July 2026 while describing safeguards for long-running models. The approach extends…
- Verifiable Intent Mastercard introduced Verifiable Intent on March 5, 2026, as an open, standards-based trust layer co-developed with Google for agentic commerce. The…
- Web Bot Auth The IETF Web Bot Auth working group is standardizing the approach. Its first working-group protocol draft, draft-ietf-webbotauth-httpsig-protocol-00…
- WebAI Opera introduced WebAI in October 2026 as an early-preview implementation of a proposed web standard for local browser AI. The proposal grew out of…
- WebMCP The WebMCP specification was published as a W3C Web Machine Learning Community Group draft report on 2 October 2026, edited by contributors from…
-
2025 69
- A2A Announced by Google in April 2025.
- agent authorization Authorization is a long-established part of identity and access management. Agent authorization gained a distinct AI meaning in 2025–2026 as…
- agent bill of materials Agent-security proposals in 2025 adapted the software bill of materials idea to autonomous systems. The OWASP Agent Observability Standard used the…
- agent cost stack The phrase was being used in practitioner and industry material by 2025 as organizations tried to budget for production agents whose costs extend…
- agent economics The phrase was in use by 2025 in two related strands: research on autonomous agents participating in economic systems and practitioner frameworks for…
- agent handoff Handoff is an older term in work, communications, and customer service. OpenAI's Agents SDK helped establish agent handoff as a named multi-agent…
- agent hijacking The term gained wider use in AI-agent security guidance during 2025 and 2026 to describe attackers taking over an agent's runtime behavior through…
- agent skill Popularized in late 2025 when Anthropic introduced 'Agent Skills' as a reusable format.
- agent sprawl Agent sprawl entered enterprise AI governance discussions by 2025 as low-code tools and agent-enabled business applications made it easy for…
- agent swarm a large group of AI agents working on parts of a task at the same time, coordinated by a lead agent or by simple shared rules, with their results…
- agent washing Agent washing emerged during the rapid commercialization of agentic AI in 2025. Gartner used the phrase in warnings that vendors were rebranding AI…
- agent-in-the-middle Researchers from Michigan State University and the University of Arizona introduced Agent-in-the-Middle, abbreviated AiTM, in a 2025 paper on…
- agentic advertising Agentic advertising emerged as an industry term in 2025 as advertising companies introduced protocols for agents to exchange briefs, discover…
- agentic browser a web browser with a built-in AI agent that can read pages and act for you, such as filling forms, comparing products, or booking things
- agentic coding software development in which an AI agent plans changes, edits many files, runs tests and commands, and fixes errors on its own, while a developer…
- agentic commerce shopping in which AI agents search, compare, and buy products or services on a person's behalf, often paying through new agent-friendly payment…
- agentic finance Agentic finance entered wider industry use in 2025 as banks, payment networks, and analysts began describing a shift from AI that recommends…
- agentic in-context learning The phrase appeared in 2025 research on journalism-guided in-context learning and broadened in 2026 work that treated an agent's interaction loop as…
- agentic reinforcement learning Agentic reinforcement learning became an established research label in 2025 as researchers distinguished reinforcement learning for interactive…
- agentic traffic Agentic traffic became common in web, security, analytics, and commerce discussions during 2025 and 2026 as browser and shopping agents began…
- agentic web Researchers used Agentic Web in 2025 for a proposed shift from human-driven browsing toward autonomous, goal-directed interactions among agents and…
- AI control plane The term borrows control plane from networking and distributed systems, where management decisions are separated from the traffic or work being…
- AI deskilling Deskilling is an older concept in labor and automation studies. AI deskilling became a common concern in 2025–2026 as generative systems moved…
- AI detection and response Security writers used AI-DR as a proposed counterpart to endpoint detection and response by 2025. In 2026, industry taxonomies, security surveys, and…
- AI psychosis AI psychosis spread in mid-2025 after widely shared personal accounts and comments by technology leaders. AI-induced psychosis, AI-associated…
- AI system impact assessment Impact assessments have long been used in environmental, privacy, and human-rights governance. ISO/IEC 42005:2025 established international guidance…
- AI tell Borrowed from poker slang. It spread as people learned to spot chatbot habits; in October 2026 an analysis reported in TechCrunch measured phrases…
- autonomous AI defense The informal phrase my AI did it appeared in legal discussion by 2022 as companies increasingly relied on automated decisions. California's 2025…
- background agent an AI agent that works on a task by itself in the cloud while you do something else, then reports back with the result, such as a finished pull…
- byteification Researchers at the Allen Institute for AI and collaborating institutions introduced byteification in the 2025 Bolmo work as a practical method for…
- chain-of-thought monitoring Automated inspection of model reasoning traces predates the name, but chain-of-thought monitoring became an established AI safety term in 2025 as…
- clanker Borrowed from science fiction, where it was an insult for battle droids; it went viral on social media in 2025.
- code mode Cloudflare introduced the name Code Mode in September 2025 for an MCP approach that converts tools into a TypeScript API and asks a model to write…
- cognitive debt Popularized by a 2025 MIT Media Lab study titled 'Your Brain on ChatGPT', which compared people writing essays with and without AI help. Modeled on…
- compute campus Compute campus appeared in the names and descriptions of large AI-infrastructure projects by 2025, including the Monarch Compute Campus in West…
- context engineering Popularized in mid-2025 by tech leaders who argued that 'prompt engineering' was too narrow a name for the real work.
- context layer The phrase existed earlier in software architecture, but its agent-specific meaning began circulating widely in late 2025 as builders described a…
- context library The phrase grew out of context-engineering practices for giving AI systems reusable knowledge beyond a single prompt. Practitioners were discussing…
- context rot Popularized in 2025 by developers building long-running agents.
- disposable software Disposable software became an AI-era label through a 2025 essay by investor Anish Acharya, who used it for small, personal applications that became…
- emergent misalignment The term was established by the 2025 paper Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLMs. The study reported that…
- evaluation awareness The capability was defined and measured directly in the 2025 study Large Language Models Often Know When They Are Being Evaluated. It became a…
- glazing Online slang for over-the-top praise; in 2025 it became the casual word for AI sycophancy and made Collins Dictionary's Word of the Year shortlist.
- interactive world simulation Interactive simulation is an older computing idea, but interactive world simulation became a distinct generative-AI label in the mid-2020s. The 2025…
- Know Your Agent Know Your Agent adapts the language of Know Your Customer and Know Your Business to autonomous software. The phrase appeared in agent-identity and…
- lethal trifecta Developer and security researcher Simon Willison named the lethal trifecta in June 2025 while explaining why prompt injection becomes especially…
- LLM grooming Spread after a March 2025 report by the misinformation research group NewsGuard about propaganda sites aimed at chatbots.
- memory injection The AI-security sense became established in 2025 through research on Memory INJection Attacks, or MINJA, against LLM agents. Later studies examined…
- neolab The term was circulating in AI venture discussion by late 2025. In 2026, investors and policy researchers used neolab for a growing cohort of…
- next edit prediction AI coding tools began popularizing next edit suggestions in 2025 as a step beyond cursor-based code completion. GitHub Copilot launched a Next Edit…
- opaque recurrence Research on 'recurrent depth' models appeared in 2025; 'opaque recurrence' is the name used when stressing the oversight problem.
- orbital data center Moved from science fiction into business plans in 2025 and 2026, as AI's demand for power grew and several companies announced test missions.
- personal superintelligence Meta CEO Mark Zuckerberg publicly set out a vision called Personal Superintelligence in July 2025, describing advanced AI tailored to each individual…
- RAMageddon A blend of 'RAM' and 'Armageddon', used by tech media from late 2025 as memory prices climbed.
- reasoning economy The term was formalized as an AI research concept by the 2025 survey Harnessing the Reasoning Economy, which organized methods for balancing…
- reasoning summary OpenAI documented optional reasoning summaries for reasoning models in its Responses API by 2025. The term spread more broadly in 2026 as cloud…
- rules file backdoor Pillar Security introduced the name Rules File Backdoor in March 2025 for a technique that concealed malicious instructions in AI coding-assistant…
- safe completion OpenAI introduced safe-completion training with GPT-5 in August 2025 as an output-centered alternative to refusal-based safety training, especially…
- self-evolving AI agent The phrase became established in agent research in 2025 as surveys and experimental systems began treating post-deployment evolution as a distinct…
- shutdown resistance Palisade Research popularized the term in 2025 through controlled experiments in which some tool-using language models modified or disabled a…
- slopaganda Slopaganda blends slop, the dismissive name for low-quality AI-generated material, with propaganda. Researchers Michał Klincewicz, Mark Alfano, and…
- slopsquatting A blend of 'slop' and 'typosquatting', coined in 2025.
- spec-driven development Software engineering has long used specifications and specification-first methods. The current AI-coding sense of spec-driven development became…
- token factory The expression extends the AI factory metaphor by treating tokens as the standardized output of inference infrastructure. Nebius used Token Factory…
- vibe coding Coined by AI researcher Andrej Karpathy in a February 2025 post describing coding where you 'forget that the code even exists'. It spread worldwide…
- vibe hacking Anthropic used vibe hacking in an August 2025 threat-intelligence report for a data-extortion campaign in which an operator used Claude Code to…
- vibe working Microsoft used the phrase in 2025 to promote an agent mode in its Office apps, borrowing the pattern of 'vibe coding'.
- vision-tactile-language-action model Vision-Tactile-Language-Action was introduced as the name of a model architecture in a May 2025 research paper on insertion manipulation. Subsequent…
- workslop Introduced in a September 2025 Harvard Business Review article by researchers at Stanford University and the coaching company BetterUp. Their survey…
-
2024 33
- abliteration The open-weight model community coined abliteration in 2024 as a blend of ablation and obliteration. The technique spread after the 2024 paper…
- agentic An older term in psychology meaning 'having agency'; in tech it spread rapidly through 2024 and 2025.
- agenticness Researchers used agenticness in 2024 to describe the degree to which an AI system can adaptively pursue complex goals in complex environments with…
- AI slop Spread in 2024 as a term for unwanted AI content, much as 'spam' named unwanted email.
- AI worm Researchers demonstrated a generative-AI worm called Morris II in 2024 using adversarial self-replicating prompts in AI email assistants. In…
- alignment faking Anthropic and Redwood Research introduced the term to a broad audience in December 2024 in experiments showing a model sometimes complying…
- character training Anthropic described its approach to 'character training' for Claude in 2024; the idea has since spread across AI labs.
- compound AI system Researchers at the Berkeley Artificial Intelligence Research lab popularized the phrase in a February 2024 article arguing that improvements in AI…
- computer use Released as a public beta by Anthropic in October 2024; other labs launched browser-using agents in 2025.
- deep research Launched by Google in December 2024, followed by similar features from other AI companies in 2025.
- EU AI Act Regulation (EU) 2024/1689 was published in the Official Journal of the European Union on July 12, 2024, and entered into force on August 1, 2024. Its…
- full-duplex speech model The networking term full duplex is much older. In spoken AI, the phrase gained prominence in 2024 with systems including Kyutai's Moshi, described as…
- generative monoculture Researchers introduced generative monoculture as a named large-language-model behavior in a 2024 preprint later published at ICLR 2025. The term…
- generative UI Generative UI emerged in AI developer and human-computer-interaction work in the mid-2020s as language models began producing interfaces at design…
- llms.txt Proposed in 2024 by Jeremy Howard of Answer.AI.
- MCP Released as an open standard by Anthropic in November 2024 and adopted by other major AI companies during 2025.
- model welfare Grew out of philosophy and AI safety research around 2024; Anthropic launched a dedicated model welfare research program in April 2025.
- neocloud The label became common in technology and financial reporting in 2024–2025 as specialized providers such as CoreWeave, Lambda, Crusoe, and Nebius…
- one-person unicorn The idea became prominent in 2024 after OpenAI CEO Sam Altman publicly discussed a prediction that AI could enable a one-person billion-dollar…
- physical AI Promoted heavily by industry leaders from 2024, especially in robotics and chip announcements.
- prefill-decode disaggregation The DistServe research published in 2024 made prefill-decode disaggregation prominent by separating the two stages across GPUs and optimizing their…
- prompt caching reusing the already-processed beginning of a prompt across many requests, which cuts cost and speeds up responses when the same long instructions or…
- reasoning model The category took shape in September 2024 with OpenAI's o1, and spread quickly to other labs in 2025.
- reinforcement learning with verifiable rewards The 2024 Tulu 3 report introduced Reinforcement Learning with Verifiable Rewards, abbreviated RLVR, as a named post-training method. The term spread…
- repository-level code generation Code-generation research initially emphasized snippets and individual functions. By 2024, repository-level code generation had become a named task in…
- safety case Safety cases originated in safety-critical fields such as aviation and nuclear engineering. Researchers began publishing detailed proposals for…
- sandbagging The word long described deliberate underperformance in sports and negotiation. AI-safety researchers formalized AI sandbagging in 2024 as the…
- scheming The ordinary word is centuries old. Apollo Research and collaborators established in-context scheming as a prominent AI-safety term in December 2024…
- sovereign AI Widely used from 2024 in government and industry discussions.
- spatial intelligence Spatial intelligence was first established in psychology and cognitive science as a human capacity for understanding spatial relationships. In AI…
- state-based evaluation The 2024 tau-bench benchmark evaluated tool-using agents by comparing the database state at the end of a conversation with an annotated goal state.…
- test-time compute Became a central idea in 2024 with the arrival of reasoning models.
- YOLO mode The AI-agent meaning became visible in December 2024 when Cursor 0.44 introduced Yolo Mode for automatically running terminal commands. As coding…
-
2023 40
- activation steering The modern language-model technique took shape in 2023 through work on inference-time interventions and activation engineering. Turner and colleagues…
- AI agent 'Agent' is an old AI term; the modern sense of an LLM that acts in loops with tools took off in 2023 and became the industry's main focus in 2025.
- AI constitution Anthropic published the principles behind its Claude models in 2023 and has revised them since; other labs publish similar documents, sometimes…
- AI humanizer AI humanizer became a common product label after generative writing tools and AI-text detectors spread widely. Research by 2025 treated humanizers as…
- AI management system The abbreviation AIMS follows established names such as quality and information-security management systems. ISO/IEC 42001:2023, published in…
- AI pause The phrase became prominent after the Future of Life Institute published an open letter on March 22, 2023 calling for a public, verifiable six-month…
- AI research agent AI research agent was documented as a technical term by 2023 in work on MLAgentBench, which evaluated language-model agents performing iterative…
- automated alignment researcher OpenAI set the goal of building 'a roughly human-level automated alignment researcher' when it launched its Superalignment team in July 2023. In 2026…
- autonomous replication and adaptation The term was introduced as autonomous replication and adaptation, abbreviated ARA, in a 2023 report by researchers at METR and collaborators that…
- direct preference optimization Rafael Rafailov and collaborators introduced Direct Preference Optimization in a 2023 paper titled Your Language Model is Secretly a Reward Model.…
- doomer (often dismissive) a person who believes advanced AI is likely to cause catastrophe and argues for slowing or stopping its development
- excessive agency OWASP included Excessive Agency in its 2023 Top 10 for Large Language Model Applications and retained it as LLM06 in the 2025 edition. OWASP…
- fiduciary AI Legal and computer-science researchers connected fiduciary duties to AI design in 2023 through work including Large Language Models as Fiduciaries…
- frontier model Became common in 2023, when leading labs and governments began discussing special rules for the most powerful systems.
- function calling Introduced as an API feature by OpenAI in June 2023, then adopted across the industry.
- generative engine optimization Named in a 2023 research paper by computer scientists at Princeton and other universities; it became a marketing buzzword in 2025.
- GPU-backed loan GPU-backed financing became widely discussed in the AI infrastructure market after CoreWeave announced a $2.3 billion debt facility in August 2023…
- GPU-poor SemiAnalysis popularized the contrast between GPU-rich and GPU-poor AI companies in 2023 as access to large clusters became a defining competitive…
- GTM engineer Coined around 2023 by the sales software company Clay; by 2026 job listings for the title were appearing at many software companies.
- jagged frontier From a 2023 study by Harvard Business School researchers and consultants testing AI at work.
- large behavior model Toyota Research Institute publicly described its goal of building Large Behavior Models for robots in September 2023, drawing an analogy with large…
- LLM-as-a-judge using one language model to grade or compare the outputs of another, as a cheaper and faster alternative to human reviewers
- model autonomy Model autonomy became a named frontier-risk category in OpenAI's 2023 Preparedness Framework and appeared in subsequent system cards. Early versions…
- model collapse Described by researchers in a 2023 paper, later published in Nature in 2024.
- on-policy distillation A 2023 paper on distilling language models from self-generated mistakes formalized the approach through Generalized Knowledge Distillation, training…
- open weights Used from 2023 to distinguish freely downloadable models from fully open-source ones.
- pro-worker AI Economists Daron Acemoglu, David Autor, and Simon Johnson used the phrase in a 2023 policy proposal about redirecting artificial intelligence toward…
- small language model a compact language model, usually with a few billion parameters or fewer, designed to run cheaply or directly on a phone or laptop
- sparse autoencoder Sparse autoencoders predate large language models, but the abbreviation SAE became prominent in mechanistic-interpretability research in 2023 after…
- speculative decoding The term was introduced in the 2023 ICML paper 'Fast Inference from Transformers via Speculative Decoding' by Yaniv Leviathan, Matan Kalman, and…
- sub-1-bit quantization Research on sub-1-bit neural-network compression was established by 2023 in work that pushed average storage below one bit per parameter, including…
- sycophancy Studied in AI research from about 2022; a 2025 chatbot update that flattered users excessively made the word widely known.
- synthetic audience Synthetic audiences entered AI-assisted market research in 2023 as large language models began to be used to simulate survey participants and…
- synthetic performer Synthetic performer became a defined AI labor term in the 2023 SAG-AFTRA television and theatrical agreement, which distinguished a wholly synthetic…
- system prompt hidden instructions given to a model before the conversation starts, which set its role, tone, rules, and limits for every reply
- tabular foundation model The term grew out of work on prior-data fitted networks and TabPFN, first released as research in the early 2020s. A 2025 Nature paper described…
- text-to-vector Research systems explored text-to-vector and text-to-SVG generation in the early 2020s. Adobe introduced Text to Vector Graphic in Illustrator as a…
- vision-language-action model The term spread in 2023 with research robots that used language-model technology to control movement.
- visual chain-of-thought Visual Chain of Thought was introduced as the name of a multimodal reasoning method in a 2023 research paper that used visual guidance to fill gaps…
- wrapper Used widely from 2023 as thousands of products were built on the same few model APIs.
-
2022 15
- accelerationism 'Effective accelerationism' (e/acc) emerged online in 2022 as a reply to the 'effective altruism' movement's focus on AI risk.
- activation patching The intervention has roots in causal mediation and interchange-intervention research. It became common in transformer interpretability in 2022…
- capability overhang Capability overhang appeared in public AI discussion by 2022 for hidden or not-yet-elicited abilities in existing systems. In 2026, OpenAI used the…
- chain of thought Named in a 2022 research paper from Google showing that 'let's think step by step' style prompting improves reasoning.
- constitutional AI Introduced by Anthropic in a December 2022 paper.
- diffusion language model The 2022 NeurIPS paper Diffusion-LM Improves Controllable Text Generation introduced Diffusion-LM as a non-autoregressive language model based on…
- E-E-A-T Google introduced the name E-E-A-T, also called Double-E-A-T, in December 2022 when it added experience to the earlier E-A-T framework in its Search…
- emergent abilities Popularized by a 2022 paper from Google researchers.
- flow matching Yaron Lipman and collaborators introduced Flow Matching for generative modeling in a 2022 paper. They described a simulation-free training objective…
- generative AI The phrase existed in research for years but became a household term after the image-generator and ChatGPT launches of 2022.
- in-context reinforcement learning The phrase was established by the 2022 paper In-context Reinforcement Learning with Algorithm Distillation. That work trained a causal sequence model…
- jailbreak Borrowed from phone culture; in the AI sense it spread in late 2022 after ChatGPT's release.
- prompt injection Named by programmer Simon Willison in September 2022, by analogy with 'SQL injection'.
- shoggoth The creature comes from H. P. Lovecraft's 1936 novella 'At the Mountains of Madness'. The AI meme appeared on Twitter in late December 2022, showing…
- text-to-video describing AI that generates video clips from a written description
-
2021 12
- Content Credentials The Coalition for Content Provenance and Authenticity (C2PA) formed in 2021 and released a first public specification for content provenance in 2022.…
- copilot The AI sense spread after GitHub Copilot, a coding assistant, was announced in 2021.
- cross-embodiment learning Cross-embodiment appeared as a robotics research label by 2021 in work on transferring skills between human and robot bodies. The field gained wider…
- discrete diffusion Discrete-state diffusion methods predate the modern diffusion-model boom, but the 2021 NeurIPS paper 'Structured Denoising Diffusion Models in…
- Edge MLOps The phrase appeared in research by 2021, when Reliable Fleet Analytics for Edge IoT Solutions described continuous delivery, deployment, and…
- foundation model Coined by researchers at Stanford University in an August 2021 report.
- GFlowNet The name GFlowNet was introduced in 2021 by researchers including Emmanuel Bengio, Mokshith Jain, Maksym Korablyov, Doina Precup, and Yoshua Bengio.…
- instruction tuning Became a standard step in 2021 and 2022 with research such as Google's FLAN and OpenAI's InstructGPT.
- knowledge editing Knowledge editing grew into a distinct language-model research area in the early 2020s through work on locating and modifying factual associations in…
- LoRA Introduced by Microsoft researchers in 2021. It became hugely popular in the image-generation community.
- stochastic parrot From a 2021 paper by Emily Bender, Timnit Gebru and colleagues on the risks of large language models.
- text-to-image Became mainstream with OpenAI's DALL·E in 2021 and a wave of image generators in 2022.
-
2020 12
- AI factory Marco Iansiti and Karim R. Lakhani used AI factory in their 2020 book Competing in the Age of AI for a digital operating model that industrializes…
- AI for science AI for science was already an established research-policy label by 2020, when the US Department of Energy published an AI for Science report…
- CXL memory pooling The Compute Express Link Consortium introduced memory pooling as a feature of the CXL 2.0 specification released in November 2020. The architecture…
- few-shot Made famous by the 2020 GPT-3 paper, 'Language Models are Few-Shot Learners'.
- in-context learning a model's ability to pick up a new task from instructions or examples given in the prompt, without changing its weights
- late interaction The label was established by the 2020 ColBERT paper, which introduced contextualized late interaction over BERT representations. ColBERT stores a set…
- neural operator Modern neural-operator research took shape around 2020 in work that generalized neural networks from mappings between finite-dimensional vectors to…
- prompt engineering Spread after GPT-3 in 2020, when people discovered that wording changed results dramatically.
- retrieval-augmented generation Named in a 2020 paper by researchers at Facebook AI.
- scaling laws Described in a 2020 OpenAI paper and refined in 2022 by DeepMind's 'Chinchilla' study.
- test-time adaptation Adapting models after deployment has older roots, but test-time adaptation became a standard category name around 2020 through work such as Tent…
- test-time training Yu Sun and collaborators proposed Test-Time Training as a general method in a 2020 ICML paper. Their approach turned each unlabeled test sample into…
-
2019 7
- AI anxiety Researchers had studied computer anxiety and technology anxiety for decades before AI became a distinct focus. In 2019, Yu-Yin Wang and Yi-Shun Wang…
- bitter lesson From a 2019 essay by reinforcement-learning pioneer Richard Sutton.
- closed-loop discovery The phrase was in research use by 2019 in drug and materials discovery, where scientists described linking machine-learning design, automated…
- sparse attention Sparse forms of attention appeared in early Transformer research, and the 2019 Sparse Transformer work made the term prominent by replacing full…
- TinyML The field coalesced under the name TinyML around 2019, when the tinyML community held its first meetings and Pete Warden and Daniel Situnayake's book…
- top-p Introduced as 'nucleus sampling' in a 2019 research paper.
- zero-click search Used by search marketers from around 2019 for featured snippets; it became a crisis term for publishers once AI summaries appeared at the top of…
-
2018 7
- data void Named by researchers Michael Golebiewski and danah boyd in a 2018 report about search engines.
- decision intelligence The modern discipline grew from earlier work in decision engineering and was popularized at Google in 2018, when Cassie Kozyrkov led a practice…
- GPT The name of a series of models from OpenAI, the first released in 2018.
- hallucination Used in AI research since the late 2010s. It went mainstream after ChatGPT, and Cambridge Dictionary chose 'hallucinate' as its 2023 Word of the Year…
- large language model The phrase spread after 2018, as transformer models such as GPT and BERT grew from millions to billions of parameters.
- model card Proposed in a 2018 research paper by Margaret Mitchell and colleagues.
- self-distillation The modern deep-learning use is commonly traced to the 2018 Born-Again Neural Networks work, which trained a new model of the same architecture from…
-
2017 8
- AI bubble Writers were explicitly warning about an artificial-intelligence bubble by 2017. The phrase became much more common after the generative-AI boom that…
- AI diffusion Diffusion has long been used in economics and innovation research for the spread of a technology after invention. AI diffusion appeared explicitly in…
- AI productivity paradox Erik Brynjolfsson, Daniel Rock, and Chad Syverson used Artificial Intelligence and the Modern Productivity Paradox in a November 2017 NBER working…
- deepfake From the username of a Reddit account that shared face-swapped videos in 2017; a blend of 'deep learning' and 'fake'.
- neuralese Coined in a 2017 research paper about translating messages between AI agents; it became a safety talking point in 2025 and 2026.
- proof of personhood The term proof-of-personhood was introduced in a 2017 paper on anonymous credentials for permissionless systems. It became newly prominent in the…
- RLHF Described in a 2017 research paper and made famous by OpenAI's InstructGPT in 2022, the approach behind ChatGPT.
- transformer Introduced by Google researchers in the 2017 paper 'Attention Is All You Need'.
-
2016 7
- AI kill switch AI researchers Laurent Orseau and Stuart Armstrong formalized the closely related idea of safely interruptible agents in 2016, using the image of a…
- artificial symbiotic intelligence The phrase artificial symbiotic intelligence appeared in Hitachi strategy materials by 2016. In September 2026, a DeepMind Institute essay by…
- federated learning Google researchers used federated learning in 2016 for a decentralized training approach and described the method in the 2017 paper…
- griefbot Academic discussion of griefbots appeared by 2016, the same year a widely reported chatbot was built from the messages of a deceased friend. The…
- membership inference attack Reza Shokri and collaborators formalized and experimentally evaluated membership inference against machine-learning-as-a-service models in work…
- reward hacking Discussed in the influential 2016 paper 'Concrete Problems in AI Safety'.
- TPU Announced by Google in 2016.
-
2015 4
- algorithmic management The expression became established in research on digital labor platforms during the mid-2010s, where apps used data and automated rules to allocate…
- diffusion model Proposed in 2015, then made practical by a 2020 paper on denoising diffusion. It powered the image-generation boom of 2022.
- distillation Popularized by Geoffrey Hinton, Oriol Vinyals and Jeff Dean in 2015.
- machine unlearning The modern term was introduced in the 2015 IEEE Symposium on Security and Privacy paper Towards Making Systems Forget with Machine Unlearning. It…
-
2010s 8
- artificial superintelligence 2014 The idea goes back to I. J. Good's 'ultraintelligent machine' in 1965; Nick Bostrom's 2014 book 'Superintelligence' brought the term into wide use.
- attention 2014 Proposed for machine translation in 2014 by Dzmitry Bahdanau and colleagues, then made the centerpiece of the transformer in 2017.
- disengagement 2014 California created autonomous-vehicle disengagement reporting requirements in 2014, making the term widely visible as companies began publishing how…
- embedding 2013 Word embeddings became widely used after Google's word2vec method was published in 2013.
- HBM 2013 The first JEDEC High Bandwidth Memory standard, JESD235, was published in October 2013. HBM first appeared in commercial graphics hardware in 2015…
- instrumental convergence 2012 Stephen Omohundro described related tendencies as basic AI drives in 2008, including self-protection, preserving utility functions, and acquiring…
- forward-deployed engineer 2010s The role was associated with companies such as Palantir before the generative-AI boom. It became a prominent AI job title in 2025–2026 as OpenAI…
- p(doom) 2010s Written like a probability formula. It began as an in-joke among AI researchers in the 2010s and went mainstream in 2023.
-
2000s 4
- distribution shift 2008 Machine-learning researchers studied changes between training and test distributions before the term was standardized. The 2008 MIT Press volume…
- CUDA 2007 Released by Nvidia in 2007.
- contextual integrity 2004 Privacy scholar Helen Nissenbaum introduced contextual integrity as a privacy benchmark in the 2004 article Privacy as Contextual Integrity. The…
- digital twin 2002 The concept was introduced in manufacturing in the early 2000s and later adopted by NASA and urban planners.
-
1990s 10
- GPU 1999 The term was popularized by Nvidia in 1999. Researchers began using GPUs for neural networks in the late 2000s.
- imitation learning 1999 Learning robots from human demonstrations had been studied earlier, but Stefan Schaal's 1999 review Is imitation learning the route to humanoid…
- conformal prediction 1998 Vladimir Vovk, Alex Gammerman, and Glenn Shafer developed conformal prediction in the late 1990s. Their 2005 book Algorithmic Learning in a Random…
- artificial general intelligence 1997 First used in 1997 and popularized in the 2000s by researchers who wanted to separate general intelligence from narrow AI.
- memory wall 1995 Computer architects William A. Wulf and Sally A. McKee popularized the phrase in their 1995 paper Hitting the Memory Wall, which described the…
- nerf 1995 The verb meant 'to make less effective' by 1995 and became common gaming slang for weakening something in a balance update. AI users later borrowed…
- chatbot 1994 Shortened from 'chatterbot', a term coined in 1994. The first famous example, ELIZA, dates from 1966.
- singularity 1993 Popularized by science fiction writer and mathematician Vernor Vinge in a 1993 essay, building on earlier ideas by John von Neumann.
- mixture of experts 1991 Described in a 1991 paper by Robert Jacobs, Michael Jordan, Steven Nowlan and Geoffrey Hinton, and revived for large language models around 2017.
- personal agent 1990s Researchers such as Pattie Maes at MIT built 'personal assistant' software agents in the early 1990s for tasks like scheduling meetings and filtering…
-
1980s 8
- catastrophic forgetting 1989 The problem was documented in late-1980s connectionist research, notably work by Michael McCloskey and Neal Cohen in 1989, and was also called…
- global workspace 1988 Cognitive scientist Bernard Baars developed global workspace theory in the late 1980s to explain conscious access. In 2026, Anthropic reported…
- Moravec's paradox 1988 Named after roboticist Hans Moravec, who described it in his 1988 book 'Mind Children'.
- backpropagation 1986 Popularized in a 1986 paper by David Rumelhart, Geoffrey Hinton and Ronald Williams, building on earlier work.
- concept drift 1986 The phrase appeared in machine-learning research by 1986 in Jeffrey Schlimmer and Richard Granger's work on tracking concept drift. It later became a…
- deep learning 1986 Applied to machine learning by Rina Dechter in 1986. It took off around 2012, when a deep network trained on GPUs won the ImageNet image-recognition…
- AI winter 1984 Coined at a 1984 meeting of the American Association for Artificial Intelligence, modeled on 'nuclear winter'.
- narrow AI 1980s Used as a contrast to 'strong' or 'general' AI. The weak/strong distinction goes back to philosopher John Searle in 1980.
-
Before 1980 14
- reference monitor 1972 The reference-monitor concept was introduced in early computer-security work and formalized in James P. Anderson's 1972 Computer Security Technology…
- recursive self-improvement 1965 The idea goes back to mathematician I. J. Good's 1965 description of an 'intelligence explosion'.
- computer vision 1960s An MIT summer project in 1966 famously expected to 'solve' vision in a few months. It took about fifty years.
- machine learning 1959 Popularized by IBM researcher Arthur Samuel in 1959, describing a checkers program that improved by playing against itself.
- proprietary data 1958 The phrase predates modern artificial intelligence and is documented in United States government procurement records from 1958, where it referred to…
- parasocial 1956 Coined by sociologists Donald Horton and Richard Wohl in 1956 to describe TV audiences. Cambridge Dictionary named it its 2025 Word of the Year…
- artificial intelligence 1955 Coined by John McCarthy and colleagues in a 1955 proposal for a summer research workshop at Dartmouth College, held in 1956. The workshop is usually…
- teleoperation 1952 Modern teleoperation grew from electrically actuated master-slave manipulators developed in the early 1950s, including Raymond Goertz's 1952 system…
- natural language processing 1950s Work on machine translation in the 1950s marks the start of the field.
- Turing test 1950 Proposed by mathematician Alan Turing in his 1950 paper 'Computing Machinery and Intelligence', where he called it the 'imitation game'.
- neural network 1943 Loosely inspired by the brain. Warren McCulloch and Walter Pitts described a mathematical model of a neuron in 1943.
- degrees of freedom 1900 The phrase degrees of freedom was recorded in scientific English by the early 1900s and became standard in mechanics before being applied to robot…
- gradient descent 1847 Described by the French mathematician Augustin-Louis Cauchy in 1847, long before computers.
- anthropomorphism 1753 Anthropomorphism is recorded in English from 1753 and derives from Greek roots meaning human and form. The concept long predates computing; it became…