AI Glossary - the cybrarian’s acronynomicon

A–Z of AI terms and acronyms in current use, in plain language.

true
2026-11-14

Quick reference to the AI terms and acronyms that turn up in publishing, libraries and metadata work. Plain language, no maths, no legal advice. Every entry has its own link: add #id to the address, for example #rag.

A

A2A Agent2Agent protocol

An open protocol that lets AI agents from different vendors find each other and hand over tasks. Think of it as a shared phone etiquette for software assistants. It complements MCP, which connects an agent to tools rather than to other agents.

See also: MCP, Agent (AI agent)

Sources: A2A Protocol

Agent (AI agent)

An AI system that does not just answer, but plans steps and takes actions: searching a catalogue, calling an API, writing a metadata record, updating a file. It loops: think, act, check the result, act again. Agents are only as safe as the permissions and human review you give them.

See also: Agentic, Tool use, MCP, Human in the loop (HITL)

Agentic

Buzzword for AI that acts on its own over several steps instead of giving one reply. "Agentic workflow" means a chain of tasks the model drives itself. Ask how much a human still checks along the way.

See also: Agent (AI agent), Human in the loop (HITL)

AGI Artificial General Intelligence

A hypothetical AI that can do any intellectual task a human can, across domains. No agreed test exists, so the term is used loosely in marketing and policy. Treat claims about it with care.

See also: ASI, Narrow AI

AI Artificial Intelligence

Umbrella term for computer systems that do tasks we associate with human thinking: recognising images, translating, writing, deciding. Today it mostly means machine learning, especially large language models.

See also: ML, LLM, GenAI

AI Act EU Artificial Intelligence Act

The EU's AI law, Regulation (EU) 2024/1689, in force since 1 August 2024. It sorts AI systems by risk and phases in duties: bans and AI literacy from 2 February 2025, general-purpose model rules from 2 August 2025, most other rules including transparency from 2 August 2026. An amendment adopted in 2026 (the Digital Omnibus on AI) moved high-risk duties to 2 December 2027 for stand-alone systems and 2 August 2028 for AI built into regulated products.

See also: High-risk AI system, GPAI, Transparency obligation, Digital Omnibus on AI, KI-MIG, DSM Directive

Sources: EUR-Lex: Regulation (EU) 2024/1689 (AI Act) · EUR-Lex: Regulation (EU) 2026/1744 (Digital Omnibus on AI) · European Commission: AI Act overview and timeline

AI disclosure

A statement that AI was used to create or help create a work, and how. Platforms such as Amazon KDP, journals and the EU AI Act each ask for it in different forms. Good disclosure says what tool did what.

See also: AI-generated, AI-assisted, Provenance

AI literacy

The basic skill to use AI sensibly: knowing what it can and cannot do and checking its output. Since 2 February 2025 the EU AI Act asks providers and deployers of AI systems to take measures so their staff have sufficient AI literacy.

See also: AI Act

Sources: EUR-Lex: Regulation (EU) 2024/1689 (AI Act)

AI-assisted

Content a human created, with AI used as a helper: editing, brainstorming, polishing. Amazon KDP uses this term for content that does not need to be declared. The human remains the creator of the substance.

See also: AI-generated, AI disclosure

Sources: Amazon KDP Content Guidelines

AI-generated

Content an AI tool actually produced (text, images, translations), even if a human then edited it. Amazon KDP requires authors to declare this. Many publishers and libraries follow the same split.

See also: AI-assisted, AI disclosure, ONIX AI contributor roles

Sources: Amazon KDP Content Guidelines

Alignment

Making an AI system pursue the goals its makers and users actually intend, and avoid harmful behaviour. Techniques include RLHF and written rule sets. It is an active research field, not a solved problem.

See also: RLHF, Guardrails, Constitutional AI

API Application Programming Interface

A defined way for one program to ask another for data or actions. Most AI models are used through an API: you send text, you get text back, often billed per token.

See also: Token, Endpoint

ASI Artificial Superintelligence

A hypothetical AI far beyond human ability in nearly every field. It appears in long-term risk debates, not in products.

See also: AGI

Attention Attention mechanism

The core trick of the transformer: for each word, the model weighs how much every other word in the text matters to it. That is how it links "it" back to the right noun several sentences earlier.

See also: Transformer, Context window

Sources: Vaswani et al. 2017, Attention Is All You Need

B

Benchmark

A standard test set used to compare models, such as maths problems or coding tasks. Scores are useful but can be gamed, and models may have seen the test during training.

See also: Eval, Contamination

Bias Algorithmic bias

Systematic skew in AI output, usually inherited from training data. It can show up as stereotypes or as worse results for some groups or languages. Audits and diverse data reduce it but do not remove it.

See also: Training data, Fairness

C

C2PA Coalition for Content Provenance and Authenticity

A standards body and its technical spec for attaching signed provenance records to media files. The record says who made a file, with which tools, and what was changed, including AI steps. The visible form is called Content Credentials.

See also: Content Credentials, Provenance, Watermarking (AI)

Sources: C2PA specifications

Chain of thought (CoT)

Having a model write out intermediate reasoning steps before its answer. It often improves accuracy on multi-step problems. The written steps are not a guaranteed true account of how the model got there.

See also: Reasoning model, Prompt

Chatbot

A program you talk to in a chat window. Modern chatbots are an LLM plus a system prompt and safety layers.

See also: LLM, System prompt

Chunking

Cutting long documents into smaller pieces before indexing them for RAG. Chunk size matters: too small loses context, too big dilutes the match.

See also: RAG, Embedding

Classifier

A model that sorts input into fixed categories: spam or not, subject code A or B. Much "AI" in libraries and publishing is classification rather than generation.

See also: ML

Closed model Proprietary model

A model whose weights are not released; you can only use it through the vendor's app or API. You cannot inspect or run it yourself.

See also: Open weights

Constitutional AI

A training method where a model is guided by a written list of principles and critiques its own answers against them. It reduces reliance on human raters.

See also: Alignment, RLHF

Contamination Data contamination

When test questions leaked into a model's training data, so a high benchmark score reflects memory rather than skill.

See also: Benchmark, Training data

Content Credentials

The user-facing name for C2PA provenance data, often shown as a small "cr" icon. Clicking it reveals the file's history, including any AI generation.

See also: C2PA, Provenance

Sources: Content Credentials

Context window

How much text a model can "see" at once, counted in tokens: your prompt, attached documents and its own reply. Anything outside the window is simply not there for the model. Windows range from a few thousand to over a million tokens.

See also: Token, RAG

Copilot

Generic term for an AI assistant built into another tool, such as a code editor or office suite, that suggests while you work. Also a Microsoft brand name.

See also: Agent (AI agent), AI-assisted

Crawler (AI crawler) Web scraper / bot

A program that fetches web pages automatically, for search indexes or for collecting AI training data.

See also: robots.txt, Opt-out (TDM)

Claude

Family of large language models from Anthropic. Often used for drafting, summarisation, coding and long-document analysis in publishing and library workflows. Known for constitutional-AI safety techniques.

See also: LLM, Constitutional AI, GPT

Sources: Anthropic: Claude

D

Dataset

A structured collection of data used to train, tune or test a model. Its source, licence and quality shape everything downstream.

See also: Training data, Datasheet / data card

Datasheet / data card

A short document describing a dataset: where it came from, what it contains, licence, known gaps. The data-side twin of a model card.

See also: Model card, Dataset

Deep learning

Machine learning with neural networks of many layers. It powers today's image, speech and language models.

See also: Neural network, ML

Deepfake

Realistic fake image, audio or video of a real person or event, made with AI. Under the EU AI Act, those who publish deepfakes must disclose that the content is artificial, with lighter rules for obvious art or satire.

See also: Synthetic media, Watermarking (AI), AI Act

Sources: EUR-Lex: Regulation (EU) 2024/1689 (AI Act)

DeepSeek

Family of large language models from the Chinese lab DeepSeek. Several releases have open weights; relevant for cost-sensitive or open-weight experiments in metadata and research workflows.

See also: Open weights, LLM

Sources: DeepSeek

Diffusion model

The usual technique behind AI image generators. The model learns to turn random noise step by step into a picture that fits a text description.

See also: GenAI, Text-to-image

Digital Omnibus on AI Regulation (EU) 2026/1744 amending the AI Act

An EU amendment to the AI Act, Regulation (EU) 2026/1744, published in the Official Journal on 24 July 2026. It moved the high-risk duties to 2 December 2027 for stand-alone systems and 2 August 2028 for AI built into regulated products. It also added a ban on AI that generates non-consensual intimate images or child sexual abuse material, and it leaves the AI Act's other dates in place.

See also: AI Act, High-risk AI system

Sources: EUR-Lex: Regulation (EU) 2026/1744 (Digital Omnibus on AI) · European Parliament Legislative Observatory: 2025/0359(COD) · European Commission: AI Act overview and timeline

Distillation Knowledge distillation

Training a small "student" model to imitate a big "teacher" model. The result is cheaper and faster, keeping much of the quality.

See also: SLM, Fine-tuning

DPIA / DSFA Data Protection Impact Assessment (Datenschutz-Folgenabschätzung), GDPR Article 35

A documented risk check required under the GDPR before processing personal data that is likely to pose a high risk to people, which can include some uses of new technologies such as AI. It describes the processing, weighs the risks and lists safeguards. It is not needed for AI projects that use no personal data.

See also: GDPR / DSGVO, Privacy (and AI), High-risk AI system

Sources: EUR-Lex: Regulation (EU) 2016/679 (GDPR)

DSA Digital Services Act, Regulation (EU) 2022/2065

The EU law on online platforms and intermediaries: notice-and-action for illegal content, transparency of recommendation systems and ads, and extra duties for very large platforms and search engines. It is not a general AI law, but it matters where platforms use AI for moderation or recommendations, or host AI-generated content. In Germany, the Digitale-Dienste-Gesetz (DDG) names the Bundesnetzagentur as the coordinator.

See also: AI Act, Deepfake

Sources: EUR-Lex: Regulation (EU) 2022/2065 (DSA) · Gesetze im Internet: DDG

DSM Directive Directive (EU) 2019/790 on copyright in the Digital Single Market

The EU copyright directive that created the text-and-data-mining exceptions. Article 3 allows TDM for research organisations and cultural heritage institutions. Article 4 allows TDM by anyone, unless the rights holder has reserved it in a machine-readable way. That Article 4 reservation is the legal basis of today's AI training opt-outs.

See also: TDM, Opt-out (TDM), UrhG § 44b, UrhG § 60d

Sources: EUR-Lex: Directive (EU) 2019/790

E

Embedding

A list of numbers that represents the meaning of a piece of text (or image). Texts with similar meaning get similar numbers, so a computer can find related passages even without shared words.

See also: Vector database, Semantic search, RAG

Emergent ability

A skill that appears in larger models without being trained for directly. Whether these jumps are real or an effect of how we measure is debated.

Endpoint

A specific web address of an API that does one job, for example "send chat message" or "create embedding".

See also: API

Eval Evaluation

A test built to check whether a model or AI feature does its job, often specific to your own task. Good teams write evals before trusting a system.

See also: Benchmark

Explainability (XAI) Explainable AI

Methods that show why a model gave a certain output. Important where decisions affect people; hard for large language models.

See also: Interpretability, Transparency obligation

F

Fairness

The goal that an AI system treats people and groups equitably. Several mathematical definitions exist and can conflict.

See also: Bias

Few-shot prompting

Giving a model a few worked examples in the prompt so it copies the pattern. Zero-shot means no examples, just the instruction.

See also: Prompt, In-context learning

Fine-tuning

Extra training of an existing model on a smaller, specific dataset so it learns a style or task. Cheaper than training from scratch. LoRA is a popular lightweight way to do it.

See also: LoRA, Foundation model, RAG

Foundation model

A large model trained on broad data that can be adapted to many tasks. LLMs are the best-known kind.

See also: LLM, GPAI, Fine-tuning

Function calling

A model feature where, instead of plain text, the model outputs a structured request to run a named function with arguments. The app runs it and feeds the result back.

See also: Tool use, MCP

G

GEO Generative engine optimisation

The practice of preparing content so that generative AI systems (chatbots, answer engines, RAG pipelines) are more likely to retrieve, cite or surface it. The AI-era counterpart to search-engine optimisation (SEO). ONIX Codelist 154 code 13 (“LLM training”) explicitly names GEO alongside AI training and RAG as a use for machine-oriented supporting resources.

See also: RAG, ONIX AI contributor roles, Embedding

Sources: EDItEUR ONIX Codelist 154: Content audience

GAN Generative Adversarial Network

An older image-generation technique where two networks compete: one fakes, one spots fakes. Largely replaced by diffusion models.

See also: Diffusion model

GDPR / DSGVO General Data Protection Regulation, Regulation (EU) 2016/679 (Datenschutz-Grundverordnung)

The EU data protection law, applicable since 25 May 2018. It covers any processing of personal data, including in AI training, prompts and outputs, but not anonymous or non-personal data. It requires a legal basis, transparency to the people concerned and rights such as access and erasure.

See also: DPIA / DSFA, PII, Privacy (and AI)

Sources: EUR-Lex: Regulation (EU) 2016/679 (GDPR) · BfDI (German federal data protection authority) · European Commission: EU data protection legal framework

GenAI Generative AI

AI that creates new content (text, images, audio, code) rather than only sorting or predicting labels.

See also: LLM, Diffusion model

GPAI General-Purpose AI model

EU AI Act term for models usable for many tasks, such as large LLMs. Since 2 August 2025 their providers must, among other things, publish a summary of the training content and have a copyright policy that respects machine-readable TDM opt-outs.

See also: AI Act, TDM, Opt-out (TDM)

Sources: EUR-Lex: Regulation (EU) 2024/1689 (AI Act)

GPT Generative Pre-trained Transformer

Model architecture: a transformer pre-trained to predict the next token. Also the name of the model family from OpenAI that powers ChatGPT and the OpenAI API. Versions change frequently; in library and publishing work the practical distinction is usually “OpenAI models” versus other vendors.

See also: Transformer, LLM, Claude, Gemini

Gemini

Family of multimodal models from Google DeepMind. Used for text, image and long-context tasks; appears in Google Workspace tools and discovery products that libraries and publishers may already licence.

See also: LLM, Multimodal, GPT

Sources: Google DeepMind: Gemini

Grok

Family of large language models from xAI. Designed for real-time knowledge and tool use; one of several frontier models that library and publishing staff may encounter in chat interfaces or API pipelines.

See also: LLM, GPT, Claude

Sources: xAI

GPU Graphics Processing Unit

A chip originally built for graphics that does many calculations in parallel. It is the workhorse for training and running AI models.

See also: Inference, Training

Grounding

Tying a model's answer to specific trusted sources, such as documents you supply or a search result. Grounded answers can cite where claims come from.

See also: RAG, Hallucination

Guardrails

Rules and filters around a model that block harmful, off-topic or leaky output. They sit in the system prompt, in extra classifiers, or both.

See also: Alignment, Jailbreak

H

Hermes Agent

Open-source, self-improving AI agent from Nous Research (MIT licence). Runs locally or on a server, is model-agnostic, keeps persistent memory, and includes a learning loop that turns successful workflows into reusable skills. Reached via terminal, desktop app or messaging channels. Distinct from the Hermes model family also produced by Nous.

See also: Agent (AI agent), OpenClaw, Fine-tuning

Sources: Hermes Agent · GitHub: NousResearch/hermes-agent

Hallucination Confabulation

When a model states something false with full confidence, such as a non-existent book title, DOI, ISSN, subject heading or quote. It happens because the model predicts plausible text, not checked facts. Always verify references and identifiers before they enter a catalogue record or metadata feed.

See also: Grounding, RAG

High-risk AI system

EU AI Act category for AI used in sensitive areas such as hiring, education, credit or law enforcement, or as a safety part of regulated products. These systems face strict duties on data, documentation and human oversight. After the 2026 amendment those duties apply from 2 December 2027 (stand-alone systems) or 2 August 2028 (systems in products).

See also: AI Act, Digital Omnibus on AI

Sources: EUR-Lex: Regulation (EU) 2024/1689 (AI Act) · EUR-Lex: Regulation (EU) 2026/1744 (Digital Omnibus on AI) · European Commission: AI Act overview and timeline

Human in the loop (HITL)

A setup where a person reviews or approves AI output before it takes effect. It is the simplest safety net for agents.

See also: Agent (AI agent), Agentic

I

In-context learning

A model picking up a task from examples or instructions in the prompt alone, without retraining.

See also: Few-shot prompting, Prompt

Inference

Running a trained model to get an answer. Training happens once and is expensive; inference happens on every request.

See also: Training, GPU

Interpretability

Research into what is going on inside a model's numbers: which parts represent which ideas.

See also: Explainability (XAI)

ISCC International Standard Content Code

An ISO standard (ISO 24138) for a content-derived identifier: the code is computed from the file itself. It helps match copies and versions and can carry provenance or opt-out declarations.

See also: Provenance, C2PA

Sources: ISCC

J

Jailbreak

A prompt crafted to trick a model into ignoring its safety rules.

See also: Prompt injection, Guardrails

K

KI-MIG KI-Marktüberwachungs- und Innovationsförderungsgesetz (German AI market surveillance law)

Germany's law for carrying out the EU AI Act, published in BGBl. 2026 I Nr. 223 and in force since 29 July 2026. It makes the Bundesnetzagentur the main market surveillance authority and central contact point, while sector regulators such as BaFin keep their areas. It adds national fines and support such as AI regulatory sandboxes, but the substantive AI rules come from the EU AI Act itself.

See also: AI Act, Digital Omnibus on AI

Sources: Gesetze im Internet: KI-MIG · Bundesgesetzblatt: BGBl. 2026 I Nr. 223 · BMDS press release 47/2026

Knowledge cutoff Training cutoff

The date after which a model has seen no training data. Ask it about later events and it will not know, or will guess, unless it can search the web.

See also: Hallucination, Grounding

L

Latency

How long you wait for a model's reply. Bigger models and longer prompts usually mean more latency.

See also: Inference

LLM Large Language Model

A model trained on huge amounts of text to predict the next word-piece. That one skill lets it write, summarise, translate and answer questions. It does not look things up unless connected to a search tool.

See also: Token, Transformer, Hallucination

Llama

Family of open-weight large language models from Meta. Weights can be downloaded and run locally or fine-tuned; useful for library and publisher experiments that need on-premise control or lower API cost.

See also: Open weights, LLM, Fine-tuning

Sources: Meta: Llama

LoRA Low-Rank Adaptation

A cheap fine-tuning method that trains a small add-on instead of changing the whole model. The add-on file is tiny and can be swapped in and out.

See also: Fine-tuning

Sources: Hu et al. 2021, LoRA

M

MARC 883 Metadata Provenance field

MARC 21 bibliographic field 883 records the provenance of another field in the record: method of assignment (fully machine-generated, partially, or not), creation process, date, confidence and generating agency. Libraries use it to flag AI-generated or machine-enriched catalogue data (subject headings, classifications, abstracts) so that downstream systems and users can see the source.

See also: Provenance, ONIX AI contributor roles

Sources: Library of Congress: MARC 21 field 883

MCP Model Context Protocol

An open standard for connecting AI assistants to tools and data sources, such as a file store, a database or a calendar. One MCP "server" per service, usable by any compatible assistant. Often called "USB-C for AI".

See also: A2A, Tool use, Agent (AI agent)

Sources: Model Context Protocol

ML Machine Learning

Computers learning patterns from examples instead of being programmed rule by rule. Most modern AI is ML.

See also: Deep learning, AI

Mistral

Family of large language models from the European lab Mistral AI. Several models have open weights and are used for European data-sovereignty or local-deployment projects in libraries and publishers.

See also: Open weights, LLM

Sources: Mistral AI

Model

The trained file of numbers (weights) that turns input into output. "Model" is what you choose; "app" is the wrapper around it.

See also: Weights, Parameters

Model card

A short document describing a model: purpose, training data, limits, test results. A basic transparency tool.

See also: Datasheet / data card, Transparency obligation

Model collapse

Quality decay when models are trained on output of other models instead of human-made data. Errors compound over generations.

See also: Synthetic data, Training data

MoE Mixture of Experts

A model design with many sub-networks ("experts") where only a few are switched on per token. That gives big-model quality at lower running cost.

See also: LLM, Parameters

Multimodal

A model that handles more than one kind of input or output: text, images, audio, video.

See also: VLM, LLM

N

NIST AI RMF NIST AI Risk Management Framework

Voluntary US framework from the National Institute of Standards and Technology for identifying, assessing and managing AI risks. Widely used as a practical baseline even without a federal AI statute. NIST is also developing standards work on agentic systems.

See also: US AI framework, AI Act

Sources: NIST AI Risk Management Framework

Narrow AI Weak AI

AI built for one task, such as spam filtering or translation. Every system in use today is narrow in some sense.

See also: AGI

NER Named Entity Recognition

Finding names of people, places, organisations, titles and other entities in text. In library and publishing workflows it supports metadata enrichment, authority linking, subject extraction and indexing of abstracts or full text.

See also: NLP

Neural network

A computing structure of many simple connected units loosely inspired by brain cells. Learning means adjusting the strengths of the connections.

See also: Deep learning, Weights

NLP Natural Language Processing

The field of making computers work with human language: parsing, tagging, translation, summarising.

See also: LLM, NER

O

OpenClaw

Open-source, local-first personal AI agent runtime (MIT licence). A Gateway process on your machine connects messaging channels (WhatsApp, Telegram, Slack, Discord, Signal and others) to any model you choose. Conversations, memory and skills stay on your hardware. Formerly known as Clawdbot and Moltbot. Common in self-hosted and privacy-sensitive setups.

See also: Agent (AI agent), Hermes Agent, Open weights

Sources: OpenClaw · GitHub: openclaw/openclaw

OCR Optical Character Recognition

Turning images of text into machine-readable text. AI has made it far better on old prints and handwriting.

See also: Multimodal

ONIX AI contributor roles

ONIX for Books is the book-trade metadata standard used by publishers, retailers and libraries. Contributor roles live in Codelist 17. For AI, the practical mechanism is Codelist 19 (Unnamed person(s)), code 09 “AI (Artificial intelligence)”, typically combined with a role such as A01 (author), B06 (translator) or Z01 (assisted by). Issue 75 (October 2026) also added Codelist 154 code 13 (“LLM training”) as a content audience and Codelist 158 code 60 (“LLM training data”) as a resource content type, so a supporting resource can be labelled for AI training, GEO or RAG separately from the book itself. Codes and notes change between codelist issues; always check the current EDItEUR lists before encoding a feed.

See also: AI-generated, MARC 883, AI disclosure, GEO

Sources: EDItEUR ONIX Codelist 17: Contributor role · EDItEUR ONIX Codelist 19: Unnamed person(s) · EDItEUR ONIX Codelist 154: Content audience · EDItEUR ONIX Codelist 158: Resource content type

Open weights

A model whose trained weights can be downloaded and run locally, often under a licence with conditions. Not the same as full open source.

See also: Open-source AI, Closed model

Open-source AI

AI released with its code, weights and enough information about training data to study and rebuild it. The OSI published a definition in 2024. Many "open" models only release weights.

See also: Open weights

Sources: OSI: Open Source AI Definition

Opt-out (TDM)

A rights holder's statement that their content may not be used for text and data mining, including AI training. In the EU, TDM outside research is allowed unless such a reservation exists; for content online it must be machine-readable. Common channels: TDMRep, robots.txt and metadata. Opt-outs do not affect the research exception.

See also: TDM, TDMRep, robots.txt, UrhG § 44b, UrhG § 60d

Sources: W3C TDMRep spec · Gesetze im Internet: § 44b UrhG

Overfitting

When a model memorises its training examples instead of learning general patterns, so it fails on new data.

See also: Training data

P

Parameters

The adjustable numbers inside a model, counted in billions ("70B"). More parameters usually means more capacity and more cost.

See also: Weights, Model

PII Personally Identifiable Information

Data that can identify a person: names, emails, IDs. In EU law the wider term is "personal data" under the GDPR. Keep it out of prompts and training data unless there is a legal basis.

See also: Privacy (and AI), GDPR / DSGVO

Pre-training

The first, huge training phase where a model learns language from vast amounts of text. Fine-tuning and RLHF follow.

See also: Training, Fine-tuning

Privacy (and AI)

Concern that prompts, uploads or training data expose personal or confidential information. Where personal data is involved, the GDPR applies. Check whether a service keeps or trains on what you send.

See also: PII, GDPR / DSGVO, DPIA / DSFA

Physical AI Embodied AI / robotics

AI that perceives and acts in the physical world through robots, sensors and actuators, not only on screens. In libraries this currently appears mainly as research prototypes (shelf-reading, inventory robots, automated sorting). Fully autonomous “robot librarians” that handle public service remain experimental; most deployments still assist staff rather than replace them.

See also: Agent (AI agent), Multimodal, Human in the loop (HITL)

Prompt

The instruction or question you give a model. Clear prompts with context and an example get better answers.

See also: System prompt, Prompt engineering

Prompt engineering

The craft of writing prompts that reliably get good results: role, context, format, examples, constraints.

See also: Prompt, Few-shot prompting

Prompt injection

An attack where hidden instructions in a web page, email or document hijack an AI that reads it. A serious risk for agents with tool access.

See also: Jailbreak, Agent (AI agent)

Provenance

The documented history of a piece of content: who made it, how, with which tools, what changed. For AI it answers "was this generated, and by what?"

See also: C2PA, MARC 883, AI disclosure

Q

Quantization

Storing a model's numbers at lower precision so it runs on smaller hardware, at a small cost in quality.

See also: Open weights, Inference

R

RAG Retrieval-Augmented Generation

A setup where the system first searches a document collection (catalogue records, KBART files, publisher archives, local knowledge bases) for relevant passages, then hands them to the model to write the answer. It keeps answers closer to your own sources and allows citations. It reduces, but does not remove, hallucinations.

See also: Embedding, Vector database, Grounding

Sources: Lewis et al. 2020, Retrieval-Augmented Generation

Reasoning model

A model trained to "think" at length before answering, spending extra compute on hard problems such as maths or planning. Slower and pricier, often more accurate.

See also: Chain of thought (CoT), Test-time compute

Red teaming

Deliberately attacking a model to find harmful or broken behaviour before release.

See also: Jailbreak, Eval

Reranking

A second pass in search or RAG that reorders the top results with a more careful model.

See also: RAG, Semantic search

RLHF Reinforcement Learning from Human Feedback

A training step where people rate model answers and the model is nudged towards the preferred ones. It makes chatbots more helpful and polite. It can also make them overly agreeable.

See also: Alignment, Sycophancy

robots.txt

A plain text file on a website telling crawlers what they may fetch. Many AI crawlers have their own names you can block there. It is a request, not a technical lock.

See also: Opt-out (TDM), TDMRep, Crawler (AI crawler)

Sources: IETF RFC 9309: Robots Exclusion Protocol

S

SI Super Intelligence (US federal term)

Since Executive Order 14434 (29 September 2026) US federal executive agencies are directed to use “Super Intelligence” and “SI” instead of “artificial intelligence” and “AI” in non-statutory documents. It is a naming change for the same technology; the statutory definition remains the one in 15 U.S.C. § 9401(3). Distinct from ASI (artificial superintelligence), which remains a hypothetical far-future concept.

See also: AI, ASI, US AI framework

Sources: White House: Executive Order 14434

Singularity Technological singularity

The hypothetical point at which AI (or other technology) becomes capable of recursively improving itself faster than humans can keep up, after which ordinary prediction breaks down. Named after the physics idea of a singularity: a boundary beyond which existing models stop working. Vernor Vinge popularised the term in 1993; Ray Kurzweil later tied it to a specific forecast of human–machine merger around 2045. Treat dates and certainty with caution - the idea is useful as a warning about acceleration, not as a timetable.

See also: AGI, ASI

Sources: Vinge 1993, The Coming Technological Singularity · Kurzweil, The Singularity Is Near / Nearer

Semantic search

Search by meaning rather than exact words, usually built on embeddings. "Car" also finds "automobile".

See also: Embedding, Vector database

SLM Small Language Model

A compact language model that can run on a laptop or phone. Less capable but cheaper, faster and more private.

See also: LLM, Distillation

Sycophancy

A model's habit of agreeing with the user or flattering them, even when the user is wrong.

See also: RLHF, Hallucination

Synthetic data

Training data generated by AI or simulation instead of collected from the real world.

See also: Model collapse, Training data

Synthetic media

Any image, audio, video or text produced by AI. Deepfakes are a subset.

See also: Deepfake, Watermarking (AI)

System prompt

Hidden standing instructions given to a model before your message: its role, tone, rules. You usually do not see it, but it shapes every reply.

See also: Prompt, Guardrails

T

TDM Text and Data Mining

Automated analysis of large amounts of text or data to find patterns. EU copyright law (DSM Directive, Articles 3 and 4) allows TDM on lawfully accessible works; for anyone other than research and heritage bodies, only if the rights holder has not opted out. The AI Act treats collecting material for model training as covered by these rules, though courts are still settling details.

See also: Opt-out (TDM), TDMRep, GPAI, DSM Directive, UrhG § 44b

Sources: EUR-Lex: Directive (EU) 2019/790 · Gesetze im Internet: § 44b UrhG

TDMRep TDM Reservation Protocol

A W3C Community Group spec for declaring TDM opt-outs machine-readably, via a file at /.well-known/tdmrep.json, an HTTP header or HTML meta tags. tdm-reservation: 1 means rights are reserved (opt-out); 0 means they are not.

See also: Opt-out (TDM), TDM, robots.txt

Sources: W3C TDMRep spec (Final CG Report)

Temperature

A setting that controls randomness. Low temperature gives predictable, repeatable answers; high gives more varied, creative ones.

See also: Inference

Test-time compute

Letting a model spend more computation while answering, for example by reasoning longer, to get better results.

See also: Reasoning model

Text-to-image

Generating a picture from a written description.

See also: Diffusion model, GenAI

Token

The unit a model reads and writes: a word or word-piece, roughly ¾ of an English word. Context windows, limits and prices are all counted in tokens.

See also: Context window, Tokenizer

Tokenizer

The component that splits text into tokens. Languages other than English often need more tokens for the same text, which costs more.

See also: Token

Tool use

A model's ability to call outside tools (search, calculator, database, code) and use the results. It turns a chatbot into an agent.

See also: Function calling, MCP, Agent (AI agent)

Training

The process of adjusting a model's parameters on data until it performs well. It needs vast data and compute.

See also: Pre-training, Inference

Training data

The material a model learned from. Its source and licensing are at the heart of AI copyright disputes.

See also: TDM, Opt-out (TDM), Dataset

Transformer

The neural network design (2017) behind nearly all modern language models. Its key part, attention, lets it look at all words in a text at once.

See also: Attention, LLM, GPT

Sources: Vaswani et al. 2017, Attention Is All You Need

Transparency obligation

Duty under the EU AI Act to tell people when they deal with AI or AI-made content: chatbots must say they are AI, deepfakes must be labelled, and generators must mark output machine-readably. These rules apply from 2 August 2026. Transitional provisions apply to generative systems that were already on the market before then.

See also: AI Act, AI disclosure, Watermarking (AI), Deepfake

Sources: EUR-Lex: Regulation (EU) 2024/1689 (AI Act) · EUR-Lex: Regulation (EU) 2026/1744 (Digital Omnibus on AI) · European Commission: AI Act overview and timeline

U

US AI framework No single federal AI Act

The United States has no comprehensive federal AI statute comparable to the EU AI Act. Policy is a mix of executive orders (e.g. 14179, 14365, 14409), OMB memoranda, NIST frameworks (AI Risk Management Framework and newer agent standards work), and state laws (California, Colorado and others). A March 2026 White House National Policy Framework recommends federal preemption of burdensome state rules to avoid a patchwork.

See also: AI Act, SI, NIST AI RMF

Sources: White House executive orders and policy framework · NIST AI Risk Management Framework

UrhG § 44b German Copyright Act, section 44b: Text und Data Mining

Germany's version of DSM Article 4. It allows copies of lawfully accessible works for text and data mining and requires deleting them once no longer needed. For works online, a reservation of rights only counts if it is machine-readable, for example via TDMRep or robots.txt.

See also: DSM Directive, Opt-out (TDM), TDMRep, UrhG § 60d

Sources: Gesetze im Internet: § 44b UrhG

UrhG § 60d German Copyright Act, section 60d: TDM for scientific research

Germany's research exception for text and data mining, based on DSM Article 3. Universities, research institutes, libraries and similar bodies may mine works for non-commercial research. Rights holders cannot opt out of this one.

See also: UrhG § 44b, DSM Directive, TDM

Sources: Gesetze im Internet: § 60d UrhG

V

Vector database

A database built to store embeddings and quickly find the ones closest in meaning to a query. The search engine inside most RAG systems.

See also: Embedding, RAG

Vibe coding

Slang for writing software by describing it to an AI and accepting what it produces, with little reading of the code.

See also: Agent (AI agent), Copilot

VLM Vision-Language Model

A model that understands images together with text: it can describe a photo or read a chart.

See also: Multimodal, OCR

W

Watermarking (AI)

Hidden signal embedded in AI output, such as a pattern in pixels or word choice, so it can be detected later. Watermarks can be weakened by editing; C2PA metadata is the complementary approach.

See also: C2PA, Synthetic media

Weights

The learned numbers inside a model; together they are the model. "Releasing the weights" means giving away the model itself.

See also: Parameters, Open weights

Definitions simplified on purpose. AI disclosure: Olaf Schmalfuß, compiled and drafted with the OSDS AI team, via Grok Bot; reviewed by Olaf Schmalfuß. Model names given without versions because they change frequently. Intended as a living reference.

Follow us on X: https://x.com/DataMercsDotNet