What is the AI Glossary?

The AI Glossary is a public reference registry of core artificial intelligence terminology within the AI Center. Each entry states a term, its definition, the named primary authority that defines it, a link to the source, and the date the definition was verified.

Entries draw on international standards such as ISO/IEC 22989:2022, developed by ISO/IEC JTC 1/SC 42; binding legislation such as Regulation (EU) 2024/1689 (the EU Artificial Intelligence Act); intergovernmental instruments such as the OECD Recommendation on Artificial Intelligence (OECD/LEGAL/0449); and national frameworks such as the NIST AI Risk Management Framework. Where a term carries distinct legal and technical meanings, each meaning is listed separately under its own authority.

Source: ISO/IEC 22989:2022 — Information technology — Artificial intelligence — Artificial intelligence concepts and terminology (ISO) ↗

Who defines what an "AI system" is?

Three anchors define the term "AI system": the OECD Recommendation on Artificial Intelligence (OECD/LEGAL/0449, revised definition), Article 3(1) of Regulation (EU) 2024/1689 (the EU AI Act), and the international standard ISO/IEC 22989:2022.

Article 3(1) of Regulation (EU) 2024/1689 defines an AI system as "a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments". Recital 12 of the same regulation states that the notion should be closely aligned with the work of international organisations working on AI. ISO/IEC 22989:2022 defines the term at clause 3.1.4, characterizing an AI system, in paraphrase, as an engineered system that produces outputs such as content, forecasts, recommendations, or decisions, in pursuit of a set of human-defined objectives.

Source: Regulation (EU) 2024/1689, Article 3(1) and Recital 12 — EUR-Lex ↗

What is the OECD definition of an AI system?

The OECD Recommendation on Artificial Intelligence (OECD/LEGAL/0449), adopted in May 2019 and updated in May 2024, defines an AI system as "a machine-based system that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments".

The OECD describes the Recommendation as the first intergovernmental standard on artificial intelligence. Recital 12 of Regulation (EU) 2024/1689 records the European Union's intent to keep its own definition closely aligned with the work of international organisations working on AI, and the two texts share the same structural elements: machine basis, explicit or implicit objectives, inference from input, and outputs that influence physical or virtual environments. The definition is maintained on the OECD AI Policy Observatory alongside the OECD AI Principles.

Source: OECD AI Principles, Recommendation of the Council on Artificial Intelligence (OECD/LEGAL/0449) — OECD.AI Policy Observatory ↗

How do legal definitions differ from technical definitions?

Legal definitions fix the scope of binding obligations; technical definitions standardize vocabulary for engineering and for the development of further standards. The same term can carry both kinds of meaning, and the glossary lists each meaning under its own authority.

Article 3(1) of Regulation (EU) 2024/1689 determines which systems fall within the regulation's scope, so its wording carries direct legal consequence. ISO/IEC 22989:2022, by contrast, is a vocabulary standard: its stated purpose, in paraphrase, is to provide the terminology and concepts on which other standards can build and which support communication among interested parties. Recital 12 of Regulation (EU) 2024/1689 connects the two registers, stating that the legal notion of an AI system should be closely aligned with the work of international organisations working on AI to ensure legal certainty and facilitate international convergence.

Source: Regulation (EU) 2024/1689, Recital 12 — EUR-Lex ↗

Which authorities anchor AI terminology?

Five authorities recur across the register: ISO/IEC JTC 1/SC 42 for international standards; the EU Artificial Intelligence Act for binding legal definitions; the OECD for the intergovernmental definition of an AI system; NIST for United States framework terminology; and IEEE for engineering standards.

ISO/IEC JTC 1/SC 42, created in 2017 with its secretariat held by ANSI (United States), is the joint ISO/IEC committee responsible for standardization in the area of artificial intelligence, including ISO/IEC 22989:2022. Regulation (EU) 2024/1689 of 13 June 2024 sets binding definitions in its Article 3. The OECD Recommendation on Artificial Intelligence (OECD/LEGAL/0449) supplies the intergovernmental definition of an AI system. NIST publishes the AI Risk Management Framework (AI RMF 1.0, NIST AI 100-1, January 2023) and the AI Resource Center glossary. IEEE maintains standards such as IEEE 7000-2021, the Standard Model Process for Addressing Ethical Concerns during System Design, published 15 September 2021.

Source: ISO/IEC JTC 1/SC 42 — Artificial intelligence (ISO committee page) ↗

How current are these definitions?

Each glossary entry carries its own as-of date and is verified against the live source. Definitions change as standards are revised and legislation is amended, so currency is recorded per term rather than for the page as a whole.

ISO/IEC 22989:2022 remains at edition 1, published July 2022, with a corrected French-language version issued in December 2025. The OECD Council adopted a revised definition of an AI system on 8 November 2023 and revised the Recommendation as a whole on 3 May 2024. The EU Digital Omnibus on AI, which amends Regulation (EU) 2024/1689, was approved by the European Parliament's plenary on 16 June 2026, adopted by the Council on 29 June 2026, signed on 8 July 2026, and published in the Official Journal of the European Union on 24 July 2026 as Regulation (EU) 2026/1744; it is in force from 27 July 2026. Glossary entries reflect Regulation (EU) 2024/1689 as amended by Regulation (EU) 2026/1744, which among other changes sets new application dates for Chapter III, Sections 1, 2 and 3.

Source: European Parliament Legislative Observatory — Procedure file 2025/0359(COD), Digital Omnibus on AI ↗

What is ISO/IEC 22989?

ISO/IEC 22989:2022, Information technology — Artificial intelligence — Artificial intelligence concepts and terminology, is the international standard that sets out the vocabulary of artificial intelligence and describes the field's core concepts. It was published in July 2022 and developed by ISO/IEC JTC 1/SC 42.

The standard is at edition 1, runs 60 pages, and holds the status International Standard published (stage 60.60); a corrected French-language version was issued in December 2025. Its abstract, in paraphrase, presents the document as a foundation on which other standards can be developed and as a common vocabulary that supports communication among a broad range of interested parties. The term "AI system" is defined at clause 3.1.4.

Source: ISO/IEC 22989:2022 — ISO ↗

Where does United States AI framework terminology come from?

The National Institute of Standards and Technology (NIST) anchors United States framework terminology through the AI Risk Management Framework (AI RMF 1.0, document NIST AI 100-1, released 26 January 2023) and the NIST AI Resource Center glossary, The Language of Trustworthy AI: An In-Depth Glossary of Terms.

The AI RMF 1.0 was developed to help manage risks to individuals, organizations, and society associated with artificial intelligence, and to improve the ability to incorporate trustworthiness considerations into the design, development, use, and evaluation of AI products, services, and systems. The glossary, hosted by the NIST AI Resource Center and released in beta format as a spreadsheet, seeks to promote a shared understanding and improve communication among individuals and organizations working to operationalize trustworthy and responsible AI; like the AI RMF, it is non-sector-specific and use-case-agnostic. As publications of a United States federal agency, these documents are freely accessible.

Source: NIST AI Resource Center — The Language of Trustworthy AI: An In-Depth Glossary of Terms ↗

Terms and definitions

accreditation

In European Union law, 'accreditation' means 'an attestation by a national accreditation body that a conformity assessment body meets the requirements set by harmonised standards and, where applicable, any additional requirements including those set out in relevant sectoral schemes, to carry out a specific conformity assessment activity'. Accreditation attests to the competence of the bodies that perform testing, inspection, and certification, not to products or AI systems themselves. ↗

adversarial machine learning

The field that studies attacks on machine-learning systems which exploit their statistical, data-dependent nature, together with the corresponding mitigations. Its taxonomy spans evasion, poisoning, and privacy attacks on predictive AI systems, and poisoning, direct prompting, and indirect prompt-injection attacks on generative AI systems. ↗

adversarial testing

Deliberate probing of an AI model or system with inputs and conditions designed to induce failure — drawing on attack classes catalogued in adversarial machine learning, including evasion inputs, data poisoning, privacy extraction, and prompting attacks that enable misuse — in order to identify and mitigate vulnerabilities before they are exploited. ↗

AI agent

An AI system that can perceive and act upon its environment with a degree of autonomy, using tools as needed to achieve specific goals and adapting to changing inputs and contexts. The OECD distinguishes AI agents from agentic AI, which generally refers to systems composed of multiple coordinated AI agents that break down tasks, collaborate, and pursue complex objectives over extended periods with minimal human supervision. ↗

AI alignment

The tendency of an AI model or system to apply its capabilities in accordance with human intentions, values, or norms; depending on context, the relevant intentions and values may be those of developers, users, specific communities, or society as a whole. The International AI Safety Report 2026 treats alignment in general as an open scientific problem, with research directions that include diversified training environments, interpretability techniques for detecting misalignment, and scalable oversight. ↗

AI assurance

The process of measuring, evaluating, and communicating the trustworthiness of AI systems. In the United Kingdom government's framing, assurance draws on mechanisms including risk assessment, algorithmic impact assessment, bias audit, compliance audit, conformity assessment, and formal verification, underpinned by global technical standards. ↗

AI audit

An independent, methodical process that gathers evidence and evaluates it objectively to determine the degree to which defined audit criteria are met, applied in this context to an organization's AI management system or AI-related processes. An audit may be internal (first party) or external (second or third party); third-party audits of AI management systems underpin certification by bodies meeting ISO/IEC 42006:2025. ↗

AI governance

The structures, policies, and processes through which an organization's governing body directs, oversees, and holds to account the organization's development and use of artificial intelligence. It extends established principles for the governance of information technology to decisions about AI across the AI system life cycle, and is distinct from the management activities that execute within the direction the governing body sets. ↗

AI hazard

An event, circumstance, or series of events in which the development, use, or malfunction of one or more AI systems could plausibly lead to an AI incident — injury or harm to health, disruption of the management and operation of critical infrastructure, violations of human rights or of protective obligations under applicable law, or harm to property, communities, or the environment — where such harm has not yet materialised. ↗

AI incident

An event, circumstance, or series of events in which the development, use, or malfunction of one or more AI systems directly or indirectly leads to actual harm: injury or harm to the health of a person or groups of people; disruption of the management and operation of critical infrastructure; violations of human rights or breaches of obligations under applicable law intended to protect fundamental, labour, and intellectual property rights; or harm to property, communities, or the environment. ↗

AI management system

The interconnected elements through which an organization establishes policies and objectives for the responsible development, provision, or use of AI systems, together with the processes designed to achieve those objectives. ISO/IEC 42001:2023 specifies requirements for establishing, implementing, maintaining, and continually improving such a system, follows the harmonized structure common to ISO management system standards, and serves as the basis for third-party certification. ↗

AI risk management

The organized, ongoing activities through which an organization identifies, assesses, treats, monitors, and communicates risk arising from the development, production, deployment, or use of products, systems, and services that employ artificial intelligence. ISO/IEC 23894:2023 provides guidance on integrating these activities into an organization's AI-related functions, applying the principles, framework, and process of ISO 31000:2018 to AI-specific sources of risk. ↗

AI safety

The characteristic of an AI system that, under defined conditions of operation, does not endanger human life, health, property, or the environment. The NIST AI Risk Management Framework identifies safety as a characteristic of trustworthy AI, supported by responsible design, development, and deployment practices; clear information to deployers on responsible use; responsible decision-making by deployers and end users; and documentation of risks based on empirical evidence from incidents. ↗

AI system

Under the EU Artificial Intelligence Act, "a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments". This definition, together with the Regulation's scope provisions, determines whether the Regulation's obligations attach to a given system. ↗

AI system impact assessment

A formal, documented process through which an organization that develops, provides, or uses AI systems examines the effects those systems may have on individuals, groups of individuals, and societies. ISO/IEC 42005:2025 provides guidance on how and when to perform and document such assessments and on integrating the assessment process into the organization's AI risk management and AI management system. ↗

AI transparency

The condition in which those who interact with or are affected by an AI system can obtain the information about it that their role requires — including the fact that AI is in use and that outputs are machine-generated. Article 50 of the EU AI Act gives this the force of law for certain systems and applies from 2 August 2026: providers must ensure that persons directly interacting with an AI system are informed of that fact and that synthetic audio, image, video, or text outputs are marked in a machine-readable format and detectable as artificially generated or manipulated, while deployers must disclose the operation of emotion recognition and biometric categorisation systems and the artificial origin of deep fakes. Under Article 111(4) of Regulation (EU) 2024/1689, as amended by Regulation (EU) 2026/1744, providers of AI systems generating synthetic audio, image, video or text content placed on the market before 2 August 2026 must comply with Article 50(2) by 2 December 2026. ↗

algorithm

A precisely specified procedure — a bounded, ordered set of unambiguous rules or steps — for solving a problem. In machine learning, a learning algorithm is the procedure that determines or adjusts a model's parameters on the basis of data. ↗

anonymisation

The processing of personal data in such a manner that the data subject is not or no longer identifiable, taking into account all the means reasonably likely to be used for identification — including objective factors such as the cost, time and available technology at the time of processing. Under GDPR Recital 26, data protection principles do not apply to such anonymous information, whereas pseudonymised data that could be attributed to a person by use of additional information remain personal data. ↗

artificial intelligence

The field of science and engineering concerned with creating the methods, mechanisms and applications that allow engineered systems to carry out tasks — producing content, predictions, recommendations or decisions — in pursuit of objectives set by people. The discipline is strongly interdisciplinary, drawing on computer science, data science, mathematics, the natural sciences and the humanities. ↗

attention mechanism

A neural-network operation that computes an output as a weighted combination of value vectors, with each weight derived from the compatibility between a query vector and a corresponding key vector. Attention lets a model weigh the relevance of every element of an input sequence when processing any given element; in its scaled dot-product, self-attention form it is the core operation of the transformer architecture. ↗

authorised representative

Under the EU AI Act, a natural or legal person located or established in the Union who has received and accepted a written mandate from a provider of an AI system or a general-purpose AI model to perform and carry out on its behalf the obligations and procedures established by the Regulation. ↗

automated decision-making

Decision-making produced by technological means without human involvement in the individual decision. EU law frames the concept in Article 22(1) of the General Data Protection Regulation, which provides: "The data subject shall have the right not to be subject to a decision based solely on automated processing, including profiling, which produces legal effects concerning him or her or similarly significantly affects him or her." ↗

benchmark

A standardized task set, dataset, or test procedure against which the performance of AI models or systems is measured, permitting comparison of systems with one another or with a fixed reference. In machine-learning practice, results on shared benchmarks provide a common basis for reporting and comparing model capabilities. ↗

benchmark contamination

The presence of a benchmark's test or development data in the corpus used to train a model, so that measured scores partly reflect memorization of test items rather than generalization to novel inputs. Contamination inflates reported benchmark performance and is a recognized methodological concern for models trained on large web-scale corpora. ↗

bias (in AI systems)

A systematic distortion in an AI system's data, design, or outputs that skews results away from representativeness, as distinct from random error. NIST characterises AI bias in three categories — systemic bias, statistical and computational bias, and human bias — arising respectively from institutional practices, from unrepresentative data and algorithmic error, and from human perception and judgment across the AI lifecycle. ↗

capability evaluation

An assessment of what an advanced AI system is able to do, conducted through techniques including automated capability assessments, red-team exercises by domain experts, human-uplift studies, and evaluations of AI agents that plan and act semi-autonomously. Capability evaluations measure performance in defined risk domains — in the originating UK programme: misuse spanning chemical, biological, and cyber capabilities, societal impacts, autonomous-system behaviour, and the strength of safeguards. ↗

capability threshold

A defined level of AI capability that, if reached or approached by a model, triggers predetermined obligations under a developer's safety framework, such as heightened security and deployment mitigations or a requirement to demonstrate that risk remains acceptable before further scaling or release. Published safety frameworks pair capability thresholds with evaluations or affirmative safety analyses that assess how far a model remains from each threshold. ↗

CE marking

A marking by which a provider indicates that an AI system conforms with the requirements of Chapter III, Section 2 of the EU AI Act and with other applicable Union harmonisation legislation providing for its affixing. Article 48 subjects it to the general principles set out in Article 30 of Regulation (EC) No 765/2008, requires it to be affixed visibly, legibly, and indelibly to high-risk AI systems (or to packaging or accompanying documentation where that is not possible), and requires that, for high-risk AI systems provided digitally, a digital CE marking be used only where it can easily be accessed via the interface from which the system is accessed or via a machine-readable code or other electronic means. ↗

certification

Attestation by an independent third party that an object of conformity assessment — such as a product, process, service, system, person, or body — fulfils specified requirements; the term excludes accreditation, which attests to the competence of conformity assessment bodies themselves. For AI, certification of AI management systems against ISO/IEC 42001:2023 is carried out by bodies meeting ISO/IEC 42006:2025. ↗

conformity assessment

In the EU AI Act, the process of demonstrating whether the requirements of Chapter III, Section 2 — covering risk management, data governance, technical documentation, record-keeping, transparency, human oversight, and accuracy, robustness, and cybersecurity — have been fulfilled for a high-risk AI system. Article 43 determines the applicable procedure: internal control by the provider (Annex VI) for most Annex III systems, or assessment of the quality management system and technical documentation with the involvement of a notified body (Annex VII) in specified cases, while systems covered by the Union harmonisation legislation listed in Section A of Annex I follow the conformity assessment procedure of the relevant sectoral act, with the Act's requirements forming part of that assessment. ↗

content provenance (Content Credentials)

The record of a digital asset's origin and history, including its interactions with the actors, tools and other assets involved in its creation and editing. The C2PA technical specification — deployed under the name Content Credentials — represents provenance as cryptographically signed, tamper-evident manifests bound to the asset, each combining assertions about the asset's creation and modification with a signed claim that can be validated by recipients. ↗

context window

The bounded span of content — measured in tokens — that a language model can take into account at one time when generating output, comprising the current input, prior conversation turns, and the text generated so far. Material outside the context window is not available to the model during generation, a run-time limit distinct from the knowledge acquired during training. ↗

dangerous-capability evaluation

An evaluation designed to determine whether an AI model possesses capabilities that could enable severe harm — examples treated in the originating paper include offensive cyber operations, persuasion and manipulation, deception, weapons acquisition, and self-proliferation — so that developers, policymakers, and other stakeholders can make informed decisions about model training, deployment, and security. It is distinguished from alignment evaluations, which assess a model's propensity to apply its capabilities harmfully. ↗

data poisoning

An adversarial attack carried out during the training stage of a machine-learning system, in which the attacker manipulates training data or its labels so that trained models acquire degraded performance, attacker-chosen behaviours, or hidden backdoor vulnerabilities activated later by specific inputs. Poisoning applies to predictive and generative AI and can be introduced at any stage of the training pipeline, including web-scraped pre-training corpora, instruction tuning, and reinforcement learning from human feedback. ↗

deep learning

A branch of machine learning in which neural networks with many intermediate ("hidden") layers are trained so that successive layers form increasingly abstract, hierarchical representations of the input. The approach is also termed deep neural network learning. ↗

deepfake

Under the EU Artificial Intelligence Act, 'deep fake' means 'AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful'. Deployers of AI systems that generate or manipulate such content must disclose that it has been artificially generated or manipulated; this obligation does not apply where the use is authorised by law to detect, prevent, investigate or prosecute criminal offences, and where the content forms part of an evidently artistic, creative, satirical, fictional or analogous work it is limited to disclosing the existence of the generated or manipulated content in a manner that does not hamper display or enjoyment of the work. ↗

deployer

Under the EU AI Act, a natural or legal person, public authority, agency or other body using an AI system under its authority, except where the AI system is used in the course of a personal non-professional activity. ↗

diffusion model

A class of latent-variable generative models consisting of three major components: a forward process, a reverse process, and a sampling procedure. The model's objective is to learn a diffusion process that generates the probability distribution of a given dataset; diffusion models are widely used in computer vision tasks including image denoising, inpainting, super-resolution, and image generation. ↗

direct preference optimization (DPO)

A preference-alignment method that trains a language model directly on human preference comparisons using a simple classification-style loss, removing the separate reward-model fitting and reinforcement-learning stages used in RLHF. The originating paper reports that DPO matches or exceeds PPO-based RLHF on the alignment tasks evaluated while being more stable, computationally lighter, and simpler to implement and train. ↗

distributor

Under the EU AI Act, a natural or legal person in the supply chain, other than the provider or the importer, that makes an AI system available on the Union market. ↗

embedding

A representation of a discrete item — such as a token, word, or document — as a vector of real numbers in a continuous space, learned so that geometric proximity reflects similarity of meaning or function. Embeddings are the form in which neural language models internally process tokens, and they are used independently for search, clustering, recommendation, and retrieval. ↗

EU declaration of conformity

A written, machine-readable, physical or electronically signed declaration that the provider must draw up for each high-risk AI system, stating that the system meets the requirements of Chapter III, Section 2 of the EU AI Act and containing the information set out in Annex V. The provider keeps it at the disposal of national competent authorities for 10 years after the system is placed on the market or put into service, and by drawing it up assumes responsibility for the system's compliance with those requirements. ↗

evaluation (eval)

An assessment that establishes how well an AI model or system satisfies defined criteria. In current practice, "eval" denotes a structured test — a defined task set, protocol, and scoring method — run against a model to measure a particular behavior, capability, or risk property. ↗

explainability

The capacity to provide a representation of the mechanisms underlying an AI system's operation. The NIST AI Risk Management Framework distinguishes explainability from interpretability, which in that framework concerns the meaning of a system's outputs in the context of its designed functional purposes — a framework-specific usage that differs from the inner-workings sense recorded under 'interpretability' in this registry; together the two characteristics assist users, operators and overseers in understanding a system's functionality, and explainable systems can be debugged and monitored more easily and lend themselves to more thorough documentation, audit and governance. ↗

fairness

A characteristic of trustworthy AI concerned with equality and equity, addressed in practice by managing harmful bias and discrimination. The NIST AI Risk Management Framework notes that standards of fairness are complex and difficult to define because perceptions of fairness differ among cultures and may shift depending on application, and that a system whose harmful biases have been mitigated is not necessarily fair. ↗

fine-tuning

The process of adapting a pretrained model to perform specific tasks or to specialize in a particular domain by further training it on task-specific data. It follows the initial pretraining phase and is often carried out as supervised learning. ↗

frontier safety framework

A published set of protocols through which a developer of frontier AI — highly capable general-purpose AI models or systems — identifies, assesses, and manages severe risks across the AI lifecycle, including thresholds at which risks would be deemed intolerable unless adequately mitigated, the mitigations attached to those thresholds, and associated governance and disclosure processes. Publication of such frameworks was a commitment made by signatory organisations under the Frontier AI Safety Commitments agreed at the AI Seoul Summit in May 2024. ↗

fundamental rights impact assessment

An assessment, required by Article 27 of the EU AI Act, of the impact on fundamental rights that the use of a high-risk AI system may produce, performed before the system is first deployed. It applies to deployers that are bodies governed by public law or private entities providing public services, and to deployers of high-risk systems for creditworthiness evaluation and credit scoring or for risk assessment and pricing in life and health insurance (Annex III, points 5(b) and (c)); high-risk systems intended to be used in the critical-infrastructure area listed in Annex III, point 2, are excluded from the obligation. The assessment covers the deployer's processes in which the system will be used, the period and frequency of use, the categories of persons likely to be affected, the specific risks of harm, the human oversight measures, and the measures to be taken if risks materialise, and the deployer must notify the market surveillance authority of the results. ↗

general-purpose AI model

Under the EU Artificial Intelligence Act, "an AI model, including where such an AI model is trained with a large amount of data using self-supervision at scale, that displays significant generality and is capable of competently performing a wide range of distinct tasks regardless of the way the model is placed on the market and that can be integrated into a variety of downstream systems or applications", excluding AI models used for research, development or prototyping activities before they are placed on the market. Provider obligations for these models are set out in Chapter V of the Regulation and have applied since 2 August 2025 (Art. 113(b)). ↗

grounding

Techniques that connect a generative AI system's outputs, or the data supplied to it, to identified and verifiable sources of information, so that generated content can be checked against source material. NIST AI 600-1 lists grounding alongside fine-tuning and retrieval-augmented generation among controls applied to data sources, information integrity, and third-party generative resources. ↗

hallucination

The production of confidently stated but erroneous or false content by a generative AI system, including outputs that diverge from the supplied input or contradict earlier statements in the same context. NIST AI 600-1 terms this phenomenon "confabulation" and describes it as a natural result of generative models producing outputs that approximate the statistical distribution of their training data. ↗

harmonised standard

'A European standard adopted on the basis of a request made by the Commission for the application of Union harmonisation legislation.' Under Article 40(1) of the EU AI Act, high-risk AI systems and general-purpose AI models that conform to harmonised standards, or parts of them, whose references are published in the Official Journal of the European Union are presumed to conform with the corresponding requirements or obligations of that Regulation, to the extent the standards cover them. ↗

high-risk AI system

An AI system subject to the requirements of Chapter III of the EU AI Act — including risk management, technical documentation, human oversight, and conformity assessment — because of its potential effect on health, safety, or fundamental rights. A system is classified as high-risk either because it is a safety component of a product (or is itself a product) covered by the Union harmonisation legislation listed in Annex I and required to undergo third-party conformity assessment, or because it falls within a use case listed in Annex III, such as biometrics, education, employment, essential services, law enforcement, migration, or the administration of justice. An Annex III system that does not pose a significant risk of harm to the health, safety, or fundamental rights of natural persons, including by not materially influencing the outcome of decision-making, is exempt from the classification under conditions set out in Article 6(3), except where the system performs profiling of natural persons. Under Regulation (EU) 2024/1689, as amended by Regulation (EU) 2026/1744, the requirements and obligations in Chapter III, Sections 1, 2 and 3 apply from 2 December 2027 to systems classified as high-risk under Article 6(2) and Annex III, and from 2 August 2028 to systems classified as high-risk under Article 6(1) and Annex I; Article 6(5) and Chapter III, Section 5 — harmonised standards, conformity assessment, certificates and registration — apply from 2 August 2026. ↗

holistic evaluation

An approach to language-model evaluation that measures models across a broad, explicitly taxonomized set of scenarios and along multiple metrics simultaneously — in the originating HELM framework: accuracy, calibration, robustness, fairness, bias, toxicity, and efficiency — so that trade-offs between properties are exposed rather than collapsed into a single headline score. ↗

human oversight (AI Act Art. 14 duty)

The EU AI Act requirement that high-risk AI systems be designed and developed, including with appropriate human-machine interface tools, so that natural persons can effectively oversee them during use, with the aim of preventing or minimising risks to health, safety, or fundamental rights. Oversight measures must be commensurate with the system's risks, level of autonomy, and context of use, and must enable the persons assigned to oversight to understand the system's capacities and limitations, remain aware of automation bias, correctly interpret output, decide not to use the system or to disregard, override, or reverse its output, and intervene in or halt its operation in a safe state. ↗

hyperparameter

A configuration value of a machine learning algorithm that shapes the learning process itself and is chosen before training rather than learned from data — for example, the learning rate, the number and width of network layers, or the choice of optimization method. Hyperparameters are commonly tuned against validation data, in contrast to parameters, which are fitted during training. ↗

importer

Under the EU AI Act, a natural or legal person located or established in the Union that places on the market an AI system bearing the name or trademark of a natural or legal person established in a third country. ↗

inference

Reasoning that moves from available premises — which in AI may be facts, rules, models, features or raw data — to conclusions; the term covers both the reasoning process and its outcome. In machine learning usage it also denotes the operational phase in which a trained model is run on new inputs to produce outputs, in contrast to the training phase. ↗

instruction tuning

A fine-tuning procedure in which a language model is further trained on a collection of tasks described via natural-language instructions. The originating study found that this procedure substantially improves zero-shot performance on unseen tasks, enabling the model to carry out tasks it was not trained on when they are presented as instructions. ↗

intended purpose

Under the EU AI Act, the use for which an AI system is intended by the provider, including the specific context and conditions of use, as specified in the information supplied by the provider in the instructions for use, promotional or sales materials and statements, and in the technical documentation. ↗

interpretability

The extent to which humans can understand the inner workings of an AI model, including why it generated a particular output or decision. The International AI Safety Report 2026 groups interpretability with safety-verification tools intended to give more rigorous assurance that models have specific safety-related properties, while noting that current methods rest on simplifying assumptions. ↗

jailbreak

A direct prompting attack intended to circumvent restrictions placed on a model's outputs, such as overriding refusal behaviour so that the model produces content its safeguards are designed to withhold. Jailbreak prompts may be crafted manually or generated automatically, and some transfer across models. ↗

knowledge distillation

A model-compression technique in which the knowledge held by a large model, or by an ensemble of models, is transferred into a single smaller model by training the smaller model to reproduce the larger system's predictive behavior. The resulting model retains much of the original system's capability in a form that is simpler and cheaper to deploy. ↗

leaderboard

A public ranking of AI models by their scores on a benchmark or by aggregated comparative judgments, updated as new models or results are submitted. A leaderboard's ordering is specific to the evaluation protocol that produces it, whether automated scoring against fixed test sets or statistical ranking of crowdsourced pairwise human-preference votes. ↗

LLM-as-a-judge

An evaluation method in which a strong language model scores or compares the outputs of other models, approximating human preference judgments at scale. Reported agreement with human raters exceeded 80 percent in the originating study, subject to documented judge biases including position, verbosity, and self-enhancement effects. ↗

machine learning

The process by which a model acquires its behaviour from data or experience: a learning algorithm adjusts the model's internal parameters so that the model's outputs come to reflect the patterns present in the material used to train it. Principal families include supervised, unsupervised, semi-supervised and reinforcement learning. ↗

making available on the market

Under the EU AI Act, the supply of an AI system or a general-purpose AI model for distribution or use on the Union market in the course of a commercial activity, whether in return for payment or free of charge. ↗

market surveillance

In European Union law, “the activities carried out and measures taken by market surveillance authorities to ensure that products comply with the requirements set out in the applicable Union harmonisation legislation and to ensure protection of the public interest covered by that legislation” (Regulation (EU) 2019/1020, Art. 3(3)). The EU AI Act applies that regime to AI: Art. 74(1) provides that Regulation (EU) 2019/1020 shall apply to AI systems covered by the Act, and Art. 70(1) requires each Member State to designate at least one market surveillance authority. It is distinct from post-market monitoring, which Art. 72 places on the provider: market surveillance is the public authority's function over systems already on the market. ↗

mechanistic interpretability

A branch of interpretability research that aims to understand the computational mechanisms underlying a neural network's capabilities — identifying the internal features and circuits a model has learned and how they combine to produce its behaviour — in pursuit of concrete scientific and engineering goals such as monitoring, predicting, and modifying model behaviour. ↗

mixture-of-experts (MoE)

A neural-network architecture in which the model selects different subsets of its parameters — the "experts" — for each input, rather than applying the same parameters to every input. This produces a sparsely activated model whose total parameter count can grow very large while the computation performed per input remains roughly constant. ↗

model (AI/ML model)

A representation, in mathematical or other logical form, of a system, entity, process, phenomenon or body of data; in machine learning, the trained construct that produces an inference or prediction when supplied with new input. The model is the artefact that training yields and is the component an AI system integrates and runs in operation. ↗

model card

A short document released alongside a trained machine-learning model that states the contexts in which the model is intended to be used, the procedures by which it was evaluated, and its benchmarked performance across relevant conditions, including different cultural, demographic, or phenotypic groups. Proposed by Mitchell et al. (2018) as a standard form of transparent model reporting. ↗

model extraction

A privacy attack in which an adversary queries a deployed machine-learning model and uses the returned outputs to reconstruct information about the model's architecture and parameters. Exact extraction of a model has been shown to be impossible, and even reconstruction of a functionally equivalent model is computationally prohibitive in the general case; practical attacks therefore aim for a model that approximates the original's performance, threatening the confidentiality of proprietary models exposed through inference interfaces. ↗

multimodal model

A model that processes and relates information from multiple sensory modalities, each a primary human channel of communication and sensation, such as vision and touch. In current generative systems, multimodal models accept or produce combinations of text, images, audio, and video within a single model. ↗

national competent authority

Under the EU AI Act, “a notifying authority or a market surveillance authority”; for AI systems put into service or used by Union institutions, agencies, offices and bodies, references to national competent authorities are construed as references to the European Data Protection Supervisor (Art. 3(48)). Art. 70(1) requires each Member State to establish or designate at least one notifying authority and at least one market surveillance authority, which shall exercise their powers “independently, impartially and without bias”. The two roles differ: a notifying authority assesses and designates the bodies that carry out conformity assessment; a market surveillance authority supervises systems already on the market. ↗

natural person

In European Union law, a human being, as distinguished from a “legal person” such as a company or other body. Neither the EU AI Act nor the GDPR defines the term itself; both operate on the distinction. The GDPR anchors it: personal data means “any information relating to an identified or identifiable natural person” (Art. 4(1)), and Recital 14 states the Regulation “does not cover the processing of personal data which concerns legal persons”. The AI Act uses the term throughout — defining provider and deployer as “a natural or legal person, public authority, agency or other body” (Art. 3(3)–(4)), and requiring that AI systems intended to interact directly with natural persons inform them they are interacting with an AI system (Art. 50(1)). ↗

neural network

A computational structure of simple processing units ("neurons") organised in one or more layers and linked by connections whose numerical weights can be adjusted, enabling the structure to map input data to outputs and to be trained. The design was originally inspired by biological neurons, but that inspiration plays little role in most current work on neural networks. ↗

notified body

Under the EU AI Act, a conformity assessment body — an organisation performing third-party conformity assessment activities, including testing, certification, and inspection — that has been notified, that is, officially designated, in accordance with the Regulation and other relevant Union harmonisation legislation. Notified bodies carry out the third-party conformity assessment procedures required for certain high-risk AI systems, assessing quality management systems and technical documentation under Annex VII. ↗

open-weight model

A machine-learning model whose trained parameters (weights) are published for download and local use, with or without licence restrictions on use, modification or redistribution. Releasing weights alone does not satisfy the Open Source Initiative's Open Source AI Definition, which additionally requires sufficiently detailed information about the training data, the complete source code used to train and run the system, and terms granting the freedoms to use, study, modify and share it. ↗

operator

Under the EU AI Act, the collective term for a provider, product manufacturer, deployer, authorised representative, importer or distributor. The Regulation uses it where a provision applies across these actor categories rather than to a single role. ↗

parameter (weights)

An internal numerical variable of a model whose value is learned from data during training and which determines how the model transforms inputs into outputs; examples include the weights of a neural network and the transition probabilities of a Markov model. Model scale is customarily stated as a parameter count, and parameters are distinct from hyperparameters, which are set before training rather than learned. ↗

parameter-efficient fine-tuning (PEFT)

A family of adaptation methods that specialize a pre-trained model by training only a small number of added or selected parameters while leaving most of the original weights unchanged. LoRA (Low-Rank Adaptation), a widely used method in this family, freezes the pre-trained model weights and adds compact trainable low-rank update matrices to each Transformer layer; compared with full fine-tuning of GPT-3 175B, the originating paper reports roughly 10,000 times fewer trainable parameters, a threefold reduction in GPU memory requirement, and no additional inference latency. ↗

personal data

Under the EU General Data Protection Regulation, 'personal data' means 'any information relating to an identified or identifiable natural person ("data subject")'. An identifiable natural person is one who can be identified, directly or indirectly, in particular by reference to an identifier such as a name, an identification number, location data, an online identifier, or to one or more factors specific to the physical, physiological, genetic, mental, economic, cultural or social identity of that natural person. ↗

placing on the market

Under the EU AI Act, the first making available of an AI system or a general-purpose AI model on the Union market. ↗

post-market monitoring

In the EU AI Act, all activities carried out by providers of AI systems to collect and review experience gained from the use of AI systems they place on the market or put into service, for the purpose of identifying any need to immediately apply necessary corrective or preventive actions. Article 72 requires providers of high-risk AI systems to establish a documented post-market monitoring system, based on a post-market monitoring plan, that actively and systematically collects and analyses data on system performance throughout the system's lifetime and supports evaluation of continuous compliance with Chapter III, Section 2. ↗

presumption of conformity

Under the EU AI Act, the legal effect by which conformity with a published harmonised standard is treated as conformity with the Act's own requirements. Art. 40(1) provides that high-risk AI systems or general-purpose AI models in conformity with harmonised standards whose references are published in the Official Journal “shall be presumed to be in conformity with the requirements” of the relevant Chapter, “to the extent that those standards cover those requirements or obligations”. Art. 41(3) extends the same effect to Commission-adopted common specifications. The presumption is bounded by coverage — it reaches only what the standard actually addresses — and is not conclusive: under Art. 82(1) a market surveillance authority that finds a compliant high-risk system nevertheless presents a risk shall require corrective measures. ↗

pretraining

A component of the training stage in which a model learns general patterns, features, and relationships from vast amounts of unlabeled data, such as through self-supervised learning. Pretraining can equip a model with knowledge of general features or patterns useful in downstream tasks and can be followed by additional training or fine-tuning that specializes the model for a specific downstream task. ↗

profiling

Under the EU General Data Protection Regulation, 'profiling' means any form of automated processing of personal data consisting of the use of personal data to evaluate certain personal aspects relating to a natural person — in particular to analyse or predict aspects concerning that natural person's performance at work, economic situation, health, personal preferences, interests, reliability, behaviour, location or movements. ↗

prohibited AI practices

AI practices banned outright in the European Union under Article 5(1) of the EU AI Act. The list covers subliminal, purposefully manipulative, or deceptive techniques that materially distort behaviour and cause or are reasonably likely to cause significant harm; exploitation of vulnerabilities linked to age, disability, or a specific social or economic situation; social scoring producing detrimental or unjustified treatment; predicting the risk of a person committing a criminal offence based solely on profiling or personality traits; untargeted scraping of facial images from the internet or CCTV to build facial recognition databases; emotion inference in workplaces and education institutions, except for medical or safety reasons; biometric categorisation to deduce sensitive attributes such as race, political opinions, or sexual orientation; and 'real-time' remote biometric identification in publicly accessible spaces for law enforcement, except where strictly necessary for exhaustively listed objectives. Article 5 has applied since 2 February 2025. Regulation (EU) 2026/1744 inserts points (ba) and (bb) into Article 5(1) and adds Article 5(1a) and (1b), which apply from 2 December 2026. ↗

prompt

The input submitted to a generative AI model to elicit an output, typically expressed in natural language. In prompt-based operation, a task is reformulated as a textual prompt for the model to complete, allowing a single pre-trained model to be adapted to many tasks with few or no task-specific training examples. ↗

prompt injection

An attack on an application built on a large language model in which crafted input alters the model's behaviour or output in unintended ways, exploiting the model's processing of instructions and data through the same channel. In direct prompt injection the adversarial input is supplied through the attacker's own prompt; in indirect prompt injection it is placed in external content — such as a web page, document, or retrieved data — that the system later processes. ↗

provider

Under the EU AI Act, a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model, or that has one developed, and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. ↗

pseudonymisation

Under the EU General Data Protection Regulation, 'pseudonymisation' means the processing of personal data in such a manner that the data can no longer be attributed to a specific data subject without the use of additional information, provided that such additional information is kept separately and is subject to technical and organisational measures ensuring non-attribution to an identified or identifiable natural person. Pseudonymised data remain personal data within the scope of the Regulation. ↗

putting into service

Under the EU AI Act, the supply of an AI system for first use directly to the deployer, or for own use, in the Union for its intended purpose. ↗

quantization

A model-compression technique that reduces the numerical precision used to represent a model's parameters — for example from 16-bit floating point to 4-bit or 3-bit representations — lowering memory footprint and inference cost. Post-training quantization methods apply this compression to an already-trained model; the GPTQ method quantizes models with 175 billion parameters to 3 or 4 bits per weight in approximately four GPU hours with negligible loss of accuracy relative to the uncompressed model. ↗

reasonably foreseeable misuse

Under the EU AI Act, the use of an AI system in a way that is not in accordance with its intended purpose, but which may result from reasonably foreseeable human behaviour or interaction with other systems, including other AI systems. ↗

red-teaming

In the AI context, "a structured testing exercise used to probe an AI system to find flaws and vulnerabilities such as inaccurate, harmful, or discriminatory outputs, often in a controlled environment and in collaboration with system developers" (NIST AI 600-1). Red-teaming may be performed before or after a model or system is made publicly available and may involve expert, general-public, or model-assisted teams. ↗

reinforcement learning from human feedback (RLHF)

A training method that aligns a model's behavior with human judgments. Human annotators supply demonstrations of desired outputs, on which the model is first fine-tuned with supervised learning, and rankings of alternative model outputs, which are then used as the preference signal for a further reinforcement-learning stage that optimizes the model toward the ranked human preferences. ↗

retrieval-augmented generation (RAG)

A system design in which a generative model is paired with an information-retrieval component that fetches relevant material from an external knowledge source at run time and supplies it to the model alongside the query. Introduced by Lewis et al. (2020) as the combination of a model's learned parametric memory with a non-parametric memory of retrieved documents, the approach allows the knowledge a system can draw on to be updated without retraining the model. ↗

scaling laws

Empirical regularities under which language-model performance, measured as cross-entropy loss, improves as a power law of model size, dataset size, and training compute, with some trends spanning more than seven orders of magnitude. Later work on compute-optimal training found that, for a fixed compute budget, model size and the number of training tokens should be scaled in roughly equal proportion. ↗

serious incident

Under the EU AI Act, an incident or malfunctioning of an AI system that directly or indirectly leads to any of the following: the death of a person or serious harm to a person's health; a serious and irreversible disruption of the management or operation of critical infrastructure; the infringement of obligations under Union law intended to protect fundamental rights; or serious harm to property or the environment. Providers of high-risk AI systems must report serious incidents to the market surveillance authorities of the Member States where the incident occurred — in any event no later than 15 days after becoming aware, shortened to 10 days where a person has died and to two days for a widespread infringement or a critical-infrastructure disruption. ↗

substantial modification

Under the EU AI Act, a change to an AI system after its placing on the market or putting into service that was not foreseen or planned in the provider's initial conformity assessment and that either affects the system's compliance with the requirements for high-risk AI systems in Chapter III, Section 2, or modifies the intended purpose for which the system was assessed. ↗

synthetic data

Data generated by algorithms or models rather than collected from real persons or events, typically constructed to reproduce the structure and statistical properties of an original dataset without containing records of real individuals. The EU Artificial Intelligence Act treats it as an alternative to personal data: providers of high-risk AI systems may exceptionally process special categories of personal data for bias detection and correction only where that purpose cannot be effectively fulfilled by processing other data, including synthetic or anonymised data. ↗

synthetic media

Content — including images, video, audio and text — that has been significantly altered or generated by algorithms, including by artificial intelligence. NIST guidance on digital content transparency uses the equivalent term 'synthetic content' and addresses its risks through provenance data tracking — implemented by digital watermarking and by metadata recording — and through synthetic content detection. ↗

system card

A published document that reports on an AI system at the system level rather than the model level, covering the deployed system together with non-model safeguards such as use policies, access controls, and abuse monitoring, and describing identified safety challenges and the interventions adopted before release. OpenAI's GPT-4 System Card (2023) applies the format to a large-language-model deployment. ↗

system prompt

Application-specific instructions provided in-context to a generative AI system by the model developer or application designer, typically prepended to other input, and which may be treated as higher-trust than other forms of input. A system prompt commonly establishes the model's role, constraints, and operating rules for a given application. ↗

systemic risk (general-purpose AI)

Under the EU AI Act, a risk specific to the high-impact capabilities of general-purpose AI models: one that has a significant impact on the Union market because of the models' reach, or because of actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or society as a whole, and that can be propagated at scale across the value chain. A general-purpose AI model is classified as presenting systemic risk under Article 51, including a presumption of high-impact capabilities where cumulative training compute exceeds 10^25 floating-point operations; classified providers carry additional obligations under Article 55, such as model evaluation, adversarial testing, systemic-risk mitigation, incident reporting, and cybersecurity protection. ↗

technical documentation (Art. 11 / Annex IV)

Documentation that a provider must draw up before a high-risk AI system is placed on the market or put into service, and keep up to date, demonstrating that the system complies with the requirements of Chapter III, Section 2 of the EU AI Act and providing national competent authorities and notified bodies with the information necessary to assess that compliance in a clear and comprehensive form. It must contain at minimum the elements set out in Annex IV, including a general description of the AI system, its development process and design specifications, monitoring and control capabilities, risk management, and a description of relevant changes over the system's lifecycle; SMEs, including start-ups, may supply the Annex IV elements in a simplified manner. ↗

testing data

Data used for providing an independent evaluation of the AI system in order to confirm the expected performance of that system before its placing on the market or putting into service, as defined in Article 3(32) of the EU Artificial Intelligence Act (Regulation (EU) 2024/1689). Testing data is held apart from both training data and validation data so that the evaluation is not influenced by the data used to build or tune the model. ↗

text and data mining

Under Directive (EU) 2019/790 on copyright in the Digital Single Market, 'text and data mining' means 'any automated analytical technique aimed at analysing text and data in digital form in order to generate information which includes but is not limited to patterns, trends and correlations'. Articles 3 and 4 of the Directive establish exceptions to copyright and related rights for such mining — one for scientific research by research organisations and cultural heritage institutions, and a general exception or limitation that applies unless rightsholders have expressly reserved their rights. ↗

token

The atomic unit of text on which a language model operates: a short character sequence — typically a word, subword fragment, punctuation mark, or symbol — drawn from the model's fixed vocabulary. Modern language models represent text as sequences of subword units so that rare and novel words can be composed from a bounded vocabulary; training-data scale and context-window length are customarily measured in tokens. ↗

tokenization

The process of segmenting text into tokens and mapping each token to an identifier in a model's fixed vocabulary, performed before training or inference and reversed when model output is converted back to text. Contemporary systems predominantly use subword tokenization methods such as byte-pair encoding, which build a vocabulary of frequent character sequences so that any input, including rare or unseen words, can be represented. ↗

tool use (function calling)

A mode of operation in which a language model issues structured calls to external software — such as search engines, calculators, or translation systems — during generation and incorporates the returned results into its subsequent output. Schick et al. (2023) demonstrated language models learning to decide which external interfaces to call, when to call them, and what arguments to pass. ↗

training (model training)

The stage of machine learning in which a learning algorithm processes training data to set and progressively refine a model's parameters; its product is a trained model. Training stands in contrast to inference, the stage in which the trained model is applied to new inputs. ↗

training compute (FLOP)

The cumulative amount of computation used to train a model, conventionally measured in floating-point operations (FLOP). Regulation (EU) 2024/1689 defines a "floating-point operation" as "any mathematical operation or assignment involving floating-point numbers, which are a subset of the real numbers typically represented on computers by an integer of fixed precision scaled by an integer exponent of a fixed base" (Art. 3(67)), and presumes a general-purpose AI model to have the high-impact capabilities that classify it as a general-purpose AI model with systemic risk under Art. 51(1), point (a), when the cumulative amount of computation used for its training, measured in floating-point operations, is greater than 10^25 (Art. 51(2)). ↗

training data

Data used for training an AI system through fitting its learnable parameters, as defined in Article 3(29) of the EU Artificial Intelligence Act (Regulation (EU) 2024/1689). In machine learning practice, training data is the portion of a dataset from which a model's parameters are estimated, as distinct from validation data and testing data. ↗

transformer (architecture)

A neural-network architecture, introduced in 2017, that processes sequences using attention mechanisms alone, dispensing with the recurrence and convolution of earlier sequence models. Stacked self-attention and feed-forward layers relate every position in a sequence to every other position and can be computed largely in parallel, properties that made the architecture the basis of contemporary large language models. ↗

trustworthy AI

AI systems exhibiting the characteristics that the NIST AI Risk Management Framework identifies as constituting trustworthiness: 'valid and reliable, safe, secure and resilient, accountable and transparent, explainable and interpretable, privacy-enhanced, and fair with harmful bias managed'. NIST AI 100-1 treats validity and reliability as a necessary condition of trustworthiness and notes that the characteristics must be balanced against one another according to the AI system's context of use. ↗

validation data

Data used for providing an evaluation of the trained AI system and for tuning its non-learnable parameters and its learning process in order, inter alia, to prevent underfitting or overfitting, as defined in Article 3(30) of the EU Artificial Intelligence Act (Regulation (EU) 2024/1689). The related term 'validation data set' denotes a separate data set or part of the training data set, either as a fixed or variable split (Article 3(31)). ↗

watermarking (AI content)

The embedding of identifying information directly into AI-generated content — in the pixels, words or audio signal itself rather than in a separate metadata channel — so the content can be recognised as artificially generated. Under Article 50(2) of the EU Artificial Intelligence Act, providers of AI systems that generate synthetic audio, image, video or text content must ensure the outputs are marked in a machine-readable format and detectable as artificially generated or manipulated, with technical solutions that are effective, interoperable, robust and reliable as far as technically feasible. ↗