AI Dictionary

June 2023

A shared understanding of the key AI concepts and terms is essential for the advancement of AI governance and negotiations.

This AI Dictionary is based on the agreed terminology of the expert group of the Joint Roadmap on Evaluation and Measurement Tools for Trustworthy AI and Risk Management, which issued a list of 65 key AI terms essential to understanding risk-based approaches to AI, along with their U.S. and EU interpretations and shared U.S.-EU definitions

You can consult defined terms and a list of terms to be defined.

Defined Terms

Term	Definition
Adversarial machine learning (adversarial attack)	A practice concerned with the design of ML algorithms that can resist security challenges, the study of the capabilities of attackers, and the understanding of attack consequences. Inputs in adversarial ML are purposely designed to make a mistake in its predictions despite resembling a valid input to a human.
Autonomy (autonomous AI system)	Systems that maintain a set of intelligence-based capabilities to respond to situations that were not pre-programmed or anticipated (i.e., decision-based responses) prior to system deployment. Autonomous systems have a degree of self government and self-directed behaviour (with the human’s proxy for decisions).
Big data	An all-encompassing term for large, complex digital data sets that need equally complex technological means to be stored, analysed, managed and processed with substantial computing power. Datasets are sometimes linked together to see how patterns in one domain affect other areas. Data can be structured into fixed fields or unstructured as free flowing information. The analysis of big datasets, often using AI, can reveal patterns, trends, or underlying relationships that were not previously apparent to researchers.
Classifier	A model that predicts (or assigns) class labels to data input.
Data poisoning	A type of security attack where malicious users inject false training data with the aim of corrupting the learned model, thus making the AI system learn something that it should not learn.
Deep learning	A subset of machine learning based on artificial neural networks that employs statistics to spot underlying trends or data patterns and applies that knowledge to other layers of analysis. Some have labelled this as a way to “learn by example” and as a technique that “perform[s] classification tasks directly from images, text, or sound” and then applies that knowledge independently.
Differential privacy	Differential privacy is a method for measuring how much information the output of a computation reveals about an individual. It produces data analysis outcomes that are nearly equally likely, whether any individual is, or is not, included in the dataset. Its goal is to obscure the presence or absence of any individual (in a database), or small groups of individuals, while at the same time preserving statistical utility.
Input data	Data provided to or directly acquired by an AI system on the basis of which the system produces an output.
Machine learning	Machine Learning is a branch of artificial intelligence (AI) and computer science which focuses on development of systems that are able to learn and adapt Without following explicit instructions imitating the way that humans learn, gradually improving its accuracy, by using algorithms and statistical models to analyse and draw inferences from patterns in data.
Model training	Process to establish or to improve the parameters of a machine learning model, based on a Machine Learning algorithm, by using training data.
Model validation	Confirmation through the provision of objective evidence, that the requirements for a specific intended use or application have been fulfilled.
Natural language processing	The ability of a machine to process, analyse, and mimic human language, either spoken or written.
Predictive analysis	The organisation of analyses of structured and unstructured data for inference and correlation that provides a useful predictive capability to new circumstances or data.
Profiling	‘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.
Reinforcement learning	A type of machine learning in which the algorithm learns by acting toward an abstract goal, such as “earn a high video game score” or “manage a factory efficiently.” During training, each effort is evaluated based on its contribution toward the goal.
Structured data	Data that has a predefined data model or is organised in a predefined way.
Unstructured data	Data that does not have a predefined data model or is not organised in a predefined way.
Synthetic data	Synthetic data is generated from data/processes and a model that is trained to reproduce the characteristics and structure of the original data aiming for similar distribution. The degree to which synthetic data is an accurate proxy for the original data is a measure of the utility of the method and the model.
Transfer learning	A technique in machine learning in which an algorithm learns to perform one task, such as recognizing cars, and builds on that knowledge when learning a different but related task, such as recognizing cats.
Supervised learning	Machine learning that makes use of labelled data during training.
Unsupervised learning	Machine learning that makes use of unlabelled data during training.
(AI) accuracy	Closeness of computations or estimates to the exact or true values that the statistics were intended to measure. The goal of an AI model is to learn patterns that generalise well for unseen data. It is important to check if a trained AI model is performing well on unseen examples that have not been used for training the model. To do this, the model is used to predict the answer on the test dataset and then the predicted target is compared to the actual answer. The concept of accuracy is used to evaluate the predictive capability of the AI model. Informally, accuracy is the fraction of predictions the model got right. A number of metrics are used in machine learning (ML) to measure the predictive accuracy of a model. The choice of the accuracy metric to be used depends on the ML task.
Test	Technical operation to determine one or more characteristics of or to evaluate the performance of a given product, material, equipment, organism, physical phenomenon, process or service according to a specified procedure.
Evaluation	Systematic determination of the extent to which an entity meets its specified criteria.
Verification	Provides evidence that the system or system element performs its intended functions and meets all performance requirements listed in the system performance specification.
Validation	Confirmation by examination and provision of objective evidence that the particular requirements for a specific intended use are fulfilled.
Test and Evaluation, Verification and Validation (TEVV)	A framework for assessing, incorporating methods and metrics to determine that a technology or system satisfactorily meets its design specifications and requirements, and that it is sufficient for its intended use.
Adaptive learning	An adaptive AI is a system that changes its behaviour while in use. Adaptation may entail a change in the weights of the model or a change in the internal structure of the model itself. The new behaviour of the adapted system may produce different results than the previous system for the same inputs.
Algorithm	An algorithm consists of a set of step-by-step instructions to solve a problem (e.g., not including data). The algorithm can be abstract and implemented in different programming languages and software libraries.
Classification	A classification system is a set of “boxes” into which things are sorted. Classifications are consistent, have unique classificatory principles, and are mutually exclusive. In AI design, when the output is one of a finite set of values (such as sunny, cloudy or rainy), the learning problem is called classification, and is called Boolean or binary classification if there are only two values.
Federated learning	Federated learning is a machine learning model which addresses the problem of data governance and privacy by training algorithms collaboratively without transferring the data to another location. Each federated device shares its local model parameters instead of sharing the whole dataset used to train it and the federated learning topology defines the way parameters are shared.
Generative adversarial network (GAN)	Generative Adversarial Networks, or GANs for short, are an approach to generative modelling using deep learning methods, such as convolutional neural networks. Generative modelling is an unsupervised learning task in machine learning that involves automatically discovering and learning the regularities or patterns in input data in such a way that the model can be used to generate or output new examples that plausibly could have been drawn from the original dataset.
Human values for AI	Values are idealised qualities or conditions in the world that people find good. AI systems are not value neutral. The incorporation of human values into AI systems requires that we identify whether, how and what we want AI to mean in our societies. It implies deciding on ethical principles, governance policies, incentives, and regulations. And it also implies that we are aware of differences in interests and aims behind AI systems developed by others according to other cultures and principles. The EU and U.S. are committed to the development of Trustworthy AI systems based on shared democratic values including the respect for the rule of law and human rights.
Human-centric AI	An approach to AI that prioritises human ethical responsibility, dynamic qualities, understanding and meaning. It encourages the empowerment of humans in design, use and implementation of AI systems. Human-Centric AI systems are built on the recognition of a meaningful human-technology interaction. They are designed as components of socio technical environments in which humans assume meaningful agency. Human-Centric AI is not designed as an end in itself, but as tools to serve people with the ultimate aim of increasing human and environmental well-being with respect for the rule of law, human rights, democratic values and sustainable development.
Language model	A language model is an approximative description that captures patterns and regularities present in natural language and is used for making assumptions on previously unseen language fragments.
Large language model (LLM)	A class of language models that use deep-learning algorithms and are trained on extremely large textual datasets that can be multiple terabytes in size. LLMs can be classed into two types: generative or discriminatory. Generative LLMs are models that output text, such as the answer to a question or even writing an essay on a specific topic. They are typically unsupervised or semi-supervised learning models that predict what the response is for a given task. Discriminatory LLMs are supervised learning models that usually focus on classifying text, such as determining whether a text was made by a human or AI
Model	A function that takes features as input and predicts labels as output. Typical phases of an AI model’s work flow are: Data collection and preparation, Model development, Model training, Model accuracy evaluation, Hyperparameters’ tuning, Model usage, Model maintenance, Model versioning.
Neural network	A computer system inspired by living brains, also known as artificial neural network, neural net, or deep neural net. It consists of two or more layers of neurons connected by weighted links with adjustable weights, which takes input data and produces an output. Whereas some neural networks are intended to simulate the functioning of biological neurons in the nervous system, most neural networks are used in artificial intelligence as realisations of the connectionist model.
Scalability	The ability to increase or decrease the computational resources required to execute a varying volume of tasks, processes, or services.
Socio-technical system	Technology is always part of society, just like society is always part of technology. This also means that one cannot understand one without the other. Technology is not only design and material appearance but also sociotechnical; that is, a complex process constituted by diverse social, political, economic, cultural and technological factors.
Technical interoperability	The ability of software or hardware systems or components to operate together successfully with minimal effort by an end user.
Value sensitive design (values-by design or ethics-by design)	A theoretically grounded approach to the design of technology that accounts for human values in a principled and systematic manner throughout the design process.
Auditability of an AI system	Auditability refers to the ability of an AI system to undergo the assessment of the system’s algorithms, data and design processes. This does not necessarily imply that information about business models and Intellectual Property related to the AI system must always be openly available. Ensuring traceability and logging mechanisms from the early design phase of the AI system can help enable the system's auditability.
Standard	Standards are a set of institutionalised agreed upon-rules for the production of (textual or material) objects. They are released by international organisations and ensure quality and safety and set product or services’ specifications. Standards are the result of negotiations among various stakeholders and are institutionalised and thus difficult to change.
Accessibility	Extent to which products, systems, services, environments and facilities can be used by people from a population with the widest range of user needs, characteristics and capabilities to achieve identified goals in identified contexts of use (which includes direct use or use supported by assistive technologies).
Accountability	Accountability relates to an allocated responsibility. The responsibility can be based on regulation or agreement or through assignment as part of delegation. In a systems context, accountability refers to systems and/or actions that can be traced uniquely to a given entity. In a governance context, accountability refers to the obligation of an individual or organisation to account for its activities, to complete a deliverable or task, to accept the responsibility for those activities, deliverables or tasks, and to disclose the results in a transparent manner.
AI (or algorithmic) bias	Harmful AI bias describes systematic and repeatable errors in AI systems that create unfair outcomes, such as placing. privileged groups at systematic advantage and unprivileged groups at systematic disadvantage. Different types of bias can emerge and interact due to many factors, including but not limited to, human or system decisions and processes across the AI lifecycle. Bias can be present in AI systems resulting from pre-existing cultural, social, or institutional expectations; because of technical limitations of their design; by being used in unanticipated contexts; or by non-representative design specifications.
Attack	Action targeting a learning system to cause malfunction.
Chatbot (conversational bot)	A computer program designed to simulate conversation with a human user, usually over the internet; especially one used to provide information or assistance to the user as part of an automated service.
Discrimination	Unequal treatment of a person based on belonging to a category rather than on individual merit. Discrimination can be a result of societal, institutional and implicitly held individual biases or attitudes that get captured in processes across the AI lifecycle, including by AI actors and organisations, or represented in the data underlying AI systems. Discrimination biases can also emerge due to technical limitations in hardware or software, or the use of an AI system that, due to its context of application, does not treat all groups equally. Discriminatory biases can also emerge in the very context in which the AI system is used. As many forms of biases are systemic and implicit, they are not easily controlled or mitigated and require specific governance and other similar approaches"
Evasion	In Evasion Attacks, the adversary solves a constrained optimization problem to find a small input perturbation that causes a large change in the loss function and results in output misclassification.
Fault tolerance	The ability of a system or component to continue normal operation despite the presence of hardware or software faults.
Feedback loop	Feedback loop describes the process of leveraging the output of an AI system and corresponding end-user actions in order to retrain and improve models over time. The AI generated output (predictions or recommendations) are compared against the final decision (for example, to perform work or not) and provides feedback to the model, allowing it to learn based on its results.
Harmful bias	Harmful bias can be either conscious or unconscious. Unconscious, also known as implicit bias, involves associations outside conscious awareness that lead to a negative evaluation of a person on the basis of characteristics such as race, gender, sexual orientation, or physical ability. Discrimination is behaviour; discriminatory actions perpetrated by individuals or institutions refer to inequitable treatment of members of certain social groups that results in social advantages or disadvantages. AI systems can reinforce harmful bias when trained on prejudiced or unrepresentative data. Most often harmful bias is unintended by developers and adopters of AI. AI actors can design AI systems to mitigate harmful bias.
Human rights impact assessment	An human rights impact assessment (HRIA) of AI identifies, understands and assesses the impact of the AI system on human rights, such as but not limited to, the right to privacy or non-discrimination. AI systems can pose risks to, as well as enhance, individual human rights.
Opacity	When AI system processes, functions, output or behaviour are unavailable or incomprehensible to all stakeholders – usually an antonym for transparency.
Red-team	A group of people authorised and organised to emulate a potential adversary’s attack or exploitation capabilities against an enterprise’s security posture. It is often used to help identify and address potential security vulnerabilities.
Reliability	An AI system is said to be reliable if it behaves as expected, even for novel inputs on which it has not been trained or tested earlier.
Resilience	The ability of an AI system to prepare for and adapt to changing conditions and withstand and recover rapidly from disruptions. Resilience includes the ability to withstand and recover from deliberate attacks, accidents, or naturally occurring threats or incidents. The ability of a system to adapt to and recover from adverse conditions.
Robustness (robust AI)	Robustness of an AI system encompasses both its technical robustness (ability of a system to maintain its level of performance under a variety of circumstances) as well as its robustness from a social perspective (ensuring that the AI system duly takes into account the context and environment in which the system operates). This is crucial to ensure that, even with good intentions, no unintentional harm can occur.
Safety	AI systems should not, under defined conditions, lead to a state in which human life, health, property, or the environment is endangered.
Security	The protection mechanisms, design and maintenance of an AI system and infrastructure’s AI systems that can maintain confidentiality, integrity, and availability through protection mechanisms.
Systemic bias	Systemic bias is a social consistent structure of harmful bias that is systemically reinforced in institutions, cultural perception and socio-technical infrastructures. AI systems can reinforce systemic biases by reproducing the discriminatory effects of systemic biases when deployed in socially important institutions, cultural production or in societal infrastructures.
Traceability	Ability to track the journey of a data input through all stages of sampling, labelling, processing and decision making.
Trustworthy AI	Trustworthy AI has three components: (1) it should be lawful, ensuring compliance with all applicable laws and regulations (2) it should be ethical, demonstrating respect for, and ensure adherence to, ethical principles and values and (3) it should be robust, both from a technical and social perspective, since, even with good intentions, AI systems can cause unintentional harm. Characteristics of Trustworthy AI systems include: valid and reliable, safe, secure and resilient, accountable and transparent, explainable and interpretable, privacy-enhanced, and fair with harmful bias managed. Trustworthy AI concerns not only the trustworthiness of the AI system itself but also comprises the trustworthiness of all processes and actors that are part of the AI system’s life cycle.

Terms – to be defined

Term	Definition
Artificial general intelligence (AGI)
Artificial intelligence (AI)
Biometric categorisation system
Biometric data
Conformity
Data governance
Data quality
Deepfake
Deployment
Developer
Distributor
Documentation
Emotion recognition system
Foundation models
Fundamental rights impact assessment
Generative artificial intelligence
Harm
Importer
Incident
Instructions for use (usability)
Life cycle of an AI system
Minimisation
Misuse
Model risk management
Operator
Provider
Pseudo-anonymization (pseudonymization)
Recall of an AI system
Rectification
Remote biometric identification system
Risk
Risk control
Sandbox
Sensitive data
Subliminal techniques
Substantial modification
Terms
Testing data
Testing in real world conditions
Training data
User
Validation