A

Accessibility (technology)

Accessibility is considering the needs of all potential users from the outset, engaging with individuals who have disabilities during the design process in order to create solutions that are genuinely usable by everyone.

It also includes assistive technologies, screen readers, voice recognition software and alternative input devices.

Agent communication

The exchange of messages between 2 or more software agents using understood conventions so the agents can share information, make requests, coordinate actions or carry out tasks.

Also known as agent-to-agent communication or inter-agent communication.

Agent-to-Agent (A2A)

An open standard that enables AI agents to communicate and work with each other, including discovering each other’s capabilities, exchanging information and coordinating tasks.

Agentic AI

AI systems designed to pursue objectives through planning, tool use and multi-step action with varying autonomy, subject to human authority and controls.

Agentic AI may introduce additional risks related to boundary controls, escalation thresholds and human oversight.

AI bias or machine learning bias or algorithm bias

Bias in AI models typically arises from two sources: the design of models themselves and the training data they use.

Models can sometimes reflect the assumptions of the developers coding them, which causes them to favour certain outcomes.

Additionally, AI bias can develop due to the data used to train the AI.

AI adoption

The process of introducing AI into an organisation’s operations, from conception and initial procurement through to deployment and ongoing monitoring.

AI adoption lifecycle

The end-to-end sequence of stages an agency follows when adopting AI, as defined in the New Zealand AI Assurance Model: AI platform selection, AI platform implementation, AI solution design, AI solution testing, and deployment and monitoring.

AI agent

Automated entity that senses and responds to its environment and takes actions to achieve goals.

AI-enabled robot

AI solutions that operate through or control robotic devices, machines, sensors, vehicles, robots or other connected equipment in the real or physical world.

AI governance

Governance involves steering responsible development, deployment, and use of AI technologies throughout their lifecycle, by creating and implementing a range of tools such as voluntary guidelines, policies, regulations, and laws, amongst others.

AI guardrail

A control, rule, configuration, process or technical restriction that ensures an AI platform or solution operates within approved boundaries.

AI guardrails may include content filters, prompt restrictions, access controls, human review steps and action whitelisting for autonomous execution of tasks or actions.

Guardrails can help reduce the risks of misuse, harmful outputs and data exposure.

AI lifecycle

An AI system lifecycle typically involves several phases that include:

• to plan and design
• collect and process data
• build model(s) and/or adapt existing model(s) to specific tasks
• test, evaluate, verify, and validate
• make available for use/deploy
• operate and monitor
• retire/decommission.

These phases often take place in an iterative manner and are not necessarily sequential. The decision to retire an AI system from operation may occur at any point during the operation and monitoring phase.

Recommendation of the Council on Artificial Intelligence — OECD

AI model

A mathematical or computational representation used to make inferences from inputs to produce outputs.

A core component of an AI system.

Examples of AI models include Large-Language-Models (LLMs), computer vision as well as other input to output functions like decision trees and neural networks.

AI platform

This definition is an extension to AI system.

An AI platform refers to the underlying technology environment or product that provides AI capabilities. Examples include cloud platforms with built-in AI services or AI capabilities used to process data, or standalone AI tools. A single AI platform can support multiple AI solutions.

In some cases, AI capabilities may be added as an extension or update to a technology platform that did not offer AI capabilities before.

AI solution

This definition is an extension to AI system.

An implemented application, service, product, workflow or operational capability that uses one or more AI systems or AI models to achieve a defined purpose or business outcome.

Examples include cloud platforms with built-in AI services or AI capabilities used to process data or standalone AI tools.

An AI solution may combine AI components with data, software, business processes, human oversight and other technologies, and may be delivered through an AI platform or as a standalone capability.

AI system

An AI system is 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.

Different AI systems vary in their levels of autonomy and adaptiveness after deployment.

AI use/AI use case

The application of AI capabilities within an organisation’s operations, whether through a procured platform, a designed solution or a combination of both.

C

Chatbot

Chatbot is a digital tool or software application designed to simulate conversation with users (primarily via text or synthesised speech). Some operate on predefined responses but advanced versions integrating GenAI provide more dynamic and responsive interactions with users.

Cybersecurity

Cybersecurity involves measures to protect systems, data, and devices from unauthorised access, and ensuring the confidentiality, integrity, and availability of information.

D

Data

Data can be defined as any information in a form capable of being communicated, analysed, or processed (whether by an individual or by computer or other automated means).

Data is useful when it can be communicated easily and analysed to gain insights. Data’s value stems from its use, re-use, and re-purposing, particularly in large volumes. To properly realise this value, data must be accurate, reliable, and free from bias.

Data contamination

Unintentional introduction of inaccurate, incomplete, outdated, biased, corrupted or irrelevant data into an AI system, model or data source, which may reduce the quality, reliability or integrity of its outputs.

An open protocol that enables AI applications or agents to connect to external tools, systems, and data sources in a standardised way.

Data ethics

Data ethics refers to the study and practice of ethical issues related to data, including its generation, recording, processing and distribution, and use. It encompasses principles and standards that guide the responsible and fair handling of data to ensure the rights and privacy of individuals are protected.

Key principles of data ethics:

• Privacy: Ensuring that personal information is collected, stored, and used in ways that protect individuals’ privacy and comply with legal requirements.
• Transparency: Being open about how the data is collected, used, and shared.
• Consent: Obtaining informed consent from individuals before collecting their data.
• Security: Protecting data from unauthorised access, breaches, or cyberattacks to maintain its confidentiality, integrity, and availability.
• Fairness: Ensuring data practices do not result in discrimination or bias and that data will be used in ways that are both equitable and just.
• Accountability: Holding organisations and individuals accountable for their data practices and ensuring there are mechanism in place to address any issues if they arise.

OECD Good Practice Principles for Data Ethics in the Public Sector — OECD

Data poisoning attack

Deliberate manipulation or contamination of the data used by an AI model, platform or solution in a way that affects AI behaviour or outputs.

Data poisoning may occur in training data, fine-tuning data, retrieved source documents or connected data sources. It may create or increase risks relating to inaccuracy, harmful bias, security and misuse.

Data transparency

Data transparency is providing clear and accessible information about the data used in AI systems. This includes understanding where the data comes from, how it has been collected, processed, and used, and making processes more open and understandable to stakeholders.

Deep learning

Deep learning is a more specialised machine learning technique in which more complex layers of data and neural networks are used to process data and make decisions.

E

Explainability

Explainability means enabling people affected by the outcome of an AI system to understand how it was arrived at. This entails providing easy-to-understand information to people affected by an AI system’s outcome that can enable those adversely affected to challenge the outcome, notably – to the extent practicable – the factors and logic that led to an outcome. Notwithstanding, explainability can be achieved in different ways depending on the context (such as, the significance of the outcomes).

Transparency and explainability — OECD.AI

G

Generative AI (GenAI)

Generative AI is a type of AI system that can create or generate new content such as text, images, video and music based off models and patterns detected in existing datasets. (OECD)

H

Hallucination

The OECD defines hallucinations as when GenAI systems create incorrect yet convincing outputs.

Generative AI: the risks and the unknowns — OECD.AI

Human-in-the-loop

Human-in-the-loop refers to the involvement of human oversight and decision-making in the processes that involve AI and automated systems.

This approach allows for critical decisions, especially those impacting individuals, to be reviewed, verified, and influenced by human judgement and expertise.

Human-on-the-loop

Human-on-the-loop allows the system to operate autonomously while supervised by a human who holistically monitors the system and intervenes when needed.

I

Indigenous data

Indigenous data refers to data that’s related to Indigenous peoples, their territories, resources, cultures, languages, and knowledge systems. Data is a critical part of indigenous identity and sovereignty, encompassing a wide range of information from traditional ecological knowledge to health and demographic data.

See also Māori Data Governance.

Indigenous data sovereignty thematic area narrative in English, Arabic, French, Portuguese and Spanish — Global Index on Responsible AI

L

Large Language Models (LLMs)

Large language models, also known as LLMs, are very large deep learning models that are pre-trained on vast amounts of data.

AI language models: Technological, socio-economic and policy considerations — OECD

M

Machine Learning

Machine learning (ML) is a type of artificial intelligence that allows machines to learn from data without being explicitly programmed. It does this by optimising model parameters (that is, internal variables) through calculations, such that the model’s behaviour reflects the data or experience.

The learning algorithm then continuously updates the parameter values as learning progresses, enabling the ML model to learn and make predictions or decisions.

Māori data governance

Māori data governance refers to the principles and practices that ensure Māori data is collected, managed, and used in a way that respects Māori values, rights, and interests.

Key aspects of Māori data governance include:

• Data Sovereignty – ensuring Māori data is subject to Māori governance and control.
• Ethical use, promoting the ethical use of data to enhance wellbeing of Māori people, language, and culture.
• Advocating for Māori involvement in the governance of data repositories and decision-making processes.
• Safeguarding the quality and integrity of Māori data.

Te Mana Raraunga has developed resources to better understand and to support the principles.

Te Mana Raraunga

Misinformation

Misinformation refers to false or inaccurate information that’s spread regardless of an intent to deceive.

Unlike disinformation, which is deliberately misleading, misinformation is often shared without malicious intent.

Model Context Protocol (MCP)

MCP can allow an AI solution to retrieve information or take actions through connected tools, and may introduce risks relating to access control, data exposure, prompt injection, tool misuse and auditability.

O

Open model

A general term for an AI model with one or more publicly available components, such as weights, code, documentation or data information.

The term does not specify which components are available or what rights apply.

Open-source AI

An AI system that anyone can use, study, modify or share, together with the code, model parameters and data information needed to do so.

Open-source AI model

An AI model with documentation, architecture, code or related artefacts available for public access under a licence that allows anyone to access, use, modify or redistribute the AI model, subject to the terms of that licence.

Open-weight AI model

An AI model whose trained parameters (weights) are publicly available, allowing others to run and often modify the model.

Availability of weights alone does not make a model open-source AI.

P

Predictive AI

Predictive AI (or predictive analytics) Involves using data, statistical algorithms, and machine learning techniques to identify the likelihood of future outcomes based on historical data.

Prompt

An instruction, question or other input provided to a generative AI system to guide the output or behaviour it should produce.

A prompt may include context, examples, constraints, conversation history or other data.

Prompt injection

An attack in which crafted or malicious input causes a generative AI system to override, disregard or misapply its instructions, resulting in unintended outputs, disclosure of information or unauthorised actions.

The input may be supplied directly by a user or indirectly through content the system processes.

An ordinary prompt is intended to guide the system towards the user’s or application’s legitimate objective. Prompt injection is an adversarial manipulation of that instruction-processing path.

R

Risk

Risk is the likelihood of adverse impact on people, the environment and society.

Robotic AI

AI solutions that operate through or control robotic devices, machines, sensors, vehicles, robots, or other connected equipment in the real world.

Robotic AI may have the capability to sense, interpret, decide, or act in robotic environments.

S

Stakeholders

As defined in the OECD Recommendation on AI: Persons or groups, or their legitimate representatives, who have rights or interests that are or could be affected by adverse impacts associated with the enterprise’s operations, products, or services.

These can include users of the AI system, civil society, workers’ representatives, service providers, and other enterprises. Anyone involved in or affected by relevant systems.

Recommendation of the Council on Artificial Intelligence — OECD

T

Transparency

Making the operation and decision-making processes of AI systems clear and understandable to users and stakeholders. Key components of transparency are:

• Openness: Clearly communicating the purpose and capabilities of an AI system. This includes explaining what the system is designed to do and any limitations it may have.
• Explainability: Providing understandable explanations of how the AI system reaches it decisions.
• Accountability: Ensuring that there’s a mechanism for tracking and verifying decisions made by the AI. This can include maintaining logs, version control and audit trails
• Data transparency: Disclosing what data is used to train and operate the AI system, including its sources and how its processed.

Transparency and explainability — OECD