AI - Artificial Intelligence
The umbrella term for computer systems that can perform tasks normally associated with human intelligence, such as analysing information, recognising patterns, generating content, making predictions, recommending actions or supporting decisions.
ANI - Narrow Artificial Intelligence
AI designed for a specific task or defined group of tasks, such as chatbots, image generation, fraud detection, translation, scheduling, document review or medical diagnosis support.
Workplace
AI governance
The structures, processes, responsibilities and controls used to manage AI safely and responsibly.
Generative AI - Generative Artificial Intelligence
AI that can create new content such as text, images, audio, video, music, code, summaries or synthetic data in response to prompts or inputs.
Agentic AI
AI systems that can pursue a goal through multiple steps and use tools or connected services to take actions. Agentic systems are in operational use, but their autonomy, reliability and safeguards vary substantially.
Adaptive AI
AI that can change aspects of its behaviour using new data, feedback or conditions after deployment. Adaptive systems are already used; “AAI” is not a universally standard acronym.
Artificial general intelligence (AGI)
A contested concept for AI with broad, general capabilities across many cognitive tasks. There is no agreed technical or legal threshold and no accepted evidence that a current system has achieved AGI.
Artificial superintelligence (ASI)
A hypothetical form of AI that would substantially exceed human capability across most or all relevant domains. No accepted evidence establishes that a current system meets this concept.
Collective and multi-agent AI
Multiple AI agents or systems coordinating, exchanging information or dividing work to pursue a goal. Multi-agent implementations are in use, while “collective AI” and “CAI” are not settled global categories.
Workplace explanationIn use now
Human-AI collaboration or hybrid intelligence
A work arrangement that combines AI outputs with human expertise, judgement or approval. The pattern is widely used, but “HAI” is not a universal acronym and does not prove that oversight is effective.
Self-aware artificial intelligence
A hypothetical AI that would possess consciousness or self-awareness. There is no accepted evidence that current AI systems are conscious or self-aware; “SAI” is not a universal acronym.
Workplace
Automated decision-making
The use of a system to make or substantially support decisions, sometimes with limited human involvement.
Workplace
AI risk appetite
The level and type of AI-related risk an organisation is prepared to accept in pursuit of its goals.
Workplace
AI inventory or register
A record of approved AI tools and systems, including purpose, owner, data used, risks, controls and review date.
Workplace
AI assurance
Testing, validation, audit or review processes that give confidence an AI system is operating as intended and within risk limits.
Workplace
AI literacy
The knowledge and capability needed to understand AI benefits, limitations, risks and safe use.
Workplace
Data governance
The management of data quality, security, access, lifecycle, legal compliance and appropriate use.
Workplace
Human oversight
Human monitoring, review, intervention or approval of AI outputs, decisions or actions.
Workplace
Shadow AI
Unapproved AI use at work without formal oversight, policy approval or risk controls.
AI developer
A person or organisation that designs, builds, trains, adapts or tests an AI model or system. One organisation can be both a developer and a deployer.
AI incident
An event or series of events involving an AI system that causes, or could cause, harm, a legal breach, a security or privacy compromise, disruption, or a material failure of the system. Incident thresholds vary by framework and organisation.
AI use case
A specific application of one or more AI systems to achieve an objective or perform a task. Risk should be assessed at the use-case level because the same tool can have low-risk and high-risk uses.
Algorithmic management
The use of tracked data and algorithmic systems to organise, allocate, monitor, supervise or evaluate work. It may use AI, but it can also rely on conventional rules-based software.
Biometric information
Information derived from physical, biological or behavioural characteristics, such as facial features, fingerprints, iris patterns or voice. Under Australia’s Privacy Act, biometric information used for automated verification or identification, and biometric templates, are sensitive information.
Contestability
The ability for a person affected by an AI-assisted outcome to question it, seek an explanation, provide relevant information and use an appropriate review or complaint pathway. The exact legal rights depend on the decision and jurisdiction.
Data poisoning
An attack or failure in which training, fine-tuning or retrieval data is manipulated so that a model learns or produces unsafe, biased or attacker-chosen behaviour.
Workplace explanation
Data residency
The physical or geographic location where data is stored or processed. Residency does not, by itself, determine every law, access right or control that applies.
Workplace explanation
Data sovereignty
The legal and practical control of data, including which laws, governmental powers, contracts and organisations can affect it. Australian data residency can support a sovereignty objective but does not guarantee it.
Deepfake and synthetic media
Audio, image or video content generated or materially altered by AI to depict events, people or statements that did not occur as presented. Synthetic media can be legitimate; deception, impersonation and misuse create the risk.
Fine-tuning
Further training a pre-trained model on selected data to adapt its behaviour or performance for a particular task or domain.
Frontier AI
A policy term for highly capable general-purpose AI near the leading edge of current capability. There is no single global threshold; definitions can change as models, evaluations and risks change.
Grounding
Supplying a model with relevant context or trusted information at inference time to improve the relevance and factual basis of its output. Grounding does not guarantee accuracy.
Hallucination or confabulation
A confident-looking AI output that is false, internally inconsistent or unsupported by the input or available evidence. NIST uses “confabulation”; “hallucination” is the common industry term.
Workplace explanationIn use now
Human-in-the-loop
A control pattern in which a person must review, approve, correct or complete a defined step before the system can continue or an outcome takes effect. It is a governance design choice, not a universal legal right.
Human-on-the-loop
A control pattern in which a system operates with some autonomy while a person monitors it and can intervene, pause or stop it. Effective oversight requires real authority, information, time and system controls.
Interpretability
The extent to which a person can understand the internal relationships or mechanics that produce a model’s output. Interpretability is related to, but not identical with, an explanation presented to a user.
Large language model (LLM)
A model trained on large quantities of language data to predict and generate sequences of tokens. LLMs can support many tasks, but fluent output is not proof of truth or reasoning.
Machine learning (ML)
A set of computational methods that improve model performance by learning patterns or parameters from data or experience rather than relying only on explicitly programmed rules.
Model drift
A decline or change in model performance because the data, population, environment or relationships encountered after deployment differ from those used to build or validate it.
Multimodal AI
AI that accepts, relates or generates more than one type of information, such as text, images, audio, video or sensor data.
Personal information
Under Australia’s Privacy Act, information or an opinion about an identified individual, or an individual who is reasonably identifiable, whether the information is true or recorded in material form or not.
Prompt
Input or instructions supplied to a generative AI system to influence the output. A prompt can include text, images, audio, documents, examples or structured data.
Prompt engineering
Designing and refining prompts, context and examples to guide a model toward a desired task or output without changing the model’s trained weights.
Prompt injection
A vulnerability in which instructions in user input or external content cause a language-model application to behave contrary to its intended rules. Indirect prompt injection can arrive through webpages, files, emails or retrieved data.
Retrieval-augmented generation (RAG)
A method that retrieves relevant material from an external source and supplies it to a generative model when producing an answer. RAG can improve relevance but does not guarantee accuracy or trusted sources.
Sensitive information
A higher-protection category of personal information under Australia’s Privacy Act. It includes specified information such as health, racial or ethnic origin, political opinions, union membership, sexual orientation, criminal record and certain biometric information.
Token
A unit of content processed by a model. A token may be a word, part of a word, punctuation, code or another meaningful unit, depending on the model.
AI agent
An automated entity that senses conditions, responds to them and takes actions to pursue goals.
AI audit
A review, done internally or independently, that checks whether an AI system meets the requirements of a chosen framework.
AI benchmark
Shared datasets, tasks and measures used as comparison points for AI system performance.
AI deployer
A person or organisation that uses or supplies an AI system in delivering a product or service, either internally or to people outside the organisation.
AI lifecycle
The stages an AI system moves through, from idea and development to testing, deployment, operation and retirement.
AI model
A representation of data, a process or a phenomenon that uses AI algorithms to interpret inputs, predict outcomes or generate responses. Machine learning models produce inferences or predictions from input data.
AI partner
An organisation or entity that provides AI-related services, such as integration, data support, evaluation or audit services.
AI technology producer
An organisation or entity that designs, develops, tests and provides AI technologies, models or components.
AI provider (product or service)
An organisation or entity that provides products or services that use one or more AI systems.
AI supply chain
The resources needed to develop, deploy and maintain AI, including data, models, hardware, compute, people and funding.
AI system
A machine-based system that uses inputs and objectives to infer outputs such as predictions, content, recommendations or decisions that can affect digital or physical environments.
AI user
An entity that uses or relies on an AI system, including an organisation, an individual or another system.
AI value chain
The broader process of creating and delivering value through AI systems, extending beyond the supply chain.
Accountable / accountability
Being answerable for actions, decisions and performance within a defined responsibility area. Accountability is the state of being accountable.
Affected stakeholder
A person, organisation, community or system affected by the decisions or behaviours of an AI system.
Algorithm
Instructions a computer follows to perform tasks or solve problems. A machine learning algorithm learns model settings from data against selected criteria.
Assurance
The process of measuring, evaluating and communicating evidence about a system, process or organisation to confirm it is operating as intended. AI assurance focuses on trustworthiness.
Bias
A systematic difference in how certain people, groups or objects are treated compared with others.
Continuous learning
Ongoing incremental training of an AI system during its operational phase.
ISO/IEC 29147
Disclosure
Giving relevant information to a party that may not already know it.
Evaluation
Assessment against criteria. This can include model evaluation, system evaluation, capability evaluation, benchmarking, testing, verification, validation, risk assessment and impact assessment.
Explainability
The ability of an AI system to present the important factors behind a result in a way people can understand.
Fairness
Treatment, behaviour or outcomes that are not driven by favouritism or unjust discrimination. Unfairness occurs when differential treatment is unjustified.
Fingerprint
A set of computable properties that can identify digital content or near-duplicate content.
Foundation model
A very large AI model trained on broad data, often using self-supervision, and usable across many contexts.
General-purpose AI
An AI system type that can address a broad range of tasks and uses, including uses not originally intended by developers.
Impact assessment (AI system)
A structured assessment of wider economic, social and environmental effects that may result from deploying and operating an AI system.
Justified / calibrated trust of AI
Trust in an AI system based on reliable evidence, avoiding both over-trust and under-trust.
Labelling
Applying classification labels to information assets. For AI content, labelling can include visible notices that content was AI-generated and information about provenance.
Measurement (of AI systems)
Using quantitative, qualitative or mixed methods to assess, benchmark and monitor AI performance, risk and impact.
Metrics
Qualitative or quantitative measures used to assess, compare and track quality or performance. Metrics may focus on the model, system behaviour, risk outcomes or broader impacts.
Performance
Measurable results.
Provenance
The history of an asset and its interactions with actors and other assets, represented through provenance data.
Red teaming
A simulated adversarial exercise, designed to reflect real-world conditions, used to assess AI system and organisational security capability.
Responsible AI
Developing and using AI in ways that create benefits while reducing the risk of negative consequences for people, groups and society.
Risk
The relationship between likelihood and consequence, including uncertainty about objectives and the chance of harm. AI consequences may be positive, negative or both.
Risk assessment (AI system)
A systematic process for identifying and evaluating the likelihood and consequences of AI-related events or actions that could cause harm.
Risk control (AI system)
A measure that maintains or changes risk, including a process, policy, device, practice or other action.
Risk mitigation (AI system)
Practices and tools used to reduce the likelihood or consequences of AI-related events or actions that could cause harm.
Test case
A specification for testing, including inputs, test steps and expected outcomes. A group of test cases is a test suite.
Testing
Executing an AI model or system to check whether it shows expected behaviours across selected test cases.
Transparency
For organisations, communicating relevant activities and decisions to stakeholders in an accessible way. For systems, making appropriate information available, such as features, performance, limits, components, assumptions, data sources and evaluation methods.
ISO/IEC 25010
Trust (of AI system)
The extent to which a stakeholder is persuaded that an AI system will behave as intended.
Trustworthiness (of AI system)
The ability of an AI system to meet stakeholder expectations in a way that can be verified.
Validation
Evidence-based confirmation that user needs have been fulfilled.
Verification
Evidence-based confirmation that specified requirements have been fulfilled.
Watermark
Information embedded in digital content, visibly or invisibly, to show provenance or indicate that content was generated or materially modified by AI.
No terms match that search or filter.