Artificial Intelligence glossary
This glossary provides direction in understanding terms throughout the responsible artificial intelligence guidance for businesses.
On this page
- AI system
- As defined by the OECD, 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.
- (AI) Actors
- Those who play an active role in the AI system lifecycle, including organisations and individuals that deploy or operate AI. These include, but are not limited to, those who:
• Develop AI systems, including through ideation, data gathering, model selection and testing
• Sell AI products to end-users
• Deploy AI systems in their business
• Use AI systems that have been deployed by others. - AI lifecycle
- According to the OECD, the AI system lifecycle involves the following phases which take place iteratively and are not necessarily one after the other:
i) ‘design, data and models’ (encompassing planning and design, data collection and processing, as well as model building);
ii) ‘verification and validation’;
iii) ‘deployment’; and
iv) ‘operation and monitoring’.
More information:
Advancing accountability in AI(external link) — OECD.AI - Agentic AI
- A system or program that can autonomously perform tasks on behalf of a user or another system by designing its workflow and using available tools. The system has “agency” (hence the name) to make decisions, take actions, solve complex problems, and interact with external environments beyond the data upon which the system’s machine learning (ML) models were trained.
- Algorithm
- The procedure, set of rules or instructions that is used to solve a problem or otherwise come to an output.
- Attack surface
- The set of possible points where a malicious actor is able to access a system and extract data.
- Audit
- A comprehensive assessment of conformance to standards, policies or legal requirements for data gathering, storage and/or usage.
- (AI) Bias
- Bias allows AI systems to determine how to treat different situations accordingly, and is therefore fundamental to its adaptive capacity when minimised and justified (so as to avoid unfairness).
- Compute
- The large-scale computer resources required for development and use of AI systems.
- Confidential and proprietary information
- Business data that a business wants to protect from public disclosure. This could include processes, formulas, designs or other secret or tightly held information.
- Cost-benefit analysis
- Comparison of the pros and cons (including costs) of a decision or action, to help determine whether it is a valuable and worthwhile activity.
- Data annotation
- The process of labelling data to support machine learning algorithms to understand and classify it.
- Data anonymisation
- Modifying personal information in a way that it can no longer be linked to a specific individual, so as to protect privacy while still enabling data analysis and research.
- Data augmentation
- The process of artificially expanding a dataset, creating new data points through modifying or utilizing the existing data.
- Data drift
- When the statistical properties of distribution of data changes over time, impacting the performance of the model.
- Data encryption
- Scrambling data to mask sensitive information from unauthorised users.
- Data poisoning
- A type of cyberattack in which a malicious actor intentionally compromises a training dataset used by an AI model to influence or manipulate the operation of that model.
- Data provenance
- Recorded history of data, including its origin, transformations and movements through various systems.
- Deep learning
- A subset of machine 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.
- Denial of service attacks
- A denial of service (DoS) attack aims to overload a website or network, with the aim of degrading its performance or even making it completely inaccessible. More information on DoS attacks and preparing for them is available from CertNZ.
Preparing for denial-of-service incidents(external link) — CertNZ - Explainability
- As defined by the OECD, explainability encompasses efforts to enable people affected by AI system outputs and outcomes to understand how they were 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.
- Generative AI
- 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.
- (AI) Governance
- Governance involves steering the development, deployment, and use of AI technologies throughout their lifecycle, in an organisation or jurisdiction by creating and implementing a range of tools such as voluntary guidelines, policies, rules, and regulations, amongst others.
- Intellectual property
- New or original innovations and creations of the mind, including but not limited to creative works, inventions, industrial designs, trade marks, and commercial names.
- Intellectual property
- (AI) Literacy
- The skills, knowledge and understanding to make informed decisions about AI use and development (as required by any individual).
- Machine learning
- A type of AI that allows machines to learn from data without being explicitly programmed. It does this by optimising model parameters (i.e. 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 machine learning model to learn and make predictions or decisions.
- (AI) Model card
- Model cards are short documents accompanying trained machine learning models that provide transparency about various aspects of the model. Some examples are: Claude 3 model card, Amazon Web Services AI Service Cards and Llama 3.1 model card.
- Large Language Model (LLM)
- A very large deep learning GenAI model that is pre-trained on vast amounts of data, allowing it to generate language responses to user inputs.
- What is Deep Learning?(external link) — Amazon Web Services
- Personal information
- Personal information (as defined by the Office of the Privacy Commissioner) is any information which tells us something about a specific individual. The information does not need to name the individual, as long as they are identifiable in other ways (for example through their home address).
- (GenAI) Prompt
- A specific user input into a GenAI tool to help convey what you want the output to do or be.
- Prompt injection
- Prompt injection manipulates an LLM’s behaviour (for example to alter response style, retrieve hidden or restricted data, or disrupt interactions) by embedding specific instructions within a prompt. This approach exploits the model’s tendency to follow instructions within the prompt sequence, even if instructions are unintended or malicious.
- Red teaming
- An organised process of generating malicious model inputs to test the system’s reaction and/or ability to produce harmful behaviour as a result.
- Retrieval augmented generation
- A technique allowing LLMs to access and reference knowledge sources to inform responses as they are being generated, enabling more up-to-date outputs.
- Stakeholder
- Stakeholders (as defined by the OECD) are 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.
- Unfairness
- In an AI context, the unjustified differential treatment (which can occur as a result of data and model bias) that preferentially benefits (or disadvantages) certain groups more than others.
- (AI) watermarking
- Embedding unique identifiers into a GenAI model output to help identify it as being AI-generated, and often to track its origin. These are often only detectable by algorithms rather than visible to human users.
- Web scraping
- Automatic extraction and organisation of information from websites or online files into a structured format for use.