Glossary for clear AI conversations

AI system

A machine-based system that produces outputs such as predictions, content, recommendations, or decisions from inputs, often with different levels of autonomy.

General-purpose AI

An AI model designed to perform a wide range of tasks rather than one narrowly defined function.

Large language model

A model trained on large amounts of text to predict and generate language. It can be useful without reliably knowing whether its words are true.

Training data

Examples used to develop a model’s patterns and behaviour. Training data may be incomplete, biased, outdated, or subject to rights and privacy limits.

Prompt

The instruction, question, context, or data given to an AI system.

Output

What an AI system returns, such as text, a label, an image, a recommendation, or an action proposal.

Hallucination

An invented or unsupported output presented as if it were a valid fact, source, quote, or explanation.

Retrieval

Finding material from a connected collection or the web to include as context for an AI response.

Grounding

Connecting an answer to specified source material or data. Grounding helps trace an answer but does not automatically make the source true.

Primary source

The original evidence closest to an event or claim, such as a dataset, experiment, official record, law, or first-hand account.

Secondary source

Material that reports on, interprets, or summarises primary evidence, such as journalism, a review, or an analysis.

Marketing claim

A statement intended to persuade people to choose a product or service. It may be useful context but has an incentive to highlight benefits.

Automation bias

The tendency to trust or follow an automated suggestion too much, especially when it looks precise or arrives under time pressure.

Human oversight

Meaningful human ability to understand, question, change, or stop an AI-supported result or action.

Personal data

Information relating to an identified or identifiable person, such as a name, email address, identifier, or location.

Sensitive data

Data that can cause greater harm or requires special care, such as health, biometric, financial, identity, or confidential information.

Bias

A systematic skew in data, a model, or a process that can produce unequal or unfair results.

High-risk AI

AI used in contexts where errors or rights impacts can be especially serious and where stronger controls may apply under law or policy.

Provider

A person or organisation that develops an AI system or model and puts it into service under its name.

Deployer

A person or organisation that uses an AI system in its work or under its authority.

Affected person

Someone who may experience a consequence of an AI-supported output or decision, whether or not they use the system.

Sycophancy

A model’s tendency to agree with a user’s stated view or desired answer instead of independently testing it.

Pattern completion

Generating a likely continuation from learned patterns, even when the system lacks enough evidence to answer accurately.

Confabulation

Another term for generating plausible but unsupported or invented content, often used in discussions of generative AI risk.

Source literacy

The ability to judge what a source is, who produced it, why, when, and how well it supports a claim.

Evidence

Information that supports or challenges a claim, with strength depending on its relevance, reliability, and proximity to what happened.

Citation

A reference pointing readers to material said to support a claim. A citation must be checked because it can be fabricated, irrelevant, or misread.

Vendor claim

A statement about a product or service made by the organisation selling or providing it.

Sponsored content

Material paid for or influenced by an interested party, even when it is presented in an editorial-looking format.

SEO content

Web content shaped mainly to rank in search results and attract traffic; visibility is not proof of quality.

Opinion

A viewpoint or interpretation rather than a directly established fact, even when expressed by an experienced person.

Verification

Checking an output or claim against appropriate source material, calculations, records, or another suitable method.

Data minimisation

Using only the data needed for a task and removing unnecessary identifiers or details.

Confidential data

Information restricted because disclosure could harm people, an organisation, a customer, or a partner.

Prompt injection

Instructions hidden in input or retrieved content that try to redirect an AI system away from the user’s intended task or controls.

Proportional review

Matching the depth and speed of checking to the possible impact, reversibility, and uncertainty of an AI-supported result.