Understanding AI terminology

A process in which an AI agent can dynamically decide, execute and revise steps towards a goal rather than following only a fixed sequence defined in advance.

An AI system that can pursue a goal by deciding what steps to take and carrying out actions, often using tools or other software. The amount of autonomy varies, so agents need permissions, monitoring and guardrails proportionate to what they can do.

Why it matters: Before asking how capable an agent is, ask what actions it is permitted to take, what gets logged, and what happens when it is wrong.

The use of testing, evaluation, audit, documentation and monitoring to provide evidence that an AI system meets relevant requirements for performance, safety, compliance and responsible use.

A structured test of how well an AI model or system performs against defined tasks, risks or quality criteria, using suitable data, metrics and, where appropriate, human judgement.

Why it matters: Vendor benchmark scores do not tell you whether an AI system performs reliably on your own documents, users and workflows.

The roles, policies, processes and controls an organisation uses to decide how AI is selected, developed, used, monitored and retired, including who is accountable for risks and outcomes.

A written set of rules explaining how people in an organisation may use AI, including approved tools and uses, data restrictions, responsibilities and how to report concerns or incidents.

A complete system that uses AI to generate outputs such as predictions, content, recommendations or decisions. It may combine one or more models with software, data sources, interfaces, tools and human processes.

A defined set of computational steps used to solve a problem or produce an output. AI uses algorithms, but most algorithms are not AI.

Processing personal data so that individuals are no longer identifiable by reasonably likely means. Properly anonymised information falls outside UK GDPR; simply removing names or obvious identifiers may not be enough.

Technology that enables a machine-based system to infer from inputs how to generate outputs such as predictions, content, recommendations or decisions. AI is a broad field that includes machine learning as well as other approaches.

Using automated processing to make a decision about a person. UK data protection law places particular safeguards around decisions based solely on automated processing that have legal or similarly significant effects, including rights to information, human intervention and challenge in relevant cases.

Why it matters: The UK rules changed in February 2026 under the Data (Use and Access) Act 2025. If a system makes significant decisions about people, work from current ICO guidance, not older summaries.

Using technology to carry out a task or process automatically, reducing manual intervention. Not all automation uses AI: a fixed rule that forwards every invoice email to finance is automation without AI.

A systematic tendency in data, a model, a system or its use that can skew outcomes. Harmful or unfair bias can arise from training data, design choices, evaluation, deployment context and human decisions, and can contribute to discriminatory outcomes.

Software that interacts with users conversationally through text or speech. Many current chatbots use language models, while others use rules, retrieval systems or combinations of techniques.

The amount of information a model can process within one interaction, usually measured in tokens. It can include instructions, conversation history, retrieved documents and other inputs; information outside the available context is not directly available to the model unless the system retrieves or reintroduces it.

Why it matters: Large documents and long conversations can exceed the available context, affecting what information the model can actually use.

An informal term for an AI assistant designed to work alongside a person inside software or a workflow. Microsoft Copilot is also the name of Microsoft's family of AI products, so the word may refer to a specific product depending on context.

A structured assessment used to identify and reduce data-protection risks before high-risk processing begins. A DPIA is required where personal-data processing is likely to result in a high risk to people's rights and freedoms; many AI uses can trigger this requirement, but not every AI project involving personal data does.

Why it matters: Assuming every AI project needs a DPIA wastes effort, and assuming none does creates legal risk. Screen each use against the ICO's high-risk criteria and record the decision either way.

How suitable an organisation's data is for a proposed AI use, including its quality, accessibility, structure, governance, security and legal basis for use. Poor data readiness can significantly limit an AI project's reliability and value.

Why it matters: Weak data quietly caps what any AI tool can deliver. Assessing readiness first is far cheaper than discovering the problem mid-project.

A type of machine learning that uses neural networks with multiple layers to learn complex representations from data. Deep learning underpins many modern systems for language, images, audio and other data.

A numerical representation of text, images or other data designed so that items with relevant similarities can be compared mathematically. Embeddings allow semantic search to find related material even when it uses different wording.

The extent to which the reasons, factors or processes behind an AI system's output can be meaningfully explained to people. The appropriate level of explanation depends on the system, its users and the consequences of its decisions.

Additional training of a pre-trained model on a targeted dataset to adapt its behaviour or performance for a particular task or domain. It changes model parameters, unlike ordinary prompting or RAG.

Why it matters: Most organisations get what they need from prompting and grounding first. Fine-tuning adds cost and governance overhead, so treat it as a decision to justify, not a default.

A model trained on broad data, usually at scale, that can be adapted for many downstream tasks. Language-model families such as GPT, Claude and Gemini are examples, but foundation models can also work with images, audio and other modalities.

AI that generates new content such as text, images, audio, video or code based on patterns learned from data. It differs from purely predictive or classificatory uses of AI, although a single system may combine both.

Supplying an AI system with relevant, trusted external information, such as documents, database records or search results, so its answer can be based on that evidence. Grounding can reduce unsupported claims but does not guarantee accuracy.

Why it matters: Grounded answers are only as good as the sources they draw on. Grounding reduces invented answers; it does not remove the need for checking.

Technical and procedural controls intended to keep an AI system within agreed boundaries, for example access permissions, input and output checks, restricted actions, escalation rules and human approval for sensitive operations.

When a generative AI system produces information that is false, unsupported or fabricated while presenting it as a plausible answer, for example inventing a citation or statistic. Grounding, verification and human review can reduce the risk, but none guarantees that every error will be caught.

A system design in which a person meaningfully reviews, decides or intervenes at defined points in an AI-assisted process. Human oversight is an important control for many consequential uses, but its form should match the risks and must be designed so that the human can genuinely challenge the system.

Using a trained model on new input to generate a prediction, classification or other output. Training creates or adapts the model; inference is what happens when the trained model is used.

A language model with a very large number of learned parameters, trained on extensive data to process and generate language. LLMs are underlying models used by applications such as chatbots and assistants; the application and the model are not the same thing.

A branch of AI in which models learn patterns from data rather than having every decision rule written explicitly by a programmer. Machine learning underpins much of today's AI, particularly modern generative AI.

The mathematical and computational component that maps inputs to outputs using a learned or specified structure and parameters. A model is usually one part of a wider AI application or system, alongside software, data, interfaces and controls.

A general term for deterioration or change in a deployed model's behaviour as the data or real-world relationships it encounters change over time. Related concepts include data drift and concept drift, which is why deployed systems should be monitored against defined performance measures.

AI that can process or generate more than one type of data, such as text, images, audio or video. For example, a model that can inspect a chart and answer questions about it in text.

The area of AI concerned with processing, analysing and generating human language, including tasks such as translation, summarisation, search and sentiment analysis.

A machine-learning model made from interconnected computational units arranged in layers, which learns patterns by adjusting numerical parameters during training. The original idea was loosely inspired by biological neurons, but modern neural networks should not be understood as simulations of the brain.

AI released under terms that give people meaningful freedom to use, study, modify and share it, with the model components and information needed to exercise those freedoms. This is different from an open-weight model, where the trained model weights are downloadable but the training data, code or licence may remain restricted.

An AI model whose trained parameters or weights are publicly available to download. Open-weight does not necessarily mean open source, because the licence, training data or training code may remain restricted.

Why it matters: Downloadable weights are not the same as open source. Check the licence and what you may legally do with the model before building on it.

Technology that extracts written or printed text from scanned documents or images and converts it into machine-readable text. Modern OCR may use AI, but OCR is not inherently an AI technique.

Numerical values learned during training that determine how a model transforms inputs into outputs. Weights are the principal learned parameters in neural networks, and the two terms are often used loosely as near-synonyms when discussing model size.

Information relating to an identified or identifiable living person, whether identification is direct or indirect. UK data-protection requirements apply to its processing regardless of whether AI is involved.

Training or adaptation performed after a model's initial pre-training to improve behaviours such as instruction-following, task performance, safety or alignment with user preferences.

The initial large-scale training stage in which a model learns broad patterns from a large dataset before being adapted for particular tasks or behaviours.

AI used primarily to estimate, classify or forecast an outcome from input data, for example fraud risk, demand or whether an email is spam, rather than to generate new content.

The input used to guide a generative AI model, which may include an instruction, question, context, examples or a requested output format. Clearer prompts often improve results, although model capability and the underlying information remain important.

An attack that uses malicious or conflicting instructions to manipulate an AI system into behaving contrary to its intended rules. The instruction may be supplied directly by a user or indirectly through content the system reads, such as a webpage, document or email.

Why it matters: Any AI system that reads content you do not control, such as inbound email or webpages, needs input checks and bounded permissions, because that content can carry instructions.

Processing personal data so it cannot be attributed to a specific person without separate additional information, such as replacing names with codes. Pseudonymised data remains personal data under UK data-protection law.

Structured adversarial testing in which testers deliberately try to make an AI system fail, behave unsafely, reveal information or support misuse, so weaknesses can be identified and mitigated.

An umbrella term for designing, developing and using AI in ways that appropriately address safety, fairness, transparency, privacy, accountability, legal compliance and the effects on people and society. The exact principles and controls depend on context.

A technique in which a system retrieves relevant information from a knowledge source and supplies it to a generative model as context for its answer. RAG can provide current or organisation-specific knowledge without retraining the model, although retrieved information and generated answers still need appropriate quality controls.

Why it matters: RAG is often a sensible way to give an AI application your organisation's knowledge without retraining anything. RAG changes what the model sees; fine-tuning changes the model itself.

The use of AI tools or services for work without the organisation's approval or governance. It can create data-protection, security and compliance risks when staff share information with unassessed services; clear policy, approved alternatives and appropriate technical controls can reduce the risk.

Data organised according to a defined schema, such as records in a relational database or a consistently structured table, making individual fields straightforward for software to query and process.

Artificially generated data designed to reproduce useful characteristics of real data. It can support testing, training or privacy-preserving workflows, but synthetic data is not automatically anonymous or free from privacy risk.

A unit into which a model's input and output are divided for processing. In text models a token may be a whole word, part of a word or punctuation; as a rough rule of thumb for English text, one token is about three-quarters of a word, but the relationship varies by model, language and content.

Why it matters: AI services usually price by token, and context windows are measured in them, so token counts drive both cost and how much a model can consider at once.

The process of adjusting a model's parameters using data and an optimisation objective so that it learns patterns or behaviours useful for later inference.

Data used to adjust a model's parameters during training. Its quality, coverage, provenance and biases strongly influence the resulting model, alongside the training objective, architecture, post-training and other design choices.

A neural-network architecture introduced in 2017 that uses attention mechanisms to model relationships between tokens in a sequence. Transformers became the dominant architecture behind modern large language models and are also used for images, audio and other data.

A central part of the UK's data-protection framework, operating alongside the Data Protection Act 2018 and as amended by legislation including the Data (Use and Access) Act 2025. It applies when an AI system processes personal data within the law's scope and is regulated by the Information Commissioner's Office.

Information that does not follow a predefined tabular schema, such as emails, documents, images, audio and call recordings. Modern AI has made it substantially easier to analyse and generate value from many kinds of unstructured information at scale.

A database or search system designed to index numerical vectors such as embeddings and retrieve items based on similarity. Vector search is a common component of RAG, although RAG does not require a dedicated vector database.

Asking a model to perform a task using instructions but no worked examples. Providing one or a small number of examples is usually described as one-shot or few-shot prompting.