Term Explorer

A set of step-by-step instructions for turning inputs into an output or action.

Attributing human-like traits, feelings, intentions or understanding to an AI system when its observable output does not establish those inner states.

A debated label for an AI system with broad capability across many intellectual tasks and domains, often compared with human general ability.

An umbrella term for computational systems using symbolic reasoning, search or learned models for tasks such as prediction, classification or content generation.

Automation uses a machine or software to carry out set steps with little or no human action after setup.

Recorded observations, measurements or examples that people or systems can store, examine and use for a purpose.

A dataset is an organised collection of examples that a system can use for analysis, training, testing or search.

A method that trains many-layered computer models to find useful patterns in large amounts of data.

A machine-learning model pretrained at large scale on broad data can be adapted for many tasks without training a new model from scratch.

A label for AI models near the current leading edge of capability or scale, especially those seen as pushing technical limits and raising new governance questions.

A type of AI that produces new content such as text, images, audio or video from patterns learned from many examples.

The idea that a machine might have subjective experience or awareness, not just produce intelligent-seeming behaviour.

A way of building computer systems that learn patterns from examples so they can make predictions, classifications or recommendations on new cases.

A parameterised mathematical function learned from data that maps inputs to outputs such as labels, scores, predictions or generated text.

An AI system built for a limited task or related tasks, rather than broad intellectual work across domains.

AI systems released under licence terms that let others inspect, use, modify and redistribute source code and other materials needed to run them.

A model whose trained numerical parameters are publicly released so others can run, inspect or adapt it, even if the training code, data or licence are not fully open.

A system that uses past data to estimate what is likely to happen next, such as a future value, category or level of risk.

An AI model whose owner uses licence terms or technical controls to restrict access to or reuse of key components, such as its weights or code.

A rule-based system applies explicit if-then rules to given inputs to produce specified outputs.

A set of coded instructions that tells a computer how to take input, process it and produce output.

A machine learning method that assigns one or more predefined category labels to an input based on learned patterns.

A method that groups similar examples together so patterns in data can be found without pre-written categories.

These words describe units within training: a batch groups examples, an iteration is an update step, and an epoch is one full dataset pass.

A feature is a piece of input information a model uses to help produce an output, such as a prediction, score or grouping.

The checked reference answer used to train or evaluate a model, so its output can be compared with the best available real-world answer.

Using a trained model on new input to produce an output such as a score, category, translation or generated text.

The expected answer attached to an example so a model can learn what output should match that input.

The result a model produces from input, such as a class, number, ranking or next token.

A supervised learning method that learns from example inputs and known target values to predict a numeric quantity for a new input.

A training method that aims to improve behaviour through trial, feedback and cumulative reward across a sequence of actions.

A training approach that creates its own learning targets from raw data, so a model can learn useful patterns without manual labels.

A learning approach that combines labelled and unlabelled examples to train a predictor for the same target task.

A method for training a model on examples paired with target outputs, so it can predict outputs for new inputs.

Training is the process of adjusting a model's parameters using data or feedback to improve performance on a learning objective.

A single row of training data used to adjust a model, often made of input values and, in some methods, a target answer.

Reusing patterns learned on one task or dataset to help a model learn a related task faster and with less new data.

A way of training a model on unlabelled data to identify possible structure, similarities, groups or unusual cases without target answers for each example.

Separate portions of data are kept aside to tune a model during development and then to check, at the end, how well it works on unseen examples.

A neural network usually uses an activation function to transform a unit’s weighted input into an output, often adding the non-linear behaviour needed to learn complex patterns.

A small calculation step in a neural network that combines input numbers, adjusts them by learned settings, and passes forward a new number.

A bias is an extra adjustable number added to a neuron's weighted input so the neuron's output can shift even when all inputs are zero.

A comparison between neural networks and brains that can help with intuition, but breaks down because AI systems use simplified maths rather than living cells and biological processes.

A layered model that reuses small filters across a grid-like input, combines local patterns into larger features, and produces task outputs from them.

The calculation that carries input numbers through a neural network’s layers to produce an output such as a score, label or next-token distribution.

These layers organise how a neural network turns input numbers into intermediate calculations and then into an output such as a score, label or predicted value.

A neural network is a layered set of connected numerical operations that turns input data into an output such as a label, score or prediction.

A number attached to a model’s possible outputs that shows how much of its total estimated likelihood the model assigns to each option.

A recurrent neural network processes a sequence by carrying a changing hidden state from one step to the next, so earlier inputs can influence later outputs.

A numerical function that turns a list of model scores into positive values that sum to 1, highlighting which outputs are larger relative to the others.

A weight is a learned number that scales how strongly one input signal affects the next calculation in a neural network.

A training method that adjusts each model setting using both recent gradient direction and the recent size of past gradients.

A method that propagates derivatives backwards through a model to compute how changes to each weight would affect the loss.

A calculus rule for tracing how a small change in one setting affects the final error through a chain of intermediate calculations.

Convergence means training updates are settling towards a stable, lower-loss solution instead of continuing to wander or blow up.

A training method that randomly switches off some units during updates to reduce reliance on particular pathways and potentially improve generalisation.

Error describes a mismatch between an output and a reference. Loss is a scalar training objective for one case or an aggregate across cases.

A gradient tells you how much a model’s error would change if each setting were nudged slightly up or down.

A method for reducing a model’s loss by repeatedly adjusting its internal settings in the direction opposite the loss gradient.

A hyperparameter is a training setting chosen before or during learning that shapes how a model is fitted, rather than a value the model learns from data.

A training setting that controls how far model parameters move after each error-based update.

A local minimum is the lowest loss in a nearby region, while a global minimum is the lowest loss anywhere in the whole set of possible model settings.

A loss function turns a model’s output and the correct target into a single number that shows how wrong the model was for that example or batch.

A training method that carries part of the previous update into the next one so parameter changes keep moving in useful directions and wobble less.

The process of adjusting a model’s internal settings to reduce loss, plus the algorithms and rules used to choose each adjustment.

Regularisation changes a model's training objective or procedure to reduce overfitting and improve performance beyond the training data.

Training can stall or blow up when error signals shrink toward zero across layers or time steps, or grow so large that updates become unstable.

The share of cases a system labels correctly out of all the cases it was tested on.

A reference method or result used to judge how a model performs against a relevant starting point.

A benchmark is a standard test set-up used to compare how well systems perform on a defined task under stated conditions.

Statistical bias is a systematic gap between a model’s expected prediction across possible training samples and the true target, often caused by restrictive assumptions.

Calibration describes how closely a system’s probability scores match the observed frequencies of the corresponding outcomes.

A dataset has this problem when one class appears far more often than another, which can make overall accuracy look strong while a model misses rare cases.

A table that compares a classifier’s predictions with the real outcomes, so you can see which kinds of mistakes it makes and how often.

A method for estimating how well a model will perform on new data by repeatedly training on one part of a dataset and checking results on another part.

Data leakage happens when future, target or held-out test information improperly influences model training or evaluation, making performance look better than it is.

Evaluation checks how well a system performs on a defined task, while metrics are the specific measures and evals are the tests used to produce them.

A single score that balances how many positive cases a system finds with how often its positive predictions are correct.

False positives are wrong yes predictions, while false negatives are wrong no predictions.

The ability of a model to give useful results on new cases, not just on the examples used to build or tune it.

Memorisation happens when a model encodes information specific to individual training examples or small subsets of them.

A model fits its training examples so closely that its performance drops on new cases.

The share of flagged results that are actually correct positives.

The share of real positive cases that a system successfully flags.

Reliability is how consistently a system gives similar quality results when it is used again on comparable cases and conditions.

The ability of a model or system to keep performing acceptably when inputs, conditions or small disturbances change.

A model misses important patterns because it is too simple, too constrained, or too weakly trained for the task.

Variance describes how much a fitted model’s predictions change when the same learning procedure is trained on slightly different samples.

In language models, apparent comprehension usually means producing text that fits patterns in data and context, not reliably building a grounded grasp of the world like a person does.

The text a language model produces after a prompt, built token by token from patterns learned during training.

The text, images or other inputs a model can currently use to shape its next output in a session.

The maximum amount of input and prior output a language model can use at one time while producing the next part of a response.

The same full input and generation conditions produce the same output every time, rather than varying across repeated runs.

A generated answer that sounds plausible but includes false, unsupported or invented details.

A language model estimates probabilities for pieces of language in context; many can use these estimates to generate text from a prompt.

A very large text-processing model predicts likely next pieces of language from a prompt and uses that pattern-matching ability to generate or transform text.

The delay between sending a request to a model and receiving its reply, or the next part of that reply.

A released file of learned numerical settings that lets compatible software run a trained model without repeating the original training process.

A multimodal model can take in or produce more than one kind of data, such as text, images or audio, within one system.

A language model produces text by repeatedly estimating probabilities for small pieces of text that could come next in the current sequence.

A parameter is a learned numerical setting inside a model that helps shape which token it is likely to produce next.

The text, image or other input that guides a model on what kind of output to produce next.

The practice of writing and refining instructions so a language model is more likely to produce a useful, safe and well-structured reply.

A reasoning model is a language model designed for stronger performance on multi-step tasks, often using extra computation while answering, though visible steps may not reflect its internal process.

A stochastic procedure that draws one next token from the model’s probability distribution, appends it to the output, and repeats until a stop condition.

A high-priority instruction that sets the assistant’s role, rules and priorities before it responds to later messages.

A setting that controls how predictable or varied a language model’s next chosen words are during generation.

Throughput measures how much text a model can process or generate in a given time, usually counted in tokens per second.

A token is one unit from a model’s vocabulary used to represent input or output, often smaller than a word.

The process of breaking text into smaller units a language model can process, such as words, parts of words, punctuation marks or special symbols.

A decoding setting that limits each next-token choice to the smallest group whose combined probability reaches a chosen threshold, such as 0.9.

The input a person gives a chat assistant to guide what kind of response it should produce.

A VLM links images and text so one model can take visual input, written input, or both, and produce text or other structured outputs about them.

The full set of tokens a language model can represent and choose from during text processing and generation.

A method that lets a model give different importance to different parts of the input while producing an output.

A number comparing one token’s query with a candidate token’s key during an attention step.

A model that assigns each token a probability based on earlier tokens in the same sequence, and generates by predicting and adding one new token at a time.

A list of numbers that places a token or other item in a mathematical space so a model can compare patterns, relationships and context.

In an encoder-decoder model, the encoder represents input, while the decoder uses those representations to generate output, often one token at a time.

A transformer mechanism that runs several attention patterns in parallel so the model can combine different kinds of relationships between tokens.

A method that supplies sequence-position information so attention can use word order, not just token content.

Three vectors produced by learned projections represent matching needs, matching features, and information that attention may pass onward.

A mechanism that lets each token reweight information from available tokens in the same sequence, including itself, before the next layer processes it.

A neural network design that processes many parts of a sequence in parallel by weighting which parts matter most to each other.

A vector is an ordered list of numbers, and a vector space is a set of same-length vectors that is closed under addition and scaling.

Chunking splits source material into smaller pieces so search and retrieval can find and return the most relevant passage, rather than an entire document or a random fragment.

A citation in an AI answer points you to a source passage or document intended to support a specific claim.

A measure of how closely two non-zero vectors point in the same direction, commonly used to compare embeddings in search.

Grounding ties an AI answer to specific retrieved source passages so the reply is based on provided material rather than only on the model's general patterns.

Indexing organises documents or document chunks into a searchable structure so a system can quickly find the most relevant passages for a query.

A search method that finds documents by matching query words to indexed text and ranking the matches, often with scoring methods such as BM25.

An organised set of trusted documents that a search system can look through to find relevant passages for an answer.

A method for finding the items most similar to a query in a vector space; approximate methods often trade a little accuracy for much faster search at scale.

A system orders possible results by estimated fit to a query, then may score and reorder a shortlist with a second method.

Retrieval returns passages that a search method ranks as relevant, so a person or later model step can use source material.

A method that first finds relevant source passages and then asks a language model to answer using that material.

A search method that finds material by similar meaning, not just shared words, and returns the most relevant passages for review or further AI use.

A measure of how close two pieces of text are in meaning, even if they use different words.

A specialised system stores numerical meaning-representations of content and quickly finds the closest matches to a query by semantic similarity.

Making an AI system's behaviour better match intended goals or preferences, often through methods applied after broad training.

A saved copy of a model’s state at a particular point in training or adaptation, so work can be resumed, compared or released later.

Distillation trains a student model to imitate a teacher model’s outputs or other signals; the student is often smaller.

Updating a pre-trained model's parameters with additional examples to improve performance on a narrower task, style or domain.

A model uses examples and instructions placed in the current prompt to imitate a pattern for that one interaction, without changing its stored parameters.

A post-pretraining adaptation step that teaches a general language model to respond to prompts in the requested format, style and task pattern.

The approximate point up to which a model’s training-derived background is expected to be current, while newer events and facts may be absent unless added another way.

A method for adapting a large model by training a small set of added weights instead of changing most of the original model.

A particular identified state of a model or model-backed system, with documented changes, capabilities and checks for use.

A way to adapt a large model to a new task by training only a small set of added or selected weights instead of updating the whole model.

A method for adjusting a model so it more often produces outputs people rank as better, safer or more useful than alternatives.

A large-scale training stage where a model learns reusable patterns from data, producing a general-purpose starting point that later steps can adapt for specific tasks.

Pruning removes less useful connections or parameters to reduce model size or runtime while trying to keep nearly the same performance.

A method for storing and running a model with lower-precision numbers so it uses less memory and often runs faster, with some possible loss in accuracy.

A method for adjusting a model so its outputs better match human preferences by learning from ranked examples and reward signals.

A pretrained model is further adapted by showing it input–output examples so its parameters better reproduce the target outputs for a chosen task or style.

These labels describe how many task examples a model gets in the prompt before it tries a new input: none, one, or a small handful.

A system that can pursue a goal through several tool-using steps, checking results and adjusting its next action with limited human direction.

An AI system that can pursue a goal by selecting and carrying out multiple steps, often using software tools, with oversight matched to its risks.

Software that uses AI to help with tasks by taking instructions, sometimes using connected tools, and returning a draft, answer or proposed action for a person to review.

An application programming interface defines rules that software uses to request data or actions from another system.

A chatbot is a system that takes messages in ordinary language and returns replies in a back-and-forth conversation.

A system capability that lets software observe what appears on a screen and operate apps by clicking, typing, scrolling and selecting items.

A connector links an AI system or app to another digital service so it can send requests, receive results, and use them in a task.

A tool that helps a person complete digital tasks by drafting, searching, and sometimes taking limited actions under human direction and review.

A process where the result of an action is fed back into the system and changes a later action, often through comparison with a goal or rule.

A method that lets a model return a structured request for an external tool or system instead of only producing free-form text.

A system setup where a person reviews, guides or approves an AI-driven step before the workflow continues or completes.

Stored information an AI system can retrieve across separate sessions to personalise, continue tasks, or reuse past facts and preferences.

A standard way for an AI application to connect with external tools and data sources through a shared, predictable interface.

A multi-agent system contains multiple software agents that interact, directly or through a control layer, while carrying out a shared or related task.

In an acting system, observations are the signals it receives, actions are the steps it takes, and the act-observe loop is the repeated cycle linking the two.

The control layer that coordinates an AI system’s steps, tools, rules and handoffs so a task can move from request to checked result.

Planning is the process of turning a goal into ordered sub-tasks, checks and possible actions so a system can act step by step.

Temporary information kept available during a current interaction so an AI system can use recent instructions, results and observations in its next step.

A system uses external software or devices during a task to look up information, calculate, carry out actions or control equipment beyond its own processing.

A workflow is an ordered set of steps that moves information or tasks from input to result, often combining software, people and checks.

Tools in this area create new speech, music or other sounds from text, audio examples or other controls, including voices that imitate a particular person.

A system produces computer code from prompts, examples or surrounding files. People should test and revise the result before using it.

A deepfake is an AI-generated or heavily altered image, audio clip or video that convincingly imitates an identifiable real person, usually in a recording or event that did not happen as shown.

A diffusion model creates media by learning how to remove noise step by step until a coherent image, audio clip or other output appears.

A training method with two neural networks, one making synthetic examples and one trying to spot fakes, with the aim of making generated output more realistic.

Software can create new pictures from prompts or other inputs by learning visual patterns from many examples, then producing a fresh image rather than retrieving one file.

False or misleading content can spread by mistake or on purpose, and the difference matters because intent changes how we respond.

Tools in this group turn spoken audio into written text, while related labels differ slightly in scope, method and how much human checking is involved.

Synthetic media is text, images, audio or video generated or substantially altered by software, often AI, so it can resemble authentic material.

A system produces new written words from a prompt, patterning its output on examples seen during training.

A generative model that compresses examples into a regularised probabilistic latent space, then reconstructs inputs or samples from that space to create new outputs.

A system that creates moving pictures from prompts, images or other guidance, producing short video clips that did not previously exist as recorded footage.

Content provenance records claims about origin and edits, while watermarking embeds a detectable signal that may indicate a source or how media was made.

A rectangular marker that shows where an object or text region appears in an image, usually by its position, width and height.

Computer vision enables software to analyse images or video and produce useful outputs such as labels, locations, measurements or text.

An organised collection of text or recorded speech used to study language or train and test language-processing systems.

A computer vision method that compares face images or representations to estimate whether they show the same person.

A task that turns an image into a short written description, often to summarise visible content for search, accessibility or workflow support.

A system assigns one or more predefined labels to an image based on patterns in its pixels.

Image segmentation divides an image into pixel-level regions; semantic segmentation labels them as roads, buildings or water.

A system that assigns a user message or action to a goal category, such as asking for help, making a booking or reporting a problem.

This status applies to a language and task when useful data, software tools or evaluation material are limited for building and assessing technology.

Machine translation uses computer systems to convert text from one language into another using learned patterns, statistical methods or hand-built rules.

A text-processing method that finds mentions such as people, organisations, places, dates or amounts and labels each mention by type.

A set of computing methods for turning human language into structured signals that software can classify, search, translate, summarise or generate.

A computer vision task that finds instances of chosen object types in an image or video and marks where each one appears.

A method that turns words in an image or scanned page into machine-readable text that software can search, copy and analyse.

The smallest addressable picture element in a digital image, usually storing colour or brightness information at one tiny location.

A computer vision method that locates important body, face or hand points in an image or video and often links them into a skeleton-like layout.

A method that sorts text, speech transcripts or sometimes images into attitudes such as positive, negative or neutral.

Summarisation produces a shorter version of text or speech that keeps the main points and drops less important detail.

A method that scans many documents and groups frequently co-occurring words into recurring themes, helping people see what a collection is mostly about.

People or software tools add task-relevant tags, notes or boundaries to data for model training, testing, analysis or monitoring.

Using on-demand computing services over a network, instead of relying only on machines you manage directly.

Compute is the processing capacity used to train and run AI systems, with different needs before deployment and during everyday use.

The price charged for each small piece of text a model processes or generates, usually counted separately for input and output.

The process of checking data, fixing problems, and putting it into a usable form for analysis or model training.

The pattern of values, categories and frequencies in a dataset, including what appears often, rarely or not at all.

Data governance is the set of rules, roles and checks that control how data is collected, stored, used, shared and corrected over time.

Rules and arrangements about where data is stored, processed and governed, and which laws and authorities apply to it.

Deployment means putting an AI system into real use so people or software can send it live inputs and receive live outputs.

These terms describe mismatches between development and live data, including changed inputs, labels or relationships, which can reduce a system's reliability.

Running AI inference on a device at or near the data source, instead of sending each task to a remote server, which can reduce delay and network dependence.

A network address where software sends requests to a service and receives responses according to an API's rules.

Energy and water use of AI means the electricity and water consumed in building and running AI systems; emissions are a separate impact.

CPUs, GPUs and TPUs are different kinds of computer chips that run calculations, with each designed for different patterns of work, speed, cost and energy use.

Information that describes other data, such as who created it, when, where, in what format, and under what rules it can be used.

Some values that should be present in a dataset are absent, leaving blanks that can distort analysis and model behaviour.

The systems and services that keep an AI model available so it can receive requests, run on computing hardware, and return outputs reliably and securely.

Data that contains errors, irrelevant variation or unclear signals, making patterns harder for people or models to detect reliably.

A control that caps how many requests or how much use a service accepts within a set time, to keep systems available, fair and affordable.

Choosing a smaller set of cases from a larger group so you can measure, inspect or train on it without using every case.

The ability of a system to handle more users, data or tasks without a serious drop in speed, quality, reliability or cost control.

Scaling laws describe the regular pattern that model performance often improves as data, parameters and compute increase, though gains usually slow and practical limits remain.

Information organised into predefined fields, often in rows and columns, so software can process it predictably.

Artificially generated records designed to resemble data from a real population or process for uses such as testing, training or sharing.

Information whose main content is not organised into consistent, predefined fields, such as emails, documents, images, audio and messages, and usually needs extra processing for analysis.

Designing tools, content and services so people with different disabilities can use them effectively, safely and with dignity.

Clear assignment of who must answer for an AI system’s effects, act when problems arise, and provide correction or remedy when duties are not met.

The unequal ability of people, organisations or countries to build, access, use or shape AI systems and their benefits.

AI ethics is the practice of checking whether AI systems are fair, safe, accountable and respectful of human rights before and during their use.

AI safety means designing, testing and using AI systems so they are less likely to cause harm, especially in situations where people could be affected.

Systematic unfair patterns in data, rules or model outputs can disadvantage some people or groups, especially when tools are used in important decisions.

The tendency to trust a system’s output too much and miss errors that a careful human review might have caught.

Confidentiality means limiting access to information so only authorised people can see, use or share it for a legitimate purpose.

A person freely agrees to a specific use of them or their information after getting clear, relevant information and a real chance to say yes or no.

The process of reviewing and acting on user content against rules to reduce harm, illegality and abuse while protecting lawful expression and human rights.

Data security means protecting digital or paper information from unauthorised access, change, loss or theft throughout its use, storage and sharing.

The gap between people who can reliably use digital tools and services and those who cannot, because of access, cost, skills, language, design or safety barriers.

Explainability means being able to show, in human-usable terms, why an AI system produced a particular output or pattern of outputs.

Designing and checking AI systems so they do not produce avoidable, unjust differences in how people or groups are treated or affected.

A guardrail is a rule, check or control that limits unsafe or unacceptable AI outputs or actions before people rely on them.

Human oversight is the processes that let people monitor AI use, review its outputs and effects, and intervene when needed.

An ongoing process to identify rights harms linked to an organisation’s actions or AI systems, address them, and account for the response.

A structured review used before or during AI use to identify who could be affected, what harms or rights risks may arise, and what controls or human decisions are needed.

Inclusion means designing and using AI so different people can access it, be represented in it, and avoid being unfairly left out or harmed.

Interpretability is the extent to which people can understand how a model works or produces outputs, sometimes by inspecting its internal parts.

Much AI depends on people who collect, label, review, filter and test data and outputs, often in outsourced or hidden forms of work.

People with real authority must have enough information and practical ability to shape, limit, change or stop AI-supported actions, especially in high-risk situations.

A principle and legal obligation to prevent unjustified harmful treatment based on protected traits or close proxies for them.

Privacy means controlling how information about identifiable people is collected, used, shared and kept, so they are not exposed, tracked or harmed unfairly.

A structured attempt to make an AI system fail, so people can find risks, harms and weak controls before real users are affected.

An approach to designing, using and governing AI so it is safe, fair, accountable and under appropriate human control.

Toxicity means abusive, hostile or degrading output, including insults, threats or dehumanising content aimed at people or groups.

Being open about where an AI system is used, what it does, what information it uses, and who remains responsible for its effects.

Access control sets rules about which people, systems or AI tools can view, use or change specific data, functions and resources.

An adversarial attack is a deliberate attempt to exploit an AI system so it fails or produces an attacker-chosen result.

A small, carefully chosen change to an input can push a model towards a wrong output while looking almost unchanged to a person.

AI security protects AI systems, their inputs, outputs and connected tools from manipulation, misuse, theft and unauthorised access.

Authentication verifies a claimed identity; authorisation checks what that requester is allowed to do.

Unauthorised copying of information out of a system, application or connected tool to a place an attacker or unapproved user can access.

An attack that quietly corrupts training data by adding, changing or removing examples, causing wrong patterns or a hidden trigger in the model.

A malicious instruction in external content tries to make a model follow it instead of the user’s or system’s intended task.

A jailbreak is a prompt or sequence of prompts designed to make an AI assistant ignore or work around its safety and behaviour limits.

Model extraction tries to copy model behaviour, while model inversion uses model information to reconstruct training-data features or examples.

A text-based attack that tries to make an AI assistant follow hidden or malicious instructions instead of the rules and task it was meant to follow.

Sandboxing runs a model, tool or file inside a restricted environment so harmful actions cannot easily reach the wider system.

The practice of storing and using passwords, tokens and keys safely so people and software can reach services without exposing those credentials.

Use of AI tools outside approved organisational processes, especially where data, security, legal or oversight rules are bypassed.

A structured check of how an AI system is governed, tested and used, producing evidence-based confidence that stated requirements are met and risks are controlled.

The rules, institutions and processes used to guide how AI is built, used and overseen so its benefits and risks are handled in the public interest.

A planned process for strengthening people, organisations and systems so they can carry out tasks well, sustainably and with less outside help.

A way of adopting text when no delegation present formally objects, so the chair can move forward without holding a vote.

Shared digital building blocks and reusable digital tools that help countries deliver public services, especially when they are designed for broad access and interoperability.

A document symbol is the official code used to identify and retrieve a specific United Nations document, while prefixes such as A/RES/ show what kind of document it is and where it came from.

A tool, model or dataset can serve beneficial civilian purposes and also be adapted for harmful, coercive or military uses.

The main UN forum where all Member States meet to debate issues, negotiate texts and adopt resolutions on matters including digital technology and AI.

A United Nations-led process for bringing governments and other relevant actors together to discuss shared approaches for managing AI across borders.

A UN-agreed framework that sets shared principles and actions for a safer, more inclusive and more coordinated digital future, including how digital technologies and AI should be governed globally.

A proposed UN-linked expert body that would review evidence on AI and provide independent scientific advice for global discussions and cooperation.

A document approved through a recognised international standards process that sets common requirements, guidance or methods for repeated use.

The ability of systems or organisations to exchange information and use it correctly under compatible rules.

Weapons that can select and apply force to targets with reduced or no direct human control at the moment of attack, raising legal, ethical and security concerns.

An authorised task or scope of work formally given to a person, group or organisation by a competent authority.

A sovereign country that belongs to an international organisation and can take part in its formal meetings, negotiations and decisions.

A way of shaping rules or shared practices with several kinds of actors, such as governments, companies, experts, civil society and affected communities.

A way of working where several countries address a shared issue through a common framework, negotiation or joint action rather than acting alone.

A formally issued or adopted text that sets shared principles, expectations or guidance for behaviour, often without creating the same binding legal duties as a regulation or treaty.

A UN-wide political agreement that sets shared priorities and follow-up actions for future global challenges, including digital cooperation and the governance of artificial intelligence.

A country’s official team at the United Nations or another international organisation represents its government in meetings, negotiations and documents on a continuing basis.

A formally counted vote where each eligible member’s position is entered into the official record, so others can see who supported, opposed or abstained.

A regulation is a binding rule made by a competent authority under rule-making powers granted by law for those within its scope.

A regulatory sandbox is a limited testing programme run by a public authority under defined conditions to learn, manage risks and shape future rules.

A formal text adopted by a UN body or other assembly to record a shared position, request action or set next steps.

A way to set stricter rules for uses more likely to cause serious harm, while applying lighter duties to lower-risk uses.

The main administrative body that supports the United Nations by preparing documents, organising meetings, carrying out mandates and helping Member States work through the UN system.

The UN body with primary responsibility for international peace and security can adopt decisions, including resolutions, that carry special political and legal weight.

A shared UN framework of 17 goals that countries agreed to use as a guide for improving people’s lives and protecting the planet by 2030.

An approach to rule-making that sets obligations by function, outcome or risk, rather than by naming one specific technology.

A terms of reference document sets the job, scope, roles, timeline and expected outputs for a group, project or process.

A United Nations change agenda that aims to make the organisation more effective, skilled, networked, data-informed and digitally capable for today’s problems.