Topic library

Artificial intelligence is software that performs tasks normally associated with human judgment, such as recognizing patterns or making predictions, rather than following a fixed set of manual rules.

Rule-based automation carries out a task using a fixed set of manually written if-then instructions, with no learning from data involved.

Machine learning is a branch of AI where a system improves its performance on a task by learning patterns from example data, rather than by explicit programming.

Deep learning is a form of machine learning that uses neural networks with many stacked layers to learn increasingly abstract patterns directly from raw data.

Generative AI refers to models that produce new content, such as text, images or audio, rather than only classifying or scoring existing input.

Narrow AI describes a system built and trained to perform one specific task well, such as translation or image classification, with no ability to generalize beyond it.

Artificial general intelligence describes a hypothetical system that could learn and perform any intellectual task a human can, across domains, without being retrained for each one.

AI washing is marketing a product as 'AI-powered' when it uses little or no actual learned-model behavior, to make it sound more advanced than it is.

A model input is the data given to a model at the moment it is asked to produce a result, such as a prompt, image or table of numbers.

A model output is the result a model produces after processing its input, such as generated text, a predicted label, or a probability score.

A token is the basic unit of text a language model reads and generates, usually a word, part of a word, or punctuation mark, not always a full word.

A context window is the maximum amount of text, measured in tokens, that a language model can consider at once when generating a response.

A prompt is the text input a user provides to instruct a language model on what task to perform or what to generate.

A system prompt is a set of instructions given to a language model before the user's message, defining its role, tone, or rules for the whole conversation.

Temperature is a setting that controls how random or predictable a language model's output is when choosing its next word.

A hallucination is a confident, fluent statement produced by a model that is factually wrong or not supported by any real source.

A large language model is a neural network trained on vast amounts of text to predict and generate language, with enough parameters to capture broad patterns in grammar, facts and style.

Next-token prediction is the core mechanism where a language model repeatedly predicts the single most likely next token given everything generated so far.

A chatbot is a conversational interface that responds to user messages in natural language, without necessarily taking any action beyond replying.

An AI assistant is a conversational system that can use tools, such as searching the web or checking a calendar, in addition to replying in natural language.

An AI agent is a system that can plan a sequence of steps and take actions toward a goal with limited ongoing human direction, rather than only responding to one request at a time.

Information retrieval is the general task of finding relevant documents or data in response to a query, independent of any specific search technique.

Keyword search finds documents by matching the exact words or phrases in a query against the words present in each document.

Semantic search finds documents based on meaning rather than exact wording, by comparing the underlying concepts of a query and each document.

A multimodal model is a single system that can process or generate more than one type of data, such as text and images together, rather than only one.

Pretraining is the initial stage where a model learns general language patterns from a huge, broad dataset, before any task-specific adjustment happens.

Fine-tuning is additional training performed on a pretrained model using a smaller, more specific dataset to adjust its behavior for a particular use case.

Instruction tuning is fine-tuning a model on examples of instructions paired with the desired response, so it learns to follow user requests rather than just continue text.

Preference tuning adjusts a model using human judgments about which of several candidate responses is better, to steer its output toward preferred styles and behaviors.

Reranking is a second pass over an initial set of retrieved results that reorders them by a more precise relevance score before they are used or shown.

A transformer is a neural network architecture that processes an entire sequence at once using self-attention, instead of reading it one element at a time like earlier models.

Self-attention is the mechanism that lets a transformer weigh how relevant every other element in a sequence is to each element it is currently processing.

Positional encoding is information added to each token so a transformer knows the order of elements in a sequence, since self-attention alone does not track position.

An embedding is a numerical vector representing an item, such as a word or document, so that related items can be placed near one another in a representation space.

Cosine similarity measures how closely two vectors point in the same direction, giving a score used to judge how similar two embeddings are in meaning.

Retrieval-augmented generation retrieves relevant documents at request time and includes them in a model's prompt, so its answer is grounded in that supplied content.

A vector database stores embeddings and can quickly find the ones most similar to a given query vector, even across millions of stored items.

Tool calling is a model's ability to request that a specific external function be run, with structured arguments, then use its result to continue responding.

An agent loop is the repeated cycle of a model deciding an action, an application executing it, and the result being fed back to the model before it decides the next action.

Agent memory is information an agent retains across steps or sessions, beyond what fits in a single prompt, so earlier context can inform later decisions.

An AI benchmark is a standardized dataset and scoring method used to measure and compare how well different models perform on a specific task.

Human evaluation is judging model output quality by having people directly review and rate responses, rather than relying only on automated scoring.

A vision-language model is trained to understand both images and text together, so it can answer questions about an image or describe it in natural language.

Context compression reduces a long piece of text into a shorter representation that preserves the information a model needs, so it fits within a limited context window.

Model serving is the infrastructure that receives a request, runs it through a trained model, and returns the result, making a model usable in a live application.

Inference latency is the time between sending a request to a model and receiving its response, a key factor in how responsive an AI product feels.

An AI guardrail is a check applied before or after a model generates output, to block or adjust responses that violate defined safety or policy rules.

A model gateway is a service layer that routes application requests to one or more underlying models, handling concerns like fallback, logging and rate limiting in one place.

A Mixture-of-Experts model contains several specialised neural-network components but activates only selected experts for each input.

Expert routing is the mechanism that decides which expert components in a Mixture-of-Experts model process each token or example.

A long-context model is designed to process very large context windows, often spanning hundreds of thousands of tokens, without requiring the input to be chunked.

Test-time compute is additional computation a model spends while answering a specific question, such as generating and comparing multiple reasoning attempts, rather than during training.

Knowledge distillation trains a smaller student model to reproduce the behavior of a larger teacher model, aiming to keep most of its capability at lower cost.

Quantisation reduces the numerical precision used to store a model's parameters, shrinking its size and speeding up computation at some cost to accuracy.

ML problem formulation is turning a business or research question into a precise prediction task, with a defined input, output, and success measure.

A training example is one single record of data, made of its features and (for supervised learning) its label, used to teach a model.

A feature is one measurable attribute of a training example that a model uses as input to make its prediction.

A label is the correct answer attached to a training example that a supervised model is trained to predict.

A dataset is the full collection of training examples gathered for a machine learning task, before it is split into training, validation and test portions.

Supervised learning trains a model on examples that each include a known correct label, so the model learns to predict that label for new examples.

Unsupervised learning finds structure or patterns in data that has no labels, such as grouping similar examples together.

Reinforcement learning trains a system to choose actions that maximize a cumulative reward signal, learned through trial and error rather than from labeled examples.

Classification is a supervised learning task where the model predicts which category, out of a fixed set, an example belongs to.

Regression is a supervised learning task where the model predicts a continuous numeric value rather than a category.

The training set is the portion of a dataset actually used to fit a model's parameters during learning.

The validation set is a portion of data held out from training, used to tune choices like hyperparameters and compare models during development.

The test set is a portion of data held out entirely from training and tuning, used only once for a final, honest estimate of model performance.

Data leakage happens when information that would not be available at real prediction time accidentally influences training, making evaluation results look better than they truly are.

Generalisation is how well a model's learned patterns apply to new data it has never seen, rather than only the data it was trained on.

Overfitting happens when a model learns patterns specific to its training data, including its noise, so closely that it performs worse on new, unseen data.

Underfitting happens when a model is too simple to capture the real patterns in the data, performing poorly on both training and new data.

Accuracy is the proportion of a classifier's predictions that are correct out of all predictions made.

Precision is the proportion of a model's positive predictions that are actually correct, out of all the times it predicted positive.

Recall is the proportion of all actual positive cases that a model correctly identifies as positive.

The F1 score is a single metric combining precision and recall into one number, using their harmonic mean, so both must be reasonably high for the score to be high.

A confusion matrix is a table showing how many predictions of each class matched or did not match the true class, breaking results into correct and incorrect predictions of each type.

Mean absolute error is the average size of a regression model's prediction mistakes, ignoring whether each mistake was too high or too low.

Mean squared error is the average of the squared differences between predicted and true values, which penalizes large errors more heavily than small ones.

A loss function measures how far a model's prediction is from the true value on a single example, and training adjusts the model to make this value smaller on average.

The bias-variance trade-off describes how a model's error splits between bias, from being too simple to capture true patterns, and variance, from being too sensitive to noise in the training data.

A classification threshold is the probability cutoff above which a model's predicted score is converted into a positive class prediction.

An ROC curve plots a classifier's true positive rate against its false positive rate across every possible classification threshold.

ROC-AUC is the area under the ROC curve, a single number summarizing how well a model ranks positive examples above negative ones across all thresholds.

Gradient descent is an optimization method that repeatedly adjusts a model's parameters in the direction that most reduces the loss function.

The learning rate controls how large a step gradient descent takes when adjusting a model's parameters at each iteration.

Linear regression predicts a continuous value as a weighted sum of input features, learning one coefficient per feature that best fits the training data.

Logistic regression predicts the probability of a binary outcome by passing a weighted sum of features through a function that squashes the result between 0 and 1.

A decision tree predicts an outcome by asking a sequence of yes/no questions about feature values, branching until it reaches a final prediction.

A random forest is an ensemble of many decision trees, each trained on a different random subset of data and features, whose predictions are averaged or voted together.

Gradient boosting builds an ensemble of decision trees one at a time, where each new tree is trained specifically to correct the errors of the trees built so far.

Ensemble learning combines predictions from multiple models into one, aiming for better and more stable performance than any single model alone.

Clustering groups examples into sets based on similarity, without any predefined labels for what each group represents.

k-means is a clustering algorithm that assigns each point to the nearest of k centroids, then repeatedly recalculates those centroids until assignments stop changing.

Feature scaling transforms numeric features onto a common range or distribution, so features measured in different units contribute fairly to a model.

One-hot encoding converts a categorical feature into several binary columns, one per possible category, each marking whether that category applies to a given example.

Missing-value imputation fills in a reasonable estimate for data points that are missing, so a model can use a row instead of discarding it entirely.

Feature engineering is creating new input features, or transforming existing ones, to better expose the patterns a model needs to learn.

Cross-validation repeatedly splits data into different training and validation portions, evaluating a model on each split and averaging the results for a more reliable estimate.

A hyperparameter is a setting chosen before training begins, such as learning rate or tree depth, that controls how a model learns rather than being learned from data itself.

Hyperparameter tuning is systematically searching over different hyperparameter values to find a combination that performs best on validation data.

Regularisation adds a penalty to a model's training objective that discourages overly complex patterns, helping reduce overfitting.

Class imbalance occurs when one class in a classification dataset is far more common than another, which can bias a model toward always predicting the majority class.

Resampling addresses class imbalance by adjusting the training data itself, either duplicating minority-class examples or removing majority-class examples, to better balance the classes.

Feature importance measures how much each input feature contributes to a model's predictions overall, ranking features by their typical influence.

A SHAP value attributes a model's specific prediction for one example to each of its input features, showing how much each feature pushed the prediction up or down.

An artificial neuron computes a weighted sum of its inputs, adds a bias term, and passes the result through an activation function to produce one output value.

An activation function is a nonlinear transformation applied to a neuron's weighted sum, allowing a network of neurons to represent more than simple linear relationships.

Backpropagation is the algorithm that efficiently computes how much each parameter in a neural network contributed to its overall error, so gradient descent knows how to adjust each one.

A convolution slides a small learned filter across an input, such as an image, computing a weighted sum at each position to detect a specific local pattern.

A convolutional neural network stacks layers of convolutions and pooling to progressively build from simple local patterns, like edges, to complex ones, like whole objects.

A recurrent neural network processes a sequence one element at a time, carrying a hidden state forward from each step to the next to retain information about earlier elements.

An attention mechanism lets a model directly weigh the relevance of every other element in a sequence to the one it is currently processing, rather than passing information step by step.

Representation learning is a model automatically learning useful internal features from raw data, rather than a person manually engineering those features beforehand.

An experiment hypothesis is a specific, testable claim about what change is expected to improve a model, stated before an experiment is run.

A control variable is anything kept identical between two experiment runs, such as the data split or random seed, so any performance difference can be attributed to the one thing actually being tested.

A random seed is a starting value that makes a process involving randomness, such as data shuffling or weight initialization, produce the exact same sequence every time it runs.

Experiment tracking is systematically recording each training run's configuration, data, and results, so past experiments can be compared and reproduced later.

Data versioning tracks changes to a dataset over time, so a specific model can always be linked back to the exact data it was trained or evaluated on.

A model registry is a system that stores trained models along with their metadata, such as version, training data, and performance, so specific models can be found and deployed reliably.

Batch inference runs a trained model over a large collection of examples all at once on a schedule, rather than responding to individual requests as they arrive.

Online inference runs a trained model on a single request as it arrives, returning a prediction immediately for that one input.

Concept drift is when the real-world relationship between input features and the correct outcome changes over time, making a previously accurate model gradually less reliable.

Data drift is when the statistical distribution of a model's input data changes over time compared to what it was trained on, even if the underlying relationship to the outcome stays the same.

Model monitoring is continuously tracking a deployed model's inputs, outputs and performance over time, to catch drift or degradation before it causes serious harm.

The Domain Name System translates human-readable domain names, like example.com, into the numeric IP addresses computers use to locate each other on a network.

An IP address is a numeric identifier assigned to a device on a network, used to route data to and from that specific device.

A URL is a structured text address that specifies a protocol, a host, and a path, telling a browser exactly what resource to request and how.

An HTTP request is a message a client sends to a server asking it to perform an action, such as retrieving or submitting data, following the HTTP protocol's structure.

An HTTP response is the message a server sends back after receiving a request, including a status code and, usually, the requested data.

A web browser is a program that sends HTTP requests, receives responses, and renders the returned HTML, CSS, and JavaScript into an interactive page.

A web server is a program that listens for incoming HTTP requests and returns an appropriate HTTP response, often by running application logic or reading files.

An HTML element is a single structural unit of a webpage, defined by a tag, that tells the browser what kind of content it contains, such as a paragraph or a link.

Semantic HTML uses elements whose tag names describe their actual meaning, like `<nav>` or `<article>`, rather than generic containers for everything.

A CSS selector is a pattern that identifies which HTML elements a set of style rules should apply to.

The CSS box model describes how an element's content, padding, border, and margin combine to determine its total rendered size and spacing.

CSS specificity is the set of rules that decide which style wins when multiple CSS rules target the same element with conflicting values.

Flexbox is a CSS layout mode that arranges elements along a single row or column, distributing space and alignment between them automatically.

CSS Grid is a layout mode that arranges elements across both rows and columns at once, defined as an explicit grid structure.

Responsive design is building a page's layout to adapt to different screen sizes, using techniques like flexible layouts and conditional styles based on viewport width.

A JavaScript variable is a named storage location that holds a value, which can be referenced and, depending on how it was declared, reassigned later.

A JavaScript function is a reusable block of code that can accept input values as parameters and optionally return a result.

A JavaScript object is a collection of key-value pairs, letting related data and behavior be grouped together under one variable.

An array is an ordered list of values in JavaScript, each accessible by its numeric position, or index, starting at zero.

A Promise is a JavaScript object representing a value that isn't available yet but will eventually resolve successfully or fail with an error.

async/await is syntax that lets code wait for a Promise to settle using a sequential, top-to-bottom style, instead of chaining '.then()' calls.

The Document Object Model is the browser's in-memory tree representation of a page's HTML, which JavaScript can read and modify to change what's displayed.

A browser event is a signal fired when something happens on a page, such as a click or a key press, that JavaScript code can listen for and react to.

A Git repository is a folder tracked by Git that stores a project's full history of changes, allowing any past version to be recovered.

A Git commit is a saved snapshot of a project's files at a specific point, recorded with a message describing what changed.

A Git branch is an independent line of commits, letting changes be developed separately from the main history until they're ready to be merged in.

A merge conflict happens when Git cannot automatically combine changes from two branches because both modified the same part of a file differently.

An API is a defined set of rules that lets one piece of software request functionality or data from another, without needing to know its internal implementation.

An API endpoint is a specific URL that an API exposes for performing one particular operation, such as fetching or creating a specific kind of data.

An HTTP method is the verb in an HTTP request, such as GET or POST, that indicates the type of action being requested on a given endpoint.

An HTTP status code is a three-digit number in an HTTP response that indicates whether a request succeeded, failed, or requires further action.

JSON is a lightweight text format for representing structured data, such as objects and lists, that is easy for both humans and machines to read and write.

The frontend is the part of an application that runs in the user's browser or device, responsible for what the user sees and directly interacts with.

The backend is the part of an application that runs on a server, handling business logic, data storage, and requests that the frontend can't safely or capably handle itself.

A full-stack application combines both frontend and backend development, with one team or codebase responsible for the complete user-facing and server-side experience.

A TypeScript type describes the shape of a value, such as which properties an object has or what kind of data a variable can hold, checked before the code ever runs.

Type inference is TypeScript automatically determining a value's type from context, without requiring the type to be written explicitly every time.

A React component is a reusable function that describes a piece of user interface, returning what should be displayed based on its inputs and internal state.

Props are the input values passed into a React component from its parent, letting the same component render differently depending on what it's given.

State is data a React component manages internally that can change over time, causing the component to re-render whenever it's updated.

A React hook is a special function, like useState or useEffect, that lets a component use features such as state or side effects without being written as a class.

An effect is code in a React component that runs in response to rendering, used for tasks like fetching data or subscribing to an external event, outside the normal render flow.

A Next.js route is a URL path automatically mapped to a specific file in the project, based on that file's location within the app's folder structure.

A Server Component in Next.js renders on the server and sends only the resulting HTML to the browser, without shipping its own JavaScript to run client-side.

A Client Component in Next.js runs in the browser, allowing it to use interactivity features like state and event handlers that Server Components cannot.

A Route Handler is a Next.js file that responds to HTTP requests directly, letting a project expose its own API endpoints alongside its pages.

Server-side rendering generates a page's HTML on the server for each request, so the browser receives a fully formed page rather than building it entirely client-side.

Static rendering generates a page's HTML once at build time, then serves that same pre-built file for every future request until it's rebuilt.

A relational database stores data in structured tables with defined relationships between them, queried using a language like SQL.

A database table stores records as rows, each with the same set of named columns defining what data every row contains.

A primary key is a column, or set of columns, whose value uniquely identifies each row in a database table.

A foreign key is a column in one table that references another table's primary key, linking a row to a related row elsewhere.

A SQL query is a statement written in SQL that requests or modifies data in a relational database, such as retrieving matching rows or updating values.

A database index is a separate structure built on one or more columns that lets the database find matching rows much faster than scanning the whole table.

A database transaction groups multiple operations so they either all succeed together or all fail together, leaving the database in a consistent state either way.

Authentication is verifying that a user is who they claim to be, typically by checking credentials like a password or a token.

Authorisation is determining what an already-identified user is permitted to do or access, after authentication has confirmed who they are.

A session is a period during which a server recognizes a user as logged in, usually tracked through a token stored on the user's device between requests.

A cookie is a small piece of data a server asks the browser to store, then automatically sends back with future requests to that same site.

Password hashing transforms a password into a scrambled, one-way value before storing it, so the original password can't be recovered even if the stored data is exposed.

A unit test checks that one small, isolated piece of code, like a single function, behaves correctly for a given input.

An integration test checks that multiple parts of a system, such as an API and a real database, work correctly together, not just individually.

An end-to-end test verifies a complete user flow through a real running application, from the interface down through the backend, as an actual user would experience it.

A test double is a stand-in object used in a test to replace a real dependency, such as a database or external service, so the test stays fast and isolated.

A design pattern is a named, reusable solution to a commonly occurring software design problem, giving developers a shared vocabulary for a familiar structure.

Separation of concerns is organizing code so each part handles one distinct responsibility, keeping unrelated logic, like data storage and display formatting, apart from each other.

Dependency injection is supplying a piece of code with the other components it needs from outside, rather than having it create or look up those components itself.

A distributed system is an application made of multiple separate computers working together over a network, rather than running as one single program on one machine.

A load balancer distributes incoming requests across multiple servers, so no single server has to handle all the traffic alone.

A cache stores a copy of data that's expensive to compute or fetch, so future requests for the same data can be served faster from the stored copy instead.

Cache invalidation is removing or updating cached data once it's known to be outdated, so future requests don't keep receiving stale results.

A message queue holds messages sent from one part of a system until another part is ready to process them, decoupling the sender from needing an immediate response.

Idempotency is a property of an operation where running it multiple times has the same effect as running it once, even if it's accidentally repeated.

Transport Layer Security is a protocol that encrypts data sent between a client and a server, so it can't be read or altered by anyone intercepting the connection.

An injection attack occurs when untrusted user input is inserted directly into a command or query, letting an attacker change its meaning to do something unintended.

Cross-site scripting is an attack where malicious script gets injected into a page and runs in another user's browser, often because untrusted input was displayed without proper handling.

Cross-site request forgery tricks a logged-in user's browser into sending an unwanted request to a site they're authenticated with, without the user intending it.

An access-control failure occurs when a system fails to properly check that a user is authorised to view or modify a specific piece of data, letting them access something that isn't theirs.

A secret is sensitive configuration data, like an API key or database password, that must be kept out of source code and version control to avoid exposure.

A container packages an application together with everything it needs to run, so it behaves the same way regardless of the machine it's deployed on.

A CI pipeline automatically builds and tests code every time changes are pushed, catching integration problems early instead of discovering them later.

A CD pipeline automatically takes code that has passed its checks and deploys it to a live environment, reducing manual steps in releasing changes.

A cloud region is a specific geographic location where a cloud provider runs its data centers, letting an application be deployed physically closer to its users.

A content delivery network stores copies of static content across many geographically distributed servers, serving each user from whichever copy is physically closest to them.

Latency is the time delay between making a request and receiving the start of its response.

Throughput is the total amount of work a system can complete in a given period, such as the number of requests it can handle per second.

A performance budget is a defined limit, such as a maximum page load time or file size, that a team commits to staying within as they build a product.

A log is a timestamped record of a discrete event that happened in a running system, such as a request being handled or an error occurring.

A metric is a numeric measurement tracked over time, such as requests per second or error rate, used to understand a system's overall behavior and health.

A trace follows a single request as it moves through multiple services in a system, showing how long each step along the way took.

Observability is how well you can understand a system's internal state and diagnose problems just from its external outputs, like logs, metrics, and traces.

A service-level indicator is a specific, measured metric used to judge how well a service is performing, such as the percentage of requests served successfully.

A service-level objective is a target value set for a service-level indicator, such as '99.9% of requests succeed', that a team aims to meet.

An error budget is the amount of allowable failure implied by a service-level objective, such as the total downtime still permitted while still meeting a 99.9% target.

An incident is an unplanned event where a system fails to meet its expected level of service, requiring a response to detect, mitigate, and resolve it.

A post-incident review is a structured examination of what happened during an incident, why it happened, and what changes could prevent or reduce it happening again.

Responsible technology is building and operating systems with deliberate attention to their real-world effects on people, not just their technical performance.

AI ethics is the study of what obligations exist when building and deploying AI systems, covering questions like fairness, harm, and who is accountable for outcomes.

A sociotechnical system is a technology considered together with the people, processes, and institutions that shape how it's actually used and its real effects.

A stakeholder is anyone with a meaningful interest in how a system behaves, including its users, operators, and people affected by its decisions even if they never use it directly.

An affected community is a group of people who experience a system's consequences, whether or not they were consulted in its design or have any control over it.

Accountability is having a clearly identifiable person or organization responsible for a system's decisions, able to answer for its outcomes.

Contestability is a system being designed so an affected person can challenge or question a decision made about them, rather than the decision being final and unexplained.

A remedy is the corrective action taken when a system has caused harm, such as reversing a wrong decision or compensating someone affected.

Dataset bias is a systematic imbalance or gap in a training dataset that causes a model to perform unevenly across different groups or situations.

Sampling bias occurs when the way data is collected systematically over- or under-represents certain groups compared to the real population the model will be used on.

Label bias occurs when the 'correct answers' used to train a model themselves reflect human prejudice or inconsistent judgment, rather than an objective ground truth.

Measurement bias occurs when a feature is measured less accurately or consistently for some groups than others, distorting what the model actually learns about them.

A proxy variable is a feature that closely correlates with a sensitive attribute, like race or gender, letting a model indirectly rely on that attribute even if it's not used directly.

Historical bias occurs when a dataset accurately reflects a real past pattern of unfair treatment, so a model trained on it learns to continue that same unfairness going forward.

Personal data is any information that relates to an identified or identifiable individual, whether directly, like a name, or indirectly, like a device identifier.

Privacy risk is the potential for personal data to be misused, exposed, or linked in ways that harm the individual it describes.

Data minimisation is collecting and retaining only the personal data actually necessary for a specific purpose, rather than gathering everything that might be useful someday.

Purpose limitation is using personal data only for the specific purpose it was originally collected for, rather than repurposing it for something unrelated later.

Consent is a person's informed, freely given agreement to a specific use of their personal data, understood clearly enough to be a real choice.

Anonymisation is processing data so it can no longer be linked back to a specific individual, even by combining it with other available information.

Pseudonymisation replaces directly identifying details with a consistent artificial identifier, so the same person's records can still be linked together without exposing their real identity directly.

Differential privacy adds carefully calibrated statistical noise to data or query results, so no individual's presence in the dataset can be confidently determined from the output.

Federated learning trains a shared model across many devices without any individual device's raw data ever leaving that device, sharing only model updates instead.

Transparency is making meaningful information available about how a system works and makes decisions, to whoever needs it to trust or evaluate the system.

Explainability is a system's ability to provide a human-understandable reason for a specific decision it made, even if that explanation is generated after the fact.

Interpretability is how directly a person can understand a model's internal logic itself, rather than only receiving a separately generated explanation of its output.

Human oversight is a person retaining meaningful ability to review, question, or override an automated system's decisions, rather than the system operating entirely unchecked.

Human-in-the-loop describes a system where a person must actively review and approve a decision before it takes effect, rather than the system acting on its own.

Human-on-the-loop describes a system that acts automatically but is monitored by a person who can intervene or halt it if something goes wrong, rather than approving each action in advance.

Automation bias is the tendency for a person overseeing a system to over-trust its output and under-scrutinize it, weakening the value of human oversight in practice.

AI literacy is having enough practical understanding of how AI systems work and where they fail, to use, evaluate, or oversee them appropriately.

AI risk is the potential for an AI system to cause harm, evaluated in terms of both how likely that harm is and how severe it would be if it occurred.

Likelihood is how probable it is that a specific risk will actually occur, one of the two core factors used to evaluate a risk's overall significance.

Severity is how serious the consequences would be if a specific risk actually occurred, the second core factor used alongside likelihood to evaluate a risk.

Risk tolerance is the level of risk an organization has explicitly decided it's willing to accept for a given system, rather than eliminate entirely.

A risk register is a maintained record listing a system's identified risks, their likelihood and severity, and the status of any mitigation for each.

A risk control is a specific measure put in place to reduce a risk's likelihood or severity, such as a guardrail, review step, or monitoring alert.

An algorithmic impact assessment is a structured process for evaluating a system's likely effects on people before it's deployed, covering risks, affected groups, and mitigations.

A risk owner is the specific person responsible for tracking a given risk, ensuring its controls stay effective, and escalating if it changes.

Demographic parity is a fairness criterion requiring a model's positive prediction rate to be roughly equal across different demographic groups.

Equal opportunity is a fairness criterion requiring that qualified individuals across different groups have an equal chance of receiving a positive prediction.

Equalised odds is a fairness criterion requiring both the true positive rate and the false positive rate to be equal across groups.

Calibration is a fairness property where a model's predicted probability actually matches the real observed outcome rate for every group, not just overall.

A fairness trade-off is the mathematical fact that, except in special cases, a model generally cannot simultaneously satisfy every fairness criterion at once, forcing a choice of which to prioritize.

Subgroup evaluation is measuring a model's performance separately for each demographic group, rather than relying only on one overall aggregate score.

Intersectional evaluation measures a model's performance for people belonging to combinations of multiple groups at once, not just one group category at a time.

A model card is a short, standardized document describing a model's intended use, performance across groups, limitations, and known risks.

A dataset datasheet documents how a dataset was collected, what it contains, and any known gaps or limitations, so others can judge whether it's fit for a given use.

A system card documents an entire deployed system, including how its models, data, and human processes work together, not just one model in isolation.

A decision record documents a specific choice made during a system's design or deployment, along with the reasoning and alternatives considered at the time.

Audit evidence is the concrete documentation, logs, and records collected to demonstrate that a system actually followed its stated processes and controls.

Content provenance is verifiable information about how a piece of media, like an image or video, was created and whether it was altered, attached to the content itself.

Content Credentials is a specific technical standard for attaching tamper-evident provenance information, like editing history, directly to a media file.

A digital watermark is a signal embedded directly into a piece of media's content, detectable by software, that can indicate the content was generated or modified by AI.

A deepfake is synthetic media, often video or audio, generated to convincingly depict a real person saying or doing something they did not actually say or do.

A responsible product requirement is a specific, checkable condition, like a fairness threshold or a required review step, that a product must meet before it ships.

An AI governance policy is an organization's documented rules for how AI systems must be reviewed, approved, and monitored before and after deployment.

An AI governance committee is a designated group of people responsible for reviewing and approving AI systems against an organization's governance policy.

An AI audit is an independent examination of a deployed system to verify it actually complies with stated policies, performance claims, and legal requirements.

Incident reporting is a defined process for recording and escalating when an AI system causes unexpected harm or behaves outside its intended bounds.

A sunset criterion is a predefined condition under which a deployed system must be retired, paused, or replaced, decided in advance rather than only in reaction to a crisis.

Red teaming is deliberately trying to make a system fail or misbehave, using an adversarial mindset, to find weaknesses before real-world misuse does.

Adversarial testing systematically probes a system with inputs specifically crafted to trigger incorrect or harmful behavior, rather than testing with typical, expected inputs.

Safety evaluation is systematically checking a system against a defined set of harm categories before deployment, to confirm it behaves acceptably across each of them.

AI regulation is legally binding rules, created by a government or governing body, that constrain how certain AI systems may be built, deployed, or used.

The EU AI Act risk classification is a framework that sorts AI systems into risk tiers, such as unacceptable, high, limited, or minimal risk, each with different legal obligations.

A technical standard is a formally published specification, often from a standards body, defining a consistent way to build, test, or document a system's specific aspect.

A conformity assessment is a formal process for checking whether a system actually meets a required standard or regulation before it can be legally offered or deployed.

A real-world outcome is the actual effect a system has on people's lives, distinct from how accurately it performs on a technical benchmark or test dataset.

An output metric measures what a system directly produces, like predictions made or content generated, rather than the effect those outputs actually have.

An outcome metric measures the actual real-world change a system contributes to, such as improved health results, rather than just the volume of outputs it produced.

Distributional impact is how a system's real-world effects vary across different groups, rather than looking only at its average effect across everyone combined.

A feedback loop occurs when a system's own past decisions influence the data it later learns from, potentially reinforcing and amplifying an early pattern over time.

An appeal mechanism is a defined process letting someone affected by a system's decision formally challenge it and request reconsideration.