ISTQB CT-AI key terms

The branch of computing that builds systems able to perform tasks usually needing human intelligence, such as recognising images, understanding language, or choosing actions.

A software system in which some behaviour or output is produced by AI components, such as trained models, rather than only by explicitly written program logic.

An approach in which algorithms learn patterns from data and adjust themselves to build a model, instead of following rules that a person writes by hand.

A machine learning technique that stacks many layers of simple processing units so the system can learn complex patterns from large volumes of data.

A rule-driven system that stores human expertise in a knowledge base and applies it to give advice in a narrow domain, without learning from examples.

An international standard that extends the SQuaRE quality model with characteristics and guidance specific to AI-based systems, helping teams specify and evaluate their quality.

The degree to which an AI-based system produces the outputs that are right for the task, measured against agreed references or ground truth.

The ability of a model or system to keep giving acceptable results when inputs are noisy, unusual, corrupted, or deliberately crafted to cause failure.

The degree to which information about an AI system, such as its purpose, data, logic, and limits, is available so stakeholders can understand and question it.

The ability to give people meaningful reasons for an AI system's output, often using additional techniques that approximate how the model reached its answer.

The property that an AI system treats individuals and groups comparably, without unjustified differences in outcomes tied to protected characteristics such as gender or ethnicity.

The overall property of an AI system being safe, fair, reliable, transparent, and secure enough that people can reasonably depend on it.

The property that an AI-based system does not cause unacceptable harm to people, property, or the environment, including when it behaves in unexpected ways.

The ordered activities from defining a problem through data work, model building, deployment, monitoring, and retirement that create and maintain an ML system.

The process of running a learning algorithm over data so it adjusts a model's internal parameters and improves its ability to perform the intended task.

Using an already trained model to produce predictions or content for new inputs, without changing the model's learned parameters.

The known or reference answer attached to a data item, used to teach a supervised model and to judge whether its predictions are right.

A setting chosen before training that controls how a learning algorithm works, such as learning rate or tree depth, and is not learned from the data.

A failure where a model learns its training data too closely, including its noise, so it scores well there but badly on new data.

A change over time in the true relationship between inputs and correct outputs, so a model that was accurate gradually becomes less accurate.

A model's ability to give good results on data it has never seen, which is the real purpose of machine learning.

The examples used to fit a model's parameters during learning; their quality, coverage, and balance strongly shape what the model learns.

Data kept apart from training and used only to measure how well a finished model performs on cases it has not encountered.

The extent to which data is accurate, complete, consistent, timely, and relevant enough for the purpose it will serve.

A systematic distortion in a dataset that pushes a model toward skewed or unfair patterns, for example through unrepresentative sampling or labelling habits.

A mistake where information from the target or from test data reaches training, giving results that look excellent but fail on real new data.

A source of expected results used to judge whether an observed output is correct; for ML this is often labelled data or a derived property.

A technique for checking a model without expected outputs, by confirming that related inputs produce outputs related in a predicted way.

Of all the items a model labelled positive, the share that really are positive; high precision means few false alarms.

Of all the items that really are positive, the share the model found; high recall means few positive cases were missed.

An input deliberately modified in a way people barely notice but which causes a model to give a wrong or unsafe result.

Testing the code, data handling, and components around a model while it is being built, as distinct from measuring the model's predictive quality.

Checking that the chain of data preparation, training, and serving steps works end to end and handles bad data or changed formats correctly.

Running a new model beside the current one on live inputs, recording its outputs without acting on them, so behaviour can be compared safely.

AI that produces new content such as text, images, audio, or code by learning patterns from large collections of existing examples.

A very large neural network trained on huge amounts of text to predict and generate language, used for answering, summarising, translating, and similar tasks.

The instructions or input text given to a generative model, which strongly influence the response that comes back.

When a generative model states something false or invented with confidence, instead of admitting that it does not know.

An attack in which hidden instructions in input or retrieved content cause a model to ignore its intended instructions or reveal protected information.

Checks and controls around a generative model, such as filters, content rules, and output limits, that keep responses safe and within scope.