PECB Certified Artificial Intelligence Professional key terms

The capability of a machine-based system to produce outputs such as predictions, recommendations, or decisions that influence real or virtual environments, given explicit or implicit objectives and varying degrees of autonomy.

A family of techniques in which models learn patterns from data rather than following hand-coded rules, improving their performance on a task through training and adjustment of internal parameters.

A mathematical or computational structure whose parameters are set through training so that it maps inputs to outputs such as classifications, scores, or generated content.

The dataset used to fit a model's parameters, whose quality, representativeness, and provenance strongly shape the accuracy, fairness, and robustness of the resulting system.

The operational phase in which a trained model is applied to new inputs to produce outputs; it is governed by performance, latency, and monitoring requirements distinct from training.

Machine learning using multi-layer artificial neural networks that learn hierarchical feature representations directly from raw data such as images, audio, or text, typically requiring large datasets and compute.

A large model trained on broad data at scale that can be adapted to many downstream tasks through fine-tuning or prompting, rather than being built for a single purpose.

Systems that produce new content such as text, images, code, or audio in response to prompts, typically built on foundation models and raising specific risks around hallucination and provenance.

A foundation model trained on very large text corpora to predict and generate language, used for summarization, question answering, and code generation, and prone to confident but incorrect outputs.

The input instruction or context supplied to a generative model that conditions its output; prompt design and prompt injection are both functional and security concerns.

Further training of a pre-trained model on a smaller, task- or domain-specific dataset to adapt its behavior, which can also introduce new bias or erode previously applied safeguards.

A generated output that is fluent and plausible but factually wrong or unsupported by any source, a characteristic failure mode of generative models that requires verification controls.

A modeling failure in which a model captures noise specific to its training data and therefore performs poorly on new, unseen data, reducing generalisation in production.

The degree to which the reasons for a system's output can be understood by people, ranging from inherently interpretable models to post-hoc techniques that approximate feature influence.

The arrangement of human review, intervention, and stop capability over an AI system so that people retain meaningful control and can correct or halt harmful behavior.

A design pattern in which a person reviews or approves system outputs before they take effect, used where errors carry significant consequences or where regulation requires intervention.

The end-to-end phases of an AI system from conception, data collection, and development through deployment, operation, monitoring, and eventual decommissioning.

Degradation of model performance over time because the statistical relationship between inputs and outcomes changes in the real world after deployment.

The effect of uncertainty on objectives arising from AI systems, encompassing harm to people, legal and regulatory exposure, financial loss, and reputational damage.

A structured process that identifies, analyses, and evaluates risks by combining likelihood and impact estimates against defined criteria, producing prioritized treatment decisions.

The selection and implementation of options to modify risk, such as avoiding an activity, reducing likelihood or impact through controls, transferring it, or consciously accepting the residual.

The risk that remains after controls and treatments have been applied, which must be compared with the organization's tolerance and formally accepted by an accountable owner.

The amount and type of risk an organization is willing to pursue or retain in pursuit of its objectives, providing the benchmark against which individual AI use cases are judged.

A documented evaluation of how an AI system may affect individuals, groups, and society, covering rights, fairness, safety, and environment, usually performed before deployment and revisited periodically.

Systematic and unfair deviation in system outputs against particular groups, arising from data, labeling, feature selection, or objective design, and measurable through disaggregated performance testing.

The property, or set of competing formal criteria, by which an AI system's outcomes are judged equitable across groups; no single metric satisfies all definitions simultaneously.

The ability of a system to maintain acceptable performance and safety when inputs are noisy, unusual, or deliberately manipulated, including adversarial examples and distribution shift.

The risk of adverse consequences from decisions based on incorrect or misused model outputs, including errors in design, data, implementation, and use outside intended scope.

Independent testing and review confirming that a model or system meets its intended purpose and performance criteria before release and after significant change.

The policies, roles, structures, and processes by which an organization directs and controls its AI activities, aligning them with strategy, law, and ethical commitments.

A formal statement of principles and mandatory requirements governing acceptable AI use, approved by senior management and translated into standards, procedures, and controls.

The assignable obligation of a named role or function to answer for AI-related decisions and outcomes, supported by authority, resources, and documented evidence.

The practice of making AI systems, their purposes, data sources, and limitations visible and understandable to affected parties, regulators, and internal reviewers.

The organization-wide rethinking of products, processes, and business models through digital technology, of which AI adoption is one component requiring change management and capability building.

The managed introduction of AI capabilities into business processes, covering use-case selection, value case, skills, integration, and readiness of the operating model.

The structured approach to helping people transition to new ways of working, addressing skills, incentives, communication, and resistance so that technology investments deliver expected value.

The knowledge and skills that staff and users need to understand AI capabilities and limitations, use systems appropriately, and recognize when outputs require scrutiny or escalation.

The human tendency to over-trust automated outputs and reduce independent verification, which weakens the practical effect of oversight controls in AI-assisted decisions.

The protection of AI systems, their data, and their supply chains against threats such as data poisoning, model theft, evasion, and prompt manipulation throughout the lifecycle.

A deliberate attempt to cause an AI system to misbehave, for example by perturbing inputs to force misclassification or by crafting prompts that bypass safeguards.

Manipulation of training or fine-tuning data to corrupt model behavior, plant backdoors, or degrade performance, often introduced through unvetted third-party data sources.

An attack in which malicious instructions embedded in user input or retrieved content cause a generative system to ignore its intended constraints and perform unintended actions.

Unauthorized extraction or copying of a model's parameters or behavior, for instance through systematic querying of a deployed service, eroding intellectual property and control.

A structured analysis of what could go wrong, who might cause it, and which controls reduce the risk, adapted for AI to cover data, model, and supply chain threats.

The set of legal and technical measures safeguarding personal data through lawful basis, purpose limitation, minimization, security, and rights of data subjects.

The principle that privacy protections are embedded into systems from the outset rather than added afterwards, through default settings, data minimization, and early risk assessment.

Unauthorized disclosure of sensitive or personal information, in AI contexts occurring through training on confidential data, memorized model output, or prompt disclosure to third-party services.

Information relating to an identified or identifiable natural person, whose processing in AI systems triggers obligations such as lawful basis, transparency, and access rights.

Artificially generated records that mimic the statistical properties of real data, used to reduce privacy exposure or augment scarce datasets while carrying its own fidelity risks.