ISTQB CT-GenAI key terms

A class of machine learning models, typically based on neural networks, that learn patterns from training data and produce new content such as text, code, images, or test cases in response to a prompt.

A generative model trained on very large text corpora to predict and produce sequences of tokens. It can answer questions, write code, and generate test artifacts when given suitable prompts and context.

The basic unit of text an LLM processes, roughly a word fragment. Token counts determine context limits and, for hosted models, much of the usage cost of a request.

The instruction and supporting input given to a generative model to steer its output. Its clarity, specificity, and included context strongly influence the relevance and quality of what is produced.

The practice of deliberately structuring prompts, including role, task, context, constraints, and output format, to obtain more accurate and usable results from a generative model, often refined iteratively.

A confident but factually wrong or fabricated model output, such as an invented API, requirement, or defect. Testers must verify generated content against authoritative sources rather than trusting it.

The maximum amount of text, measured in tokens, that a model can consider in one interaction. Content beyond it is truncated or ignored, which can cause relevant details to be missed.

The dataset used to fit a model's parameters. Its coverage, quality, and biases shape what the model can generate, including whether it knows relevant domains, languages, or technologies.

A systematic skew in model outputs caused by imbalanced or unrepresentative training data or design choices. It can produce unfair or incomplete test ideas and must be considered when assessing results.

The property that the same prompt can yield different outputs on repeated runs, due to sampling settings and model behavior. It makes generated artifacts hard to reproduce exactly and requires versioning and review.

A sampling setting that controls output randomness. Lower values make responses more focused and repeatable; higher values increase variety and creativity but also the risk of irrelevant or incorrect content.

A technique that retrieves relevant documents from a knowledge source and supplies them to the model with the prompt, improving factual grounding and reducing hallucination for domain-specific tasks.

Anchoring model output in verifiable sources such as requirements, specifications, logs, or retrieved documents, so that generated content can be checked and traced rather than accepted on trust.

The deliberate review, correction, and approval of generative AI output by a qualified person. It is a core control for accuracy, safety, and accountability in AI-assisted testing activities.

The difficulty of determining the correct expected result for a test. Generative AI can propose expected outcomes, but its suggestions still require a trustworthy source to confirm correctness.

Using a generative model to summarize, decompose, and inspect requirements for ambiguity, gaps, and testability, producing candidate test conditions that a human then validates against the source.

A testable aspect of a component or system identified as relevant for testing. Generative AI can suggest candidate conditions from requirements, which testers review for completeness and correctness.

Producing test cases, including preconditions, steps, and expected results, from inputs such as requirements or user stories. AI-generated cases are drafts that need review, refinement, and traceability checks.

A black-box technique dividing inputs into classes expected to be handled the same way. Generative AI can propose partitions, but testers must confirm boundaries and business rules are correct.

A technique testing values at the edges of input partitions, where defects cluster. AI can enumerate candidate boundaries, yet testers must verify them against actual specifications and constraints.

A requirement, risk, or code area not exercised by existing tests. Generative AI can help identify such gaps by comparing test suites with requirements, subject to human confirmation.

The documented link between requirements, test conditions, and test cases. AI can help build and check these links, but the resulting matrix must be validated to avoid invented or missing references.

Creating input data for tests, including valid, invalid, and edge values. Generative AI can synthesize realistic datasets quickly, but privacy, masking, and representativeness must be verified.

Artificially produced data that mimics real data patterns without using actual records. It supports testing where real data is sensitive or scarce, provided its realism and coverage are assessed.

Protecting personal and confidential information from unauthorized exposure. Prompts to external AI services must exclude sensitive data unless approved safeguards and policies are in place.

Using generative AI to draft automated test scripts, page objects, or fixtures from descriptions or examples. Output must be reviewed, executed, and adapted to project standards before use.

Using tools and scripts to execute tests and compare outcomes automatically. Generative AI can accelerate script creation and maintenance, but reliability and maintainability still depend on engineering review.

An expression that identifies a UI element for automation, such as a CSS selector or XPath. AI can suggest locators, but fragile or ambiguous ones cause flaky tests and need validation.

A test that passes and fails inconsistently without code changes, often due to timing, environment, or unstable locators. AI can help diagnose patterns, but root causes require investigation.

Automation that detects broken element references and proposes or applies alternative locators at runtime. It reduces maintenance effort but needs monitoring so that real defects are not silently masked.

Interpreting results, logs, and failure output to determine causes. Generative AI can summarize failures and suggest likely root causes, which testers must confirm with evidence.

Determining the underlying reason a defect or failure occurred. AI can cluster similar failures and propose hypotheses, but conclusions must be verified against logs and code.

The danger that generated content reproduces protected material or that submitted input leaks confidential assets. Teams need policies on tool use, licensing, and review of AI-assisted output.

Communicating test progress, results, and quality status to stakeholders. Generative AI can draft summaries and dashboards text from raw results, but accuracy and context must be checked.

Drafting defect records with summary, steps, and evidence. AI can structure and clarify reports from notes or logs, yet testers must confirm reproducibility and remove speculative claims.

A quantitative measure such as defect density, pass rate, or coverage used to judge quality. AI can compute and explain metrics, but definitions and data sources must be verified.

The ongoing refinement of processes and artifacts based on feedback and results. AI can analyze retrospectives and defect trends to suggest improvements, which teams then evaluate and adopt.

Assessing a generative model's output quality for a testing task using defined criteria and sample cases. It helps decide whether and how to rely on the tool in a given context.

Including a small number of worked examples in a prompt so the model imitates the desired style or format. It often improves consistency of generated test artifacts.