CT-AI — AI Testing glossary

156 terms the exam expects you to know, defined in plain language.

A

A/B Testing(split testing): A statistical hypothesis testing method used to compare two versions of a system by exposing them to similar user groups and measuring a target metric.
AI-specific quality characteristics(ai-specific quality characteristics, robustness, transparency, fairness, safety, iso/iec 25059): Quality attributes particularly relevant to AI systems, including robustness, transparency, fairness, and safety.
Acceptance Criteria for AI Systems(acceptance criteria for ai systems, ai acceptance criteria): Conditions that an AI system must meet to be accepted by stakeholders, including functional metrics and non-functional requirements like robustness and explainability.
Accuracy(accuracy): A classification metric computed as (TP + TN) / (TP + TN + FP + FN), measuring overall correctness.
Accuracy Constraint(accuracy): A constraint that verifies data correctness against a trusted source.
Adjusted R-squared(adjusted r², adjusted r-squared): A modified version of R² that penalizes the addition of irrelevant predictors, preferred for model comparison.
Adversarial Examples(adversarial examples, adversarial example): Inputs crafted with small, often imperceptible perturbations to mislead a model while appearing normal to humans.
Adversarial Patch(adversarial patches): A physical adversarial attack that places a crafted patch in the scene to fool a model.
Adversarial Robustness(adversarial robustness): The resilience of a model against adversarial examples, often measured by accuracy on a test suite of attacks.
Adversarial Testing(adversarial testing): A testing method that applies small perturbations to input data to create adversarial examples that can fool a model.
Adversarial Training(adversarial training): A technique that includes adversarial examples in the training data to improve model robustness.
Adversarial testing for explainability(adversarial testing): A method that probes a model's decision boundaries with small perturbations to verify explanation stability and faithfulness.
Automated Evaluation Metrics(automated evaluation metrics): Metrics like BLEU and ROUGE used to assess output quality, though they often do not align with human judgment.

B

BLEU metric(bleu): An automated metric that measures precision of n‑grams for text generation output quality.
Back-to-Back Testing: A testing technique that runs the same inputs through multiple implementations and compares their outputs to detect inconsistencies.
Bias and Fairness Testing(bias and fairness testing): Testing with diverse and adversarial inputs to uncover unexpected bias and fairness issues in generative AI systems.
Bias detection techniques(bias detection technique, diverse inputs, adversarial inputs): Methods using diverse and adversarial inputs to uncover unexpected discrimination in AI systems.

C

Canary Testing(canary testing, canary release): A deployment pattern that gradually routes a small percentage of live user traffic to a new model, exposing it to real users while monitoring performance.
Carlini & Wagner Attack(carlini & wagner attack, carlini and wagner): An adversarial attack method that generates examples by optimizing a custom loss function.
Class imbalance(imbalanced dataset, imbalanced classes, class imbalance): A situation where classes have significantly different numbers of samples, influencing the choice of averaging method for performance metrics.
Classification Metamorphic Relations(classification mrs): Metamorphic relations for classification models, often checking label invariance under small perturbations.
Cloud hosting(cloud, cloud hosting): AI system hosted on a cloud provider, offering managed services and scalability but with variable performance.
Coefficient of Determination(r², r-squared): The proportion of variance in the target explained by the model, computed as 1 - (SS_res/SS_tot).
Completeness Constraint(completeness): A constraint that ensures no missing values in critical fields of a dataset.
Confusion matrix(confusion matrix, true positives, true negatives, false positives, false negatives): A table that compares actual class labels with predicted class labels, containing TP, TN, FP, FN cells.
Consistency Constraint(consistency): A constraint that checks data follows defined formats and ranges.
Consistency under transformations(consistency under transformations, metamorphic relation): A metamorphic relation that adding a constant to features should shift regression output by that constant.
Cost-sensitive metric selection(metric selection based on business context, error cost trade-off, metric interpretation in business context): The practice of choosing performance metrics based on the relative costs of false positives and false negatives in the application domain.
Cross-Referencing(cross-referencing): A technique for testing accuracy by comparing data with external databases.
conformity assessment(conformity assessment obligations): A mandatory evaluation process for high-risk AI systems to verify compliance with regulatory requirements.

D

Data Accuracy(data accuracy): A dataset constraint ensuring data correctly represents real-world values.
Data Augmentation(data augmentation): Technique that creates modified versions of existing data to increase dataset diversity.
Data Completeness(data completeness): A dataset constraint ensuring all necessary data is present.
Data Consistency(data consistency): A dataset constraint ensuring data is uniform and free of contradictions.
Data Drift(data drift): A change in data distribution over time that may require updating dataset constraints.
Dataset Constraint Testing(dataset constraint testing): The process of verifying that data meets predefined quality rules such as completeness, consistency, and accuracy.
Dataset Constraints(dataset constraint, dataset constraints, completeness, consistency, accuracy): Quality requirements such as completeness, consistency, and accuracy that must be tested for input data.
Differential Privacy(differential privacy): A technique that provides mathematical guarantees against re-identification when generating synthetic data.

E

EU AI Act(european union's ai act, risk classification, risk level classification): A risk-based regulatory framework that classifies AI systems by risk level and imposes conformity assessment obligations.
Entailment model for hallucination detection(entailment models): A natural language inference model that checks whether generated output is entailed by a source text.
Ethical Considerations in AI Testing(ethical considerations, ethical testing of ai systems, ai ethics): Focus on fairness, accountability, transparency, and bias mitigation to ensure non-discrimination and explainability in AI systems.
Ethics in AI(ethics): Cross-cutting quality concern ensuring AI systems respect human rights, privacy, and societal values.
Explainability(explainability): A quality characteristic of AI systems that enables providing human-understandable reasons for model decisions.
Explainability Testing(explainability testing): Verification that a model's decisions can be understood and justified, building trust and enabling bias detection.
Exploratory Testing(exploratory testing): An approach to testing generative AI that involves exploring the output space.
Exploratory Testing for AI(exploratory testing): A testing technique where testers probe the model's behavior with diverse inputs to uncover unexpected issues.

F

F1-score(f1-score, f1 score, f1): A classification metric computed as the harmonic mean of precision and recall: 2 × (Precision × Recall) / (Precision + Recall).
FID score(fid, fréchet inception distance): A metric measuring visual quality and diversity of generated images via Fréchet Inception Distance.
FPGA (Field-Programmable Gate Array)(fpga, field-programmable gate array): A reconfigurable hardware device that can be customized for specific AI tasks.
Factuality score(factuality scores, factual accuracy metrics): A metric that compares generated statements against a trusted knowledge source to detect hallucinations.
Fairness(fairness): The absence of bias that leads to unjust or discriminatory outcomes across different user groups in an AI system.
Fairness (AI)(fairness): The absence of bias leading to unjust or discriminatory outcomes across different user groups.
Fairness in AI(fairness, bias): The absence of bias that leads to unjust or discriminatory outcomes across different user groups.
False Negative(false negatives, fn): A cell in the confusion matrix counting instances incorrectly predicted as negative.
False Negatives (FN)(false negatives, fn): Cells in a confusion matrix counting incorrectly predicted negative instances.
False Positive(false positives, fp, false positives (fp), false positive): A cell in the confusion matrix counting instances incorrectly predicted as positive.
Fast Gradient Sign Method (FGSM)(fast gradient sign method, fgsm): An attack method that adds perturbations in the direction of the gradient of the loss function to create adversarial examples.
Feature importance analysis(feature importance analysis, feature importance): A technique that identifies which input features most influence a model's prediction, often using SHAP or LIME.
Follow-up Test Cases(follow-up test cases): Transformed versions of source test cases created by applying metamorphic relations.
Follow-up test case(follow-up test case): The input obtained after applying a transformation to a source test case.
Functional Appropriateness(functional appropriateness): The degree to which an AI system's behavior matches user expectations in context.
Functional correctness(functional correctness): The ability of an AI system to produce correct outputs given inputs, considering inherent uncertainty.
Functional correctness (AI)(functional correctness, functional correctness in ai): The ability of an AI system to produce correct outputs for given inputs, considering inherent uncertainty in predictions.

G

GPU (Graphics Processing Unit)(gpu, graphics processing unit): A hardware accelerator that performs many calculations simultaneously for deep learning workloads.
Generative AI testing(testing generative ai): The specialized process of evaluating generative AI models, often complicated by non-deterministic outputs.
Generative Adversarial Networks(gans, generative adversarial networks): A class of machine learning frameworks that generate synthetic data through adversarial training.
Ground truth dataset(ground truth datasets): A reference dataset of human‑expert outputs used as a benchmark for automated evaluation metrics.

H

Human evaluation for generative AI(human evaluation): A primary method where human raters assess outputs on criteria like fluency, coherence, relevance, and factual accuracy.
Hybrid hosting(hybrid, hybrid hosting): AI system hosted in a combination of on-premises and cloud configurations.

I

IEEE 7000 series(ieee 7000): A set of standards addressing ethical concerns like transparency and accountability in AI systems.
ISO/IEC 23053(framework for ai systems using machine learning): A standard providing a framework for AI systems using machine learning, covering data quality and model validation.
ISO/IEC 25010(iso/iec 25010): The base software quality model that ISO/IEC 25059 extends for AI-specific characteristics.
ISO/IEC 25059(iso/iec 25059): An international standard extending ISO/IEC 25010 to define quality characteristics for AI-based systems.
ISO/IEC 38507(governance of ai): A standard providing guidelines for the governance of artificial intelligence.
ISTQB Glossary(istqb glossary): An authoritative source defining key terms for AI testing, such as bias and fairness, used across ISTQB exams.
Imbalanced Dataset(imbalanced dataset): A dataset where classes are not equally represented, making accuracy misleading and F1-score more informative.
Inception Score(is, inception score): A metric that evaluates image generation quality based on classifier confidence and diversity.
Input Data Testing(input data testing): Testing performed on input data to ensure quality and suitability for machine learning.
Interpretability(interpretability): The degree to which a human can understand the cause of a model's decision, often traded off against model complexity.
Invariance under data augmentation(invariance under data augmentation, invariance relation): A metamorphic relation that small perturbations like rotation should not change the predicted class.

L

LIME(lime): A technique that explains predictions by locally approximating a complex model with an interpretable surrogate model.
Label Integrity Preservation(label integrity): The requirement that data transformations do not change the original label of the training example.

M

METEOR metric(meteor): An automated metric that aligns generated text with references using synonyms and stemming.
Machine Learning Development Testing(machine learning development testing): Testing integrated into the machine learning development lifecycle.
Machine Learning Functional Performance Metrics(machine learning functional performance metrics): Metrics used to evaluate the functional performance of machine learning models.
Machine Learning Model Testing(machine learning model testing): Testing conducted on machine learning models to assess their performance.
Macro-averaging(macro-averaging, macro averaging, macro-averaged metric, macro-averaged): An averaging method that computes the metric per class and then takes the unweighted average across all classes.
Macro-averaging and Micro-averaging(macro-averaging, micro-averaging): Methods for aggregating metrics across classes in multi-class classification.
Mean Absolute Error(mae): Mean of absolute differences between predicted and actual values, in the same unit as the target.
Mean Absolute Error (MAE)(mae, mean absolute error): A metric that measures the average absolute difference between predicted and actual values, in the same unit as the target variable.
Mean Squared Error(mse): The average squared difference between actual and predicted values, computed as (1/n) Σ(actual - predicted)².
Mean Squared Error (MSE)(mse, mean squared error): A metric that averages the squared differences between predictions and actuals, penalizing larger errors more heavily.
Metamorphic relation(mr, metamorphic relation, metamorphic relations): A property stating that given a set of inputs, the corresponding outputs should satisfy a specific relationship.
Metamorphic testing(mt, metamorphic testing): A testing technique that uses metamorphic relations to generate test cases and verify outputs without a traditional oracle.
Metric Interpretation in Business Context(metric selection based on business cost, cost of false positives vs false negatives): The practice of selecting metrics based on the cost of false positives vs. false negatives in the application domain.
Micro-averaging(micro-averaging, micro averaging, micro-averaged metric, micro-averaged): An averaging method that aggregates contributions from all classes to compute the overall metric.
Misleading explanation(misleading explanation, misleading explanations): An explanation that appears plausible but does not accurately reflect the model's internal logic.
Model Complexity(model complexity): A property of machine learning models referring to the number of parameters or capacity, often traded off against interpretability.
Model Complexity vs Interpretability Trade-off(trade-off between model complexity and interpretability): The balance between model performance and the ability to understand model decisions.
Monotonic relationship(monotonicity): A property where increasing an input feature should not decrease the predicted output in a regression model.

N

NIST AI Risk Management Framework(u.s. nist ai risk management framework): A US framework proposing guidelines for managing risks associated with AI systems.
Non-determinism in Generative AI(non-determinism, non-determinism in generative ai): The property that the same input can yield different outputs, requiring testers to evaluate distributions rather than single results.
Non-determinism in ML training(non-determinism in ml training, non-determinism): Variability in training results caused by hardware and framework choices, making exact reproduction difficult.

O

On-premises hosting(on-premises, on-premises hosting): AI system hosted locally, giving full control over data security and latency but requiring significant capital investment.
Oracle problem(oracle problem): The challenge of verifying test outputs without expected results, addressed by metamorphic testing.
Output Quality Evaluation(output quality evaluation): The process of assessing generative AI outputs across dimensions such as relevance, coherence, factual correctness, and safety.
Output Quality Evaluation for Generative AI(output quality evaluation for generative ai): Assessment of generative AI outputs along dimensions such as relevance, coherence, factual correctness, and safety.
Overly Strict Constraint(overly strict constraints): A constraint that rejects valid data, potentially leading to reduced dataset size and model underfitting.

P

Pairwise Testing(all-pairs testing): A combinatorial test design technique that tests all possible pairs of input parameter values to reduce the number of test cases.
Precision(precision): A classification metric computed as TP / (TP + FP), measuring how many positive predictions are correct.
Projected Gradient Descent (PGD)(projected gradient descent, pgd): An iterative attack method that applies FGSM with clipping to stay within a small perturbation budget.
PyTorch(pytorch): An ML framework from Facebook with a dynamic computation graph, popular in research for ease of debugging.

Q

Quality Characteristics for AI-Based Systems(quality characteristics for ai-based systems): Attributes used to evaluate the performance and reliability of AI systems.

R

R-squared (R²)(r², r-squared): A metric that represents the proportion of variance in the target variable explained by the model, ranging from -∞ to 1.
RAG Testing(rag testing): Testing focused on the quality and relevance of retrieved context in retrieval-augmented generation systems.
RMSE(rmse): A regression metric that measures the square root of the average squared differences between predicted and actual values.
ROC Curve and AUC(roc curve, auc): Tools derived from the confusion matrix to evaluate classifier performance across thresholds.
ROUGE metric(rouge): An automated metric that measures recall of n‑grams for text generation output quality.
Recall(recall, sensitivity, recall (sensitivity)): A classification metric computed as TP / (TP + FN), measuring how many actual positives are captured.
Red Teaming(red teaming): A structured attempt to make a generative AI system produce undesirable outputs to identify safety and security weaknesses.
Regression Metamorphic Relations(regression mrs): Metamorphic relations for regression models, such as verifying monotonic relationships.
Regression metamorphic relation(regression metamorphic relation): Metamorphic relations designed for regression models to verify consistent behavior under transformations, such as monotonicity.
Robustness(robustness): The ability of an AI system to handle invalid inputs or adversarial perturbations.
Robustness (AI)(robustness): The ability of an AI system to maintain performance under adverse conditions such as adversarial inputs or distributional shift.
Root Mean Squared Error(rmse): The square root of Mean Squared Error, bringing the metric back to the original unit of the target variable.
Root Mean Squared Error (RMSE)(rmse, root mean squared error): The square root of MSE, bringing the metric back to the original unit, used when large errors are undesirable.
Rule-Based Check(rule-based checks): A technique for testing constraints by applying predefined rules to detect violations.

S

SHAP(shap): A feature importance technique based on Shapley values that explains individual predictions by quantifying feature contributions.
SQuaRE series(square series, squre series, square, software quality requirements and evaluation series): The Software Quality Requirements and Evaluation series of standards, of which ISO/IEC 25059 is a part.
Safety(safety, safety in ai systems, ai safety, safety testing for ai systems): The ability of an AI system to avoid causing harm, especially in high-stakes applications like autonomous driving or healthcare.
Safety (AI)(safety): The ability of an AI system to avoid causing harm, especially in high-stakes applications like autonomous driving.
Safety in AI(safety): Cross-cutting quality concern focusing on preventing physical or financial harm from AI systems.
Safety testing: An AI testing technique that assesses whether the model produces unsafe or harmful outputs.
Sampling Bias(sampling bias): A systematic error where training data does not represent the real-world distribution, leading to unfair outcomes.
Shadow Testing(shadow testing, shadow model deployment): A deployment pattern where a new model runs in parallel with the production model, processing live data but not affecting user-facing decisions.
Source test case(source test cases, source test case): The initial input used in metamorphic testing before applying transformations.
Specificity(specificity, true negative rate): A classification metric computed as TN / (TN + FP), measuring the proportion of actual negatives correctly identified.
Statistical Profiling(statistical profiling): A technique for testing constraints using statistical analysis of data distributions.
Surrogate model(surrogate models, surrogate model): A simpler, interpretable model that approximates a complex model's decisions to provide explanations.
Synthetic Data(synthetic data): Artificially generated data that mimics statistical properties of real data without containing actual personal information.
Synthetic Data Fidelity(fidelity): The degree to which synthetic data accurately represents the statistical properties of real data.
Synthetic Data Generation(synthetic data generation): Technique to artificially create data to address scarcity and imbalance.
Synthetic Data Privacy(privacy, re-identification risk): The property that synthetic data cannot be re-identified to any individual, often protected by differential privacy.
scikit-learn(scikit-learn): An ML framework for traditional algorithms, often used for preprocessing and baseline models.

T

TPU (Tensor Processing Unit)(tpu, tensor processing unit): A custom ASIC designed by Google specifically for neural network operations.
TensorFlow(tensorflow): An ML framework from Google with a static computation graph, suited for production deployment.
Testing Generative AI Systems(testing generative ai systems): The process of evaluating generative AI models for quality and correctness.
Trade-offs in AI Quality(trade-offs): The balancing of AI quality characteristics, where improving one may degrade another, requiring context-dependent management.
Transferability of Adversarial Examples(transferable, transferability, transferability of adversarial examples): The property that an adversarial example generated for one model may also fool a different model trained on similar data.
Transparency(transparency, transparency in ai): The degree to which an AI system's behavior and decisions can be understood by humans, including explainability and interpretability.
Transparency (AI)(transparency, explainability, ai transparency): The degree to which an AI system's behavior and decisions can be understood by humans, including explainability.
True Negative(true negatives, tn, true negatives (tn)): A cell in the confusion matrix counting instances correctly predicted as negative.
True Positive(true positives, tp, true positives (tp)): A cell in the confusion matrix counting instances correctly predicted as positive.

V

Validation of Explanations(validate explanations, explanation validation, validate explanations against actual model behavior, validation of model explanations, validation of explanations): The process of verifying that explanation methods accurately reflect the actual behavior of a machine learning model.
Validity of Metamorphic Relations(valid mrs): The requirement that metamorphic relations must be carefully designed for the specific ML task.
Variational Autoencoders(vaes, variational autoencoders): A type of generative model that learns to encode and decode data for synthetic data generation.