Free CT-AI: Certified Tester AI Testing Exam Questions and Answers

A sentiment classifier labels the source test input "The delivery was quick and the staff were friendly." as Positive with a confidence of 0.93. A tester wants to derive a follow-up test case by applying metamorphic testing, where no independent expected result is available. Which of the following follow-up test cases is a valid application of the technique? Select ONE option.

• AA. Translate the review into French and expect a confidence of exactly 0.93 for the label Positive
• BB. Replace "quick" with "fast" and expect the label Positive to be returned again
• CC. Delete the clause about the delivery and expect the label Negative to be returned
• DD. Replace "friendly" with "rude" and expect the label Positive to be returned again

A model trained in the cloud with 32-bit floating-point weights is converted to 8-bit quantised weights so that it can run on a battery-powered handheld scanner. Which of the following activities BEST addresses the risk introduced by this conversion? Select ONE option.

• AA. Measure neuron coverage of the original model with the cloud-hosted test suite and compare it with the target value
• BB. Repeat the unit tests of the code in the data pipeline that prepared the original training data in the cloud
• CC. Run the functional performance tests on the converted model on the scanner and compare the metrics with those of the original model
• DD. Retrain the original model with a larger training dataset and confirm that it still meets the accuracy target in the cloud

Which ONE of the following is the purpose of the test dataset in the ML workflow? Select ONE option.

• AA. Comparing candidate hyperparameter values to choose a model configuration
• BB. Measuring the functional performance of the final model on data not used to build it
• CC. Adjusting the weights of the model during each training iteration
• DD. Increasing the variety of samples by generating modified copies of existing data

A bank is introducing an ML-based system that flags fraudulent card transactions for manual review. About 0.4% of historical transactions were fraudulent. The test team must agree an acceptance criterion for the model's functional performance with the business. Which ONE of the following acceptance criteria is the MOST appropriate? Select ONE option.

• AA. Overall accuracy of at least 99% measured on a test dataset reflecting the production transaction mix
• BB. Recall of at least 92% for fraudulent transactions measured on a test dataset reflecting the production mix
• CC. Training loss below 0.01 at the final epoch of the last training run before release
• DD. Neuron coverage of 100% reached by the test suite on the model selected for deployment

Which ONE of the following BEST describes a hallucination produced by a large language model (LLM)? Select ONE option.

• AA. Different wording in the responses to the same prompt because of a sampling randomness setting
• BB. A fluent and plausible response containing statements that are factually wrong or unsupported by the source
• CC. A drop in response quality after the language used by the users changes over several months
• DD. A refusal to answer a prompt because a safety filter classified the request as harmful

A logistics company is developing software to control a robotic arm in a warehouse. No labelled examples are available. The software tries sequences of movements, receives a reward each time a parcel is placed correctly on a shelf and a penalty each time a parcel is dropped, and gradually improves its choice of movements. Which ONE of the following types of ML is being used? Select ONE option.

• AA. Supervised learning
• BB. Unsupervised learning
• CC. Semi-supervised learning
• DD. Reinforcement learning

A hospital provides patient records to train an ML model that predicts length of stay. The agreed data constraints are: 1. Age must be between 0 and 110 years, inclusive. 2. The discharge date must not be earlier than the admission date. 3. Heart rate must be between 30 and 220 beats per minute, inclusive. Which ONE of the following records violates at least one of the constraints? Select ONE option.

• AA. Age 0, admitted 2025-05-11, discharged 2025-05-11, heart rate 140
• BB. Age 45, admitted 2025-06-20, discharged 2025-06-14, heart rate 88
• CC. Age 67, admitted 2025-03-02, discharged 2025-03-09, heart rate 48
• DD. Age 110, admitted 2025-01-04, discharged 2025-01-30, heart rate 220

A model that classifies incoming emails as phishing or legitimate was tested with 200 emails. The results are shown below. Actual phishing, predicted phishing: 45 Actual legitimate, predicted phishing: 5 Actual phishing, predicted legitimate: 15 Actual legitimate, predicted legitimate: 135 Treating phishing as the positive class, what is the recall of the model? Select ONE option.

• AA. 0.75
• BB. 0.82
• CC. 0.90
• DD. 0.96

A neural network has 50 neurons. A threshold of 0.6 is used for threshold coverage. A test suite of three tests is run, with these results: - Test 1: activation value above 0.6 for 18 neurons. - Test 2: activation value above 0.6 for 14 neurons, of which 10 are among the 18 neurons from Test 1. - Test 3: activation value above 0.6 for 12 neurons, none of which were exceeded in Test 1 or Test 2. What is the threshold coverage achieved by the test suite? Select ONE option.

• AA. 36%
• BB. 68%
• CC. 76%
• DD. 88%

A team has trained a fraud-detection model but has not yet connected it to the payment application. They want to measure precision and recall on a test dataset and check whether the model overfits. At which test level are these activities performed? Select ONE option.

• AA. Input data testing
• BB. Component integration testing
• CC. Model testing
• DD. Acceptance testing