Trusted AI Safety Expert (TAISE)

Last checked November 10, 2026 against the official Cloud Security Alliance exam page.

1. Question 1Introduction to AI

   A bank wants a model that predicts whether a loan applicant will repay, using thousands of past applications where the repayment outcome is already recorded. Which approach best fits this requirement?

   1. AClustering applicants into unlabeled segments for later manual review
   2. BPrincipal component analysis that compresses the raw application fields into fewer features before scoring
   3. CReinforcement learning that rewards the model each time a loan is approved
   4. DSupervised learning trained on historical applications with recorded repayment outcomes
   Show answer and explanationHide answer

   Correct answer: D

   Recorded repayment outcomes act as labels, so the task is supervised learning. Clustering works without labels and does not predict a known outcome, reinforcement learning needs an interactive reward signal, and PCA reduces dimensionality without predicting anything.

2. Question 2Generative AI Architecture and Design

   A team wants a chatbot to answer questions from its internal policy documents without retraining the underlying foundation model whenever those documents change. Which architecture pattern best fits?

   1. AHard-coding each policy answer as a fixed rule in a decision tree before deployment, with no language model involved
   2. BRetraining the entire foundation model from scratch on the policy documents every night
   3. CRaising the sampling temperature so the model draws on more of its stored parameters
   4. DRetrieval-augmented generation, which fetches relevant passages at query time and adds them to the prompt
   Show answer and explanationHide answer

   Correct answer: D

   Retrieval-augmented generation keeps knowledge in an external store that can be updated without retraining, and the retrieved passages ground the answer at query time. Nightly full retraining is impractical, temperature only changes randomness in token selection rather than adding knowledge, and a fixed decision tree does not use a generative model at all.

3. Question 3Generative AI Use Cases and Applications

   A logistics company wants staff to paste long shipment-incident reports into a tool that produces a one-paragraph summary for managers. Which use case category does this represent?

   1. AImage generation, which creates new pictures from text prompts supplied by a user
   2. BText summarization, which condenses long documents into shorter versions while keeping their key points
   3. CAnomaly detection, which flags sensor readings and transactions that deviate from learned normal patterns over time
   4. DSpeech synthesis, which converts written text into spoken audio for playback
   Show answer and explanationHide answer

   Correct answer: B

   Condensing a long report into a short paragraph is a summarization task, which is a common generative AI use case. Image generation, anomaly detection, and speech synthesis solve different problems and do not match the input and output described.

4. Question 4Ethics, Transparency, and Explainability in AI

   A lender uses a credit-scoring model and must explain adverse decisions to applicants. The team wants reasons for each individual decision rather than a general description of the model. Which approach best fits?

   1. AA single overall accuracy figure published in the model documentation for the full test population
   2. BLocal attribution methods, such as SHAP values, that show which features drove one specific prediction
   3. CPublishing the training source code so applicants can rerun the model privately on their own data
   4. DAdding more layers to the network so that each reason becomes easier to read off the architecture
   Show answer and explanationHide answer

   Correct answer: B

   Local attribution methods estimate how much each input feature contributed to one prediction, which is what a per-decision reason needs. Overall accuracy describes the model globally, publishing code does not explain a specific outcome, and adding layers makes a network harder to interpret, not easier.

5. Question 5AI Model Lifecycle and Threat Taxonomy

   A team is about to release a new fraud model. Which lifecycle activity should define the monitoring metrics, acceptance thresholds, and rollback criteria before go-live?

   1. APre-deployment validation, where release gates and rollback conditions are agreed and tested
   2. BData collection, when analysts first attach fraud labels to historical transactions
   3. CModel retirement, when the previous version is archived for audit purposes
   4. DHyperparameter search, when engineers tune learning rates and regularization strength using only the training split
   Show answer and explanationHide answer

   Correct answer: A

   Release gates, monitoring plans, and rollback criteria belong in pre-deployment validation, before the model serves live traffic. Labeling, retirement, and tuning are real lifecycle activities, but they do not set the conditions for a safe release.