Guidelines for Responsible AI

175. A hiring-screening model is found to select candidates from one demographic group at a substantially lower rate. Which responsible AI dimension does this represent, and which tool helps investigate it?

1. A Cost; AWS Cost Explorer
2. B Latency; Amazon CloudWatch
3. C Fairness and bias; Amazon SageMaker Clarify
4. D Durability; Amazon S3 Versioning

176. Which practice most improves transparency for a foundation model deployed in a regulated environment?

1. A Reducing the model's temperature setting
2. B Encrypting the model artifacts at rest
3. C Running inference on Spot Instances
4. D Publishing model cards that document intended use, limitations, and evaluation results

177. A loan model is trained on historical approvals that reflect past discriminatory lending. What is the likely consequence?

1. A The model corrects the bias automatically during training
2. B The model reproduces the historical bias, because it learns the patterns present in its training data
3. C The bias affects only training, not predictions
4. D The bias disappears if the dataset is large enough

178. Which practice best supports human oversight of an AI system making consequential decisions about people?

1. A Publishing the model's parameter count
2. B Increasing the model's confidence threshold
3. C A human-in-the-loop review step with authority to override the model
4. D Logging all predictions for later analysis

179. What is the purpose of an AI service card or model card?

1. A To document intended use cases, limitations, and known risks so users can judge appropriate use
2. B To list the AWS Regions where the model is available
3. C To record the billing history of the service
4. D To store the model's training data

180. Which responsible AI dimension is concerned with a user being able to understand why a model produced a particular output?

1. A Durability
2. B Throughput
3. C Explainability
4. D Elasticity

181. A generative application occasionally produces offensive language in responses to ordinary questions. Which control addresses this most directly?

1. A Content filters that evaluate model output before it reaches the user
2. B Reducing the number of output tokens
3. C Adding more retrieval documents
4. D A longer system prompt asking the model to be polite

182. Why is dataset diversity important when training a facial analysis model?

1. A A dataset that under-represents some groups produces a model that performs worse for those groups
2. B Diversity removes the need for evaluation
3. C A diverse dataset always trains faster
4. D Diversity reduces the storage cost of the dataset

183. Which responsible AI dimension concerns a system continuing to perform reliably when it encounters inputs unlike those in its training data?

1. A Transparency
2. B Privacy
3. C Explainability
4. D Robustness

184. A company designs an AI interface without considering users who rely on screen readers. Which responsible AI consideration has been overlooked?

1. A Inference latency
2. B Data residency
3. C Inclusivity and accessibility
4. D Model accuracy

185. Which practice supports veracity in a generative AI application?

1. A Increasing the temperature for more varied answers
2. B Grounding responses in verifiable sources and surfacing citations to the user
3. C Removing the system prompt
4. D Generating longer responses

186. Which responsible AI dimension is addressed by giving users a mechanism to stop or correct an AI system's behaviour?

1. A Sustainability
2. B Fairness
3. C Privacy
4. D Controllability

187. Why does AWS publish AI Service Cards?

1. A To record which customers use each service
2. B To provide source code for the underlying models
3. C To list the pricing of each AI service
4. D To document intended use cases, limitations, and responsible design considerations for individual AI services

188. Several responsible AI dimensions must be matched to the concern each one addresses. (Match each dimension to its concern.)

1. Fairness
2. Explainability
3. Robustness
4. Controllability

1. A Whether a person can understand why a particular output was produced.
2. B Whether the system continues to behave reliably on unexpected or adversarial input.
3. C Whether the system's behaviour can be monitored, guided, and stopped.
4. D Whether outcomes differ systematically across groups of people.

189. Which responsible AI dimension concerns whether a system's outcomes differ systematically across groups of people?

1. A Fairness
2. B Privacy
3. C Robustness
4. D Governance

190. Which service applies configurable content filters and denied topics to a foundation model application?

1. A AWS WAF
2. B Amazon Inspector
3. C Amazon Macie
4. D Amazon Bedrock Guardrails

191. Why does dataset diversity matter when training or fine-tuning a model?

1. A Because larger datasets always train faster
2. B Because diverse datasets require less storage
3. C Because diversity removes the need for evaluation
4. D Because a dataset that under-represents a group produces a model that performs worse for that group

192. Which legal risk arises from a generative AI system producing content resembling copyrighted material?

1. A Automatic transfer of the model's ownership
2. B Intellectual property infringement, which may attach to the content the system generates
3. C Loss of the organization's own trademarks
4. D Expiry of the organization's data licences

193. What is the purpose of human oversight in a responsible AI system?

1. A To review and intervene in decisions where the consequence of an error is significant
2. B To replace automated monitoring entirely
3. C To eliminate the need for evaluation
4. D To approve every output before it is returned

194. What does a model card document?

1. A The model's hourly cost
2. B The model's source code
3. C A model's intended use, limitations, evaluation results, and known failure modes
4. D The model's training data in full

195. Which service computes bias metrics and per-prediction feature attributions for a model?

1. A Amazon SageMaker Clarify
2. B Amazon SageMaker Debugger
3. C Amazon SageMaker Model Monitor
4. D Amazon SageMaker Pipelines

196. Why is an interpretable model sometimes preferred over a more accurate one?

1. A Because interpretable models require less data
2. B Because accuracy is never important
3. C Because interpretable models always train faster
4. D Because a requirement to explain individual decisions may outweigh a small accuracy gain

197. Which practice increases transparency for users of an AI system?

1. A Hiding the system's use of AI to avoid confusion
2. B Increasing the model's response length
3. C Disclosing that they are interacting with an AI system and what its limitations are
4. D Publishing the model's weights

198. Which responsible AI dimension concerns a system continuing to behave correctly on inputs unlike its training data?

1. A Privacy
2. B Transparency
3. C Fairness
4. D Robustness

199. Which responsible AI dimension concerns whether a system's outputs are accurate and truthful?

1. A Governance
2. B Veracity
3. C Controllability
4. D Inclusivity

200. Which risk arises from training a model on data that under-represents a group of users?

1. A The model consumes more storage
2. B The model takes longer to train
3. C The model requires a larger context window
4. D The model performs worse for that group, producing unfair outcomes

201. Which practice reduces the risk of a generative AI system producing harmful content?

1. A Shortening the model's responses
2. B Applying content filters that evaluate both the request and the response before it reaches the user
3. C Instructing the model in the prompt to be helpful
4. D Increasing the model's temperature

202. Which environmental consideration applies to training large models?

1. A Training has no measurable environmental impact
2. B Environmental impact applies only to inference
3. C Environmental impact depends only on the storage used
4. D Training consumes substantial energy, so reusing an existing model where possible reduces environmental impact

203. Which practice supports identifying bias before a model is deployed?

1. A Increasing the size of the training dataset
2. B Evaluating performance separately for each relevant group rather than only in aggregate
3. C Reducing the model's complexity
4. D Measuring the model's overall accuracy

204. Which consideration applies to using a foundation model's output as the basis for a decision affecting an individual?

1. A The output should be used only if the confidence score exceeds ninety percent
2. B The output requires no record to be kept
3. C The output can be used directly because models are trained on large datasets
4. D A human should review the decision where its consequences are significant, since the output is probabilistic

205. What does feature attribution tell a reviewer about a model's prediction?

1. A How accurate the model is overall
2. B How much the model cost to train
3. C How long the model took to produce the prediction
4. D Which input features most influenced that particular prediction

206. Which practice supports transparency about an AI system's limitations?

1. A Publishing the size of the training dataset
2. B Publishing the conditions under which the system performs poorly alongside its intended use
3. C Publishing only the system's accuracy figure
4. D Publishing the model's architecture

207. Why might a simpler model be chosen over a more accurate one for a regulated decision?

1. A The simpler model always trains faster
2. B The simpler model's decisions can be explained in terms a regulator and an affected individual will accept
3. C The simpler model requires less data
4. D The simpler model costs less to run

208. Which responsible AI consideration applies when a model is trained on data collected over many years?

1. A Data age has no effect on a model's behaviour
2. B Older data always improves a model's accuracy
3. C Older data reduces the model's training cost
4. D Historical data may encode past patterns that are no longer appropriate to reproduce

209. What does controllability mean as a responsible AI dimension?

1. A The ability to monitor an AI system's behaviour and intervene or stop it when necessary
2. B The ability to control the system's operating cost
3. C The ability to control who may use the system
4. D The ability to control which Region the system runs in