Ethical Considerations of AI practice questions

Remove race and age features entirely to ensure fairness.

Why wrong: Age may be clinically relevant; removal could reduce accuracy unfairly.

Replace age with an age group bucket to reduce granularity.

Why wrong: This does not address bias from race; still problematic.

Use the model as is because predictions are accurate.

Why wrong: Accuracy does not justify ethical or legal risk.

Increase the fallback threshold for all languages.

Why wrong: This would increase fallback for English too, not fair.

Collect more data from Spanish-speaking customers and retrain the English model.

Why wrong: Retraining English model on Spanish data will not fix English model's language limitation.

Disable sentiment analysis for non-English conversations.

Why wrong: Disabling loses functionality; better to support.

Add a post-processing adjustment to increase female candidates' scores.

Why wrong: Post-processing may not address model's internal bias.

Accept the bias as a reflection of historical data.

Why wrong: Historical bias should be actively mitigated, not accepted.

Remove the gender feature from the model.

Why wrong: Removing the feature does not prevent correlation with other features.

The training data may not be representative

Why wrong: The team has already used a representative dataset, so this is not the gap.

Customer consent was not obtained

Why wrong: Customer consent is important, but the policy specifically requires explainability.

Bias testing was not performed

Why wrong: The team performed bias testing, so this is not the gap.

Increase model complexity

Why wrong: Increasing complexity may amplify existing biases, not address the root cause.

Add more features to the model

Why wrong: Adding features without bias analysis may introduce or reinforce bias.

Remove zip code feature

Why wrong: Removing the feature alone does not guarantee fairness if other features correlate with zip code.

Cost of implementation

Why wrong: Cost of implementation is a business concern, not an ethical one.

Model accuracy

Why wrong: While model accuracy is important, it is a technical metric and not the primary ethical concern; data privacy takes precedence.

Transparency of predictions

Why wrong: Transparency of predictions is important for trust, but data privacy is more foundational when handling sensitive customer data.

Underfitting

Why wrong: Underfitting would result in poor performance across all queries, not just minority languages.

Lack of compute resources

Why wrong: Insufficient compute does not specifically cause biased performance on minority languages.

Overfitting

Why wrong: Overfitting would lead to poor performance on new data overall, not specifically minority languages.

Use only demographic features

Why wrong: Relying solely on demographics can reinforce biases and stereotypes.

Remove all user data

Why wrong: Removing data defeats the purpose of personalized recommendations.

Increase model depth

Why wrong: Model depth relates to complexity, not fairness.

Low latency

Why wrong: Latency is a technical performance requirement, not ethical.

High precision

Why wrong: Precision is a performance metric, not a primary ethical requirement.

Use of external data sources

Why wrong: External data can introduce privacy and consent issues.

Increase model regularization

Why wrong: Regularization prevents overfitting but does not directly mitigate bias.

Remove the gender feature

Why wrong: Removing the feature is insufficient if other features are correlated with gender.

Reduce model complexity

Why wrong: Reducing complexity may hurt accuracy without solving bias.

Inclusion

Why wrong: Inclusion focuses on fair treatment of diverse groups, not monitoring practices.

Accountability

Why wrong: Accountability is about who is responsible for outcomes, not about disclosure.

Safety

Why wrong: Safety concerns physical or psychological harm, not necessarily surveillance.

Accountability

Why wrong: Accountability is about who is responsible for the model's performance, not the performance itself.

Overfitting

Why wrong: Overfitting would cause poor performance even within the training region's test data.

Privacy violation

Why wrong: Privacy is not the issue; the model is trained on legitimate data from one region.

Using only internal data for training

Why wrong: Internal data may not represent diverse populations.

Automating all decisions without human review

Why wrong: Human oversight is necessary for high-risk decisions.

Using customer data without explicit consent

Why wrong: This violates privacy and consent principles.

Use a single fairness metric to evaluate the model

Why wrong: Fairness is multidimensional; one metric is inadequate.

Deploy the model quickly to gather real-world data

Why wrong: Rapid deployment without fairness checks can harm marginalized groups.

Remove all sensitive attributes from the data

Why wrong: Sensitive attributes may be proxied by other features; removal alone is insufficient.

Performance

Why wrong: Performance is not indicated in the policy and is less critical than transparency. The policy focuses on explainability, not performance metrics.

Data privacy

Why wrong: Data privacy is important but the policy does not mention any privacy violations, so it is not the most critical risk here.

Model accuracy

Why wrong: Model accuracy is not addressed in the policy. While accuracy is important, transparency is more fundamental for ethical AI.

Increase the accuracy threshold

Why wrong: Accuracy threshold does not directly affect fairness.

Remove the model

Why wrong: Removal may be premature; retraining is a better first step.

Deploy as is

Why wrong: A ratio of 0.6 indicates significant bias; deploying would be unethical.

Safety

Why wrong: Safety is about preventing harm, not providing explanations.

Fairness

Why wrong: Fairness concerns bias, not explainability.

Accountability

Why wrong: Accountability is about who is responsible, not the lack of explanation.

Remove gender-related features from the model.

Why wrong: Removing features may not eliminate proxy variables that encode gender.

Re-train the model with only male candidates.

Why wrong: Training only on males would introduce more bias, not solve the problem.

Increase the model's training data size.

Why wrong: More data does not guarantee bias reduction; the data may still be biased.

Allow the system to update its features dynamically without documentation.

Why wrong: Undocumented changes reduce accountability.

Restrict the system's use to only internal employees.

Why wrong: Restricting use does not address explainability.

Use a black-box neural network for highest accuracy.

Why wrong: Black-box models are not transparent.

Reduce the model's complexity to improve performance.

Why wrong: Complexity reduction does not address data bias.

Use a completely different model for small businesses.

Why wrong: Different models may lead to inconsistent and unfair treatment.

Ignore small businesses as they contribute little revenue.

Why wrong: Ignoring a segment is unfair.