Ethical Considerations of AI practice questions

Remove the model from production immediately.

Why wrong: Removing without analysis may be unnecessary and disrupt business.

Ignore the issue because the model predictions are accurate overall.

Why wrong: Accurate overall does not mean fair; disparity is still a concern.

Add more features about customer income.

Why wrong: Adding income features could exacerbate bias, not reduce it.

The bot is routing to the wrong department.

Why wrong: Routing is not the cause of disproportionate escalation.

The escalation threshold is set too low.

Why wrong: Threshold affects volume, not demographic disparity.

The bot is not properly connected to the escalation queue.

Why wrong: The bot escalates, so connection works.

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.

Run a holdout test to check prediction accuracy.

Why wrong: Accuracy test does not directly address bias.

Retrain the model with balanced data.

Why wrong: Retraining is a fix, not a validation step.

Review the model's confidence intervals.

Why wrong: Confidence intervals measure uncertainty, not bias.

Evaluating model performance across different groups.

Why wrong: Evaluation helps detect bias, so it reduces risk.

Excluding gender from the model features.

Why wrong: Excluding protected attributes reduces ethical risk.

Documenting model limitations and assumptions.

Why wrong: Documentation reduces risk by promoting transparency.

Remove the model and use a rule-based system.

Why wrong: Rule-based systems can also be biased and less effective.

Use SHAP values to explain predictions.

Why wrong: Explanation does not change biased outcomes.

Add a disclaimer that recommendations may be biased.

Why wrong: Disclaimers do not mitigate harm.

The model is not interpretable.

Why wrong: Interpretability is separate from fairness.

The model has low accuracy for Spanish reviews.

Why wrong: Low accuracy is a symptom, but the core is fairness.

The model violates privacy regulations.

Why wrong: No privacy issue mentioned.

Limiting access to model outputs to only a few people.

Why wrong: Limiting access reduces transparency.

Using complex deep learning models without explanation.

Why wrong: Lack of explanation reduces transparency.

Automatically retraining models weekly.

Why wrong: Retraining does not directly relate to transparency.

The complexity of the model architecture.

Why wrong: Complexity does not indicate bias.

The date the model was last deployed.

Why wrong: Deployment date is not a bias factor.

Use the AI model as the sole decision-maker.

Why wrong: Sole reliance on AI can be unfair.

Use all available data including protected attributes.

Why wrong: Using protected attributes may cause bias.

Ignore adverse impact if the model is accurate.

Why wrong: Accuracy does not justify adverse impact.

The model uses too many features.

Why wrong: Number of features is not an ethical concern by itself.

The model is not explainable.

Why wrong: The model type is not specified; interpretability is not directly addressed.

The training data is imbalanced.

Why wrong: Imbalance is a technical concern, but the ethical concern is the impact of low recall.

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.

Remove sensitive attributes like race and gender from the training data.

Why wrong: Removing attributes does not eliminate bias from correlated features.

Ignore the disparity because the model's overall accuracy is acceptable.

Why wrong: Fairness requires equitable performance across groups.

Retrain the model with more complex algorithms to improve accuracy.

Why wrong: Complex algorithms may amplify bias if data issues persist.

Privacy

Why wrong: Privacy is about data protection, not discriminatory impact.

Accountability

Why wrong: Accountability is about oversight, not the specific ethical breach.

Transparency

Why wrong: Transparency relates to model interpretability, not the unfair outcome.

Use unsupervised learning to allow the bot to learn naturally from conversations.

Why wrong: Unsupervised learning lacks control over output safety.

Remove all toxicity detection to avoid false positives.

Why wrong: Removing detection increases risk of offensive content.

Train the model only on customer service transcripts without review.

Why wrong: Unreviewed data may contain biased or toxic language.

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.

Remove zip code from the feature set and retrain.

Why wrong: Other features may still correlate with race.

Keep zip code but add a fairness penalty to the loss function.

Why wrong: Retaining proxy may still cause disparate impact.

Increase transparency by publishing the model's decision criteria.

Why wrong: Transparency does not correct bias.

Ensure the system achieves high accuracy and ignore other metrics.

Why wrong: Accuracy alone is insufficient for ethical AI.

Deploy immediately and monitor for issues.

Why wrong: Deployment without prior testing may cause harm.

Test the system only on a small dataset to expedite launch.

Why wrong: Small testing may not reveal biases.

Transparency

Why wrong: Transparency is about explaining decisions, not performance equity.

Accountability

Why wrong: Accountability is about who is responsible, not the ethical breach itself.

Privacy

Why wrong: Privacy is not directly related to group performance disparity.

Limit access to the model's logic to protect intellectual property.

Why wrong: Limiting access reduces transparency.

Use a complex deep learning model for higher accuracy.

Why wrong: Complex models are often less interpretable.

Remove feature importance to simplify the model.

Why wrong: Removing feature importance reduces transparency.