AI Associate Ethical Considerations of AI Practice Questions

Explanation: The correct first step is to check the training data for representation and bias because the model's tendency to recommend high-priced items to affluent areas suggests the training data may be skewed or contain historical biases. Einstein Prediction Builder relies on historical data to learn patterns, and if the data over-represents affluent users or under-represents others, the model will perpetuate those biases. Auditing the data for fairness and representation is the foundational step before any remediation, as per responsible AI practices.

Explanation: Option A is correct because the sentiment analysis model likely exhibits bias due to training data that does not adequately represent the linguistic patterns, idioms, or expressions of non-native English speakers. This causes the model to misinterpret neutral or positive statements from these users as negative or frustrated, leading to disproportionate escalations. A non-representative dataset is a common source of algorithmic bias in AI systems.

Explanation: Removing protected attributes is a common step, but if they are proxies for legitimate medical factors, they may be retained with monitoring. Option A is too aggressive. Option C ignores that age can be medically relevant. Option D violates transparency and accountability.

Explanation: Option D is correct because analyzing the distribution of scores across industry segments directly validates whether the model exhibits systematic bias. By comparing score distributions, the associate can identify if certain industries are consistently under-scored, which would indicate a biased pattern rather than random variation. This approach aligns with ethical AI practices that require transparency and fairness assessment before any model adjustments.

Explanation: Option D is correct because including gender as a feature in a predictive model for employee performance can introduce bias and lead to unfair or discriminatory outcomes. Even if the model's accuracy improves, using protected attributes like gender may violate ethical guidelines and regulations such as GDPR or anti-discrimination laws, as it could perpetuate historical biases or result in disparate impact.