Bias in generative models in Prompt Engineering / GenAI - Practice Problems & Coding Challenges

Challenge - 5 Problems

🧠 Conceptual

intermediate

Understanding Bias Sources in Generative Models

Which of the following is the most common source of bias in generative AI models?

A The hardware used to train the model

B The choice of activation function in the neural network

C The training data containing unbalanced or stereotypical examples

D The programming language used to implement the model

💡 Hint

Think about what influences the model's learned patterns the most.

Bias in generative models mainly comes from the training data. If the data has stereotypes or is unbalanced, the model learns and reproduces those biases.

❓ Predict Output

intermediate

Output of a Biased Text Generation Example

Given a generative model trained on biased data, what is the most likely output of this prompt?

Prompt: "The nurse said that"

Prompt Engineering / GenAI

def generate_text(prompt):
    # Simulated biased output based on stereotypical training data
    if prompt == "The nurse said that":
        return "she will take care of you soon."
    else:
        return "No output."

output = generate_text("The nurse said that")

A "she will take care of you soon."

B "the data is missing."

C "they will arrive later."

D "he will fix the machine."

💡 Hint

Consider common gender stereotypes in training data.

The model reflects gender bias by associating nurses with female pronouns, showing bias in output.

❓ Metrics

advanced

Evaluating Bias with Fairness Metrics

Which metric is best suited to measure bias in a generative model's output across different demographic groups?

A Demographic Parity Difference

B Mean Squared Error

C Cross-Entropy Loss

D BLEU Score

💡 Hint

Look for a metric that compares output distributions between groups.

Demographic Parity Difference measures if the model's outputs are equally likely across groups, indicating bias.

🔧 Debug

advanced

Identifying Bias Amplification in Model Output

Consider a generative model that produces the following outputs for the prompt "The CEO is":

["a man", "a woman", "a man", "a man", "a woman"]

What is the bias amplification issue here?

A The model output is random and unbiased

B The model under-represents male CEOs compared to the real-world distribution

C The model output is perfectly balanced

D The model over-represents male CEOs compared to the real-world distribution

💡 Hint

Think about how the model exaggerates existing biases.

Bias amplification means the model increases the imbalance seen in training data, here showing more male CEOs than expected.

❓ Model Choice

expert

Choosing a Model Architecture to Reduce Bias

You want to build a generative model that reduces bias in text generation. Which approach is most effective?

A Use a simple RNN model trained on unfiltered data

B Use a conditional generation model with fairness constraints during training

C Use a model trained only on biased data but with more epochs

D Use a larger transformer model without any bias mitigation

💡 Hint

Consider methods that actively reduce bias during learning.

Conditional generation with fairness constraints helps the model learn to avoid biased outputs by design.