Output filtering and safety checks in Agentic AI - Practice Problems & Coding Challenges

Challenge - 5 Problems

🧠 Conceptual

intermediate

Why is output filtering important in AI systems?

Imagine you have an AI assistant that answers questions. Why should the system include output filtering and safety checks before showing answers to users?

A To ensure the AI only shares safe, accurate, and appropriate information with users.

B To make the AI respond faster by skipping complex answers.

C To allow the AI to generate any content without restrictions.

D To reduce the AI's memory usage during processing.

💡 Hint

Think about what could happen if the AI shares harmful or wrong information.

Output filtering helps prevent harmful, biased, or incorrect content from reaching users, keeping interactions safe and trustworthy.

❓ Predict Output

intermediate

What is the output of this safety check code snippet?

Given this Python code that filters out unsafe words from AI output, what will be printed?

Agentic AI

unsafe_words = ['badword', 'danger']
output = 'This is a safe message.'
filtered_output = ' '.join(word for word in output.split() if word.lower() not in unsafe_words)
print(filtered_output)

A This is a safe message.

B This is a message.

C This is a safe.

D badword danger

💡 Hint

Check if any words in the output match the unsafe words list.

The output contains no unsafe words, so the original message prints unchanged.

❓ Model Choice

advanced

Which model architecture best supports output safety checks?

You want to build an AI that can self-monitor and filter its own outputs for safety. Which model design helps most?

A A convolutional neural network trained on images.

B A pipeline combining a language model with a separate safety classifier module.

C A single large language model without any internal safety layers.

D A simple linear regression model.

💡 Hint

Think about separating generation and safety checking tasks.

Using a pipeline with a safety classifier allows the system to detect and block unsafe outputs effectively.

❓ Hyperparameter

advanced

Which hyperparameter adjustment can reduce unsafe outputs in text generation?

When generating text with a language model, which hyperparameter change helps reduce risky or harmful content?

A Increasing temperature to a high value like 1.5.

B Using a batch size of 64.

C Increasing max token length to 1000.

D Setting temperature to a low value like 0.2.

💡 Hint

Lower temperature makes outputs more focused and less random.

Lower temperature reduces randomness, making outputs more predictable and less likely to produce unsafe content.

🔧 Debug

expert

Why does this output filtering code fail to block unsafe words?

Review this Python code meant to block unsafe words from AI output. Why does it fail to filter 'Danger'?

Agentic AI

unsafe_words = ['danger']
output = 'This is a Danger.'
filtered_output = ' '.join(word for word in output.split() if word.lower() not in unsafe_words)
print(filtered_output)

A Because the code uses lower() but does not apply it to output words.

B Because the unsafe_words list is empty.

C Because 'Danger.' includes punctuation, it does not match 'danger' exactly.

D Because join() is used incorrectly.

💡 Hint

Check how punctuation affects string matching.

The word 'Danger.' has a period, so it does not match 'danger' exactly, causing the filter to miss it.