Output filtering and safety checks in Agentic AI - Model Metrics & Evaluation

Metrics & Evaluation - Output filtering and safety checks

Which metric matters for output filtering and safety checks and WHY

For output filtering and safety checks, the key metrics are False Positive Rate and False Negative Rate. False positives mean safe content is wrongly blocked, which hurts user experience. False negatives mean unsafe content is missed, which can cause harm. Balancing these is critical to keep outputs safe without blocking too much useful content.

Confusion matrix for output filtering

          | Predicted Safe | Predicted Unsafe
    ------|----------------|-----------------
    Actual Safe   |      TN=85     |      FP=15
    Actual Unsafe |      FN=10     |      TP=90
    

Here, TP means unsafe content correctly blocked, FP means safe content wrongly blocked, TN means safe content correctly allowed, and FN means unsafe content missed.

Precision vs Recall tradeoff with examples

Precision measures how many blocked outputs are truly unsafe. High precision means few safe outputs are blocked (low false positives).

Recall measures how many unsafe outputs are caught. High recall means few unsafe outputs slip through (low false negatives).

Example: If you block too much (high recall), users get annoyed by safe content blocked (low precision). If you block too little (high precision), unsafe content may appear (low recall). Finding the right balance depends on the use case.

What good vs bad metric values look like

Good: Precision and recall both above 90%, meaning most unsafe content is blocked and most safe content is allowed.

Bad: Precision below 70% means many safe outputs blocked, hurting user trust. Recall below 50% means many unsafe outputs missed, risking harm.

Common pitfalls in output filtering metrics

• Accuracy paradox: If unsafe content is rare, a model blocking nothing can have high accuracy but be useless.
• Data leakage: If test data leaks into training, metrics look better than real performance.
• Overfitting: Model blocks training unsafe content well but fails on new unsafe content.

Self-check question

Your output filter has 98% accuracy but only 12% recall on unsafe content. Is it good for production? Why or why not?

Answer: No, it is not good. The low recall means it misses 88% of unsafe content, which can cause harm. High accuracy is misleading because most content is safe, so blocking nothing looks accurate but unsafe outputs slip through.

Key Result

Balancing false positives and false negatives is key to effective output filtering and safety checks.

Practice

1. What is the main purpose of output filtering in AI systems?

easy

A. To stop unsafe or unwanted AI results from reaching users

B. To speed up the AI model training process

C. To increase the size of the AI model

D. To add more data to the training set

Solution

1. Step 1: Understand output filtering

   Output filtering is designed to prevent unsafe or unwanted content from being shown to users.

2. Step 2: Compare options

   Only To stop unsafe or unwanted AI results from reaching users describes stopping unsafe or unwanted results, which matches the purpose of output filtering.

3. Final Answer:

   To stop unsafe or unwanted AI results from reaching users -> Option A

4. Quick Check:

   Output filtering = stopping unsafe results [OK]

Hint: Output filtering blocks bad or unsafe AI outputs [OK]

Common Mistakes:

• Confusing filtering with training speed
• Thinking filtering adds data
• Assuming filtering changes model size

2. Which of the following is a correct way to check if an AI output contains a banned word in Python?

easy

A. if output_text.contains(banned_word):

B. if output_text == banned_word:

C. if output_text.index(banned_word):

D. if banned_word in output_text:

Solution

1. Step 1: Recall Python syntax for substring check

   In Python, to check if a substring is in a string, use the 'in' keyword.

2. Step 2: Evaluate options

   if banned_word in output_text: uses 'if banned_word in output_text:', which is correct. if output_text == banned_word: checks equality, not containment. if output_text.contains(banned_word): uses a method that doesn't exist in Python strings. if output_text.index(banned_word): uses index incorrectly and can cause errors.

3. Final Answer:

   if banned_word in output_text: -> Option D

4. Quick Check:

   Substring check in Python uses 'in' [OK]

Hint: Use 'in' keyword to check substring in Python strings [OK]

Common Mistakes:

• Using equality instead of containment
• Using non-existent string methods
• Using index without error handling

3. Given this Python code snippet for filtering AI output:

output = "Hello user!"
banned_words = ["bad", "ugly"]
filtered = any(word in output for word in banned_words)
print(filtered)

What will be the printed output?

medium

A. False

B. True

C. Error

D. None

Solution

1. Step 1: Understand the code logic

   The code checks if any banned word is in the output string using 'any()' with a generator expression.

2. Step 2: Check banned words in output

   Output is "Hello user!". Neither "bad" nor "ugly" is in this string, so 'any()' returns False.

3. Final Answer:

   False -> Option A

4. Quick Check:

   None of banned words in output = False [OK]

Hint: If no banned words found, 'any' returns False [OK]

Common Mistakes:

• Assuming 'any' returns True by default
• Confusing 'any' with 'all'
• Expecting an error from generator expression

4. This code is meant to filter AI output for banned words but causes an error:

output = "Safe text"
banned_words = ["bad", "ugly"]
for word in banned_words:
    if output.index(word):
        print("Banned word found")
        break

What is the error and how to fix it?

medium

A. Syntax error due to missing colon after for loop

B. index() raises ValueError if word not found; use 'in' instead

C. TypeError because output is not a list

D. No error; code works fine

Solution

1. Step 1: Identify the error cause

   Using output.index(word) raises ValueError if word is not found in output string.

2. Step 2: Suggest fix

   Replace 'output.index(word)' with 'word in output' to safely check containment without error.

3. Final Answer:

   index() raises ValueError if word not found; use 'in' instead -> Option B

4. Quick Check:

   index() error fixed by 'in' check [OK]

Hint: Use 'in' to check substring safely, not index() [OK]

Common Mistakes:

• Ignoring ValueError from index()
• Thinking output must be a list
• Missing colons in loops (not in this code)

5. You want to build a safety filter that blocks AI outputs containing banned words or outputs longer than 100 characters. Which approach correctly combines these checks in Python?

hard

A. if banned_words in output or len(output) == 100:
block_output()

B. if all(word in output for word in banned_words) and len(output) < 100:
block_output()

C. if any(word in output for word in banned_words) or len(output) > 100:
block_output()

D. if output.contains(banned_words) or output.length > 100:
block_output()

Solution

1. Step 1: Understand filtering conditions

   The filter should block if any banned word is present OR output length exceeds 100 characters.

2. Step 2: Evaluate options for correct logic and syntax

   if any(word in output for word in banned_words) or len(output) > 100: block_output() uses 'any' to check banned words and 'or' for length > 100, which is correct. if all(word in output for word in banned_words) and len(output) < 100: block_output() uses 'all' and 'and' incorrectly. if output.contains(banned_words) or output.length > 100: block_output() uses invalid methods. if banned_words in output or len(output) == 100: block_output() uses wrong containment and equality checks.

3. Final Answer:

   if any(word in output for word in banned_words) or len(output) > 100: block_output() -> Option C

4. Quick Check:

   Use 'any' + 'or' for combined filter [OK]

Hint: Use 'any' with 'or' to combine banned words and length checks [OK]

Common Mistakes:

• Using 'all' instead of 'any' for banned words
• Using 'and' instead of 'or' to combine conditions
• Using invalid string methods like contains()