Skip to main content

Ethics and Verification Techniques in AI

Understanding AI Ethics​

Why Do We Need AI Ethics?​

Imagine if your GPS always directed people from certain neighborhoods through longer routes, or if a job application system automatically rejected resumes with certain names. These scenarios show why we need ethical guidelines for AI.

Key Ethical Concerns:

Fairness and Bias​

What it means: AI should treat all people fairly, regardless of their background.

Real-world example: In 2018, Amazon discovered their AI recruiting tool was biased against women because it was trained on resumes from previous years when most hires were men. The AI learned to prefer male candidates, which wasn't fair.

Why it happens: AI learns from data created by humans, and if that data contains human biases, the AI will learn those biases too.

Privacy and Data Protection​

What it means: AI systems often need personal information to work, but they should protect our privacy.

Real-world example: When you use a fitness app that tracks your daily steps, it knows your location, health habits, and daily routine. This information should be kept private and secure.

Key questions:

  • What data is being collected?
  • How is it being used?
  • Who has access to it?
  • Can you control or delete your data?

Transparency and Explainability​

What it means: People should understand how AI makes decisions that affect them.

Real-world example: If a bank's AI system denies your loan application, you should know why. Was it because of your credit score, income, or some other factor? You deserve an explanation.

The "Black Box" Problem: Sometimes AI systems are so complex that even their creators don't fully understand how they make decisions. This is like having a judge who can't explain their verdict.

Accountability and Responsibility​

What it means: Someone should be responsible when AI systems make mistakes or cause harm.

Real-world example: If a self-driving car causes an accident, who is responsible? The car manufacturer? The software company? The owner? This question becomes more important as AI becomes more autonomous.

Human Agency and Oversight​

What it means: Humans should maintain control over important decisions, even when AI is involved.

Real-world example: While AI can help doctors diagnose diseases by analyzing X-rays, the final decision about treatment should always involve human doctors who can consider the full context of a patient's situation.

Ethical Frameworks for AI​

The "Four Principles" Approach​

  1. Beneficence: AI should do good and benefit humanity
  2. Non-maleficence: AI should not cause harm ("do no harm")
  3. Autonomy: AI should respect human choice and decision-making
  4. Justice: AI should be fair and treat people equally

The "Rights-Based" Approach​

  • Right to privacy
  • Right to explanation
  • Right to human oversight
  • Right to non-discrimination

Ethics in Programming and Software Development​

As future professionals, you may work with or create AI systems. Here's how ethics applies to programming:

Programming Decisions That Matter​

Data Collection:

Simple Example:
A mobile app asks for phone contacts "to find friends"
But actually uses them for targeted advertising

Ethical Programming Choice:
- Only collect data you actually need
- Be clear about how data will be used
- Give users control over their data

Algorithm Design:

Example Problem:
A job matching app sorts candidates by "compatibility score"

Ethical Considerations in Code:
- What factors does the score include?
- Does it accidentally favor certain demographics?
- Can the scoring logic be explained to users?

Code Review for Ethics​

Just like checking for bugs, programmers should check for ethical issues:

Before Writing Code:

  • Who will this affect?
  • What data am I using?
  • Could this create unfair outcomes?

During Development:

  • Am I being transparent about what the system does?
  • Have I tested with diverse user groups?
  • Can users understand and control the system?

After Deployment:

  • Am I monitoring for unintended consequences?
  • Can users report problems?
  • Am I fixing issues when they arise?

Programming Examples​

Example 1: A Search Algorithm

Bad Approach:
function searchProfiles(keyword) {
    // Only shows results from users with premium accounts first
    return sortByPremiumStatus(results);
}

Better Approach:
function searchProfiles(keyword) {
    // Shows most relevant results regardless of account type
    return sortByRelevance(results);
    // Note: Clearly document why results are ordered this way
}

Example 2: User Recommendation System

Problematic:
// Recommends content based only on past behavior
// This can create "filter bubbles"

More Ethical:
// Includes diversity in recommendations
// Allows users to see why items were recommended
// Lets users adjust their preferences

Case Studies in AI Ethics​

Case Study 1: Programming a Content Moderation System​

Scenario: You're asked to create an AI system that automatically removes "inappropriate" social media posts.

Programming Challenges:

  • Training Data: If your training data contains biased human decisions, your AI will learn those biases
  • Edge Cases: How do you handle sarcasm, cultural differences, or context?
  • Transparency: Can you explain to users why their post was removed?
  • Appeals Process: How do users challenge the AI's decision?

Ethical Programming Practices:

  • Test with diverse content and user groups
  • Build in human review processes
  • Provide clear explanations for decisions
  • Allow users to appeal automated decisions

Case Study 2: Developing a Credit Scoring Algorithm​

Scenario: A bank wants you to create an AI system that automatically approves or denies loan applications.

Programming Ethical Considerations:

  • Feature Selection: Which data points should influence the decision? (Income: yes, Zip code: maybe problematic)
  • Historical Bias: Past lending data might reflect discriminatory practices
  • Explainability: Applicants have a right to know why they were denied
  • Testing: You must verify the system doesn't discriminate against protected groups

Code Implementation Considerations:

Ethical Checkpoint Questions:
1. What features am I using? Are any potentially discriminatory?
2. How diverse is my training data?
3. Can I explain each decision the system makes?
4. Have I tested for bias across different demographic groups?
5. Is there a human review process for disputed decisions?

Verification Techniques in AI​

What is AI Verification?​

AI verification is like quality control for AI systems. Just as we test cars for safety before they hit the road, we need to test AI systems before they make important decisions.

Think of it like this: Before a pilot flies a plane, they go through a pre-flight checklist to make sure everything is working properly. Similarly, AI systems need thorough checking before we trust them with important tasks.

Why Do We Need to Verify AI?​

Real-world Consequences​

  • Medical AI: An AI that misdiagnoses cancer could delay life-saving treatment
  • Financial AI: An AI that makes bad investment decisions could lose people's retirement savings
  • Transportation AI: A self-driving car that misjudges distances could cause accidents

Types of AI Verification​

Testing with Known Examples​

What it means: We give the AI problems where we already know the correct answer.

Daily life example: It's like giving a student a practice test before the real exam. If they get the practice questions wrong, they're not ready for the real test.

How it works:

  • Prepare a set of examples with known correct answers
  • Have the AI make predictions on these examples
  • Compare the AI's answers to the correct answers
  • Calculate accuracy percentage

Cross-Validation​

What it means: We test the AI on data it hasn't seen before.

Daily life example: Imagine you're learning to cook. You practice with recipes from a cookbook, but the real test is whether you can cook a good meal using a completely different recipe.

How it works:

  • Split data into "training" and "testing" groups
  • Train AI on the training data
  • Test AI on the testing data (which it has never seen)
  • This tells us how well the AI will perform on new, real-world data

Bias Testing​

What it means: We check if the AI treats different groups of people fairly.

Daily life example: If you're a teacher, you want to make sure your grading is fair to all students, regardless of their background. You might review your grades to see if there are any unfair patterns.

How it works:

  • Test the AI's performance on different demographic groups
  • Look for significant differences in accuracy or outcomes
  • If the AI performs much worse for certain groups, it may be biased

Robustness Testing​

What it means: We check if the AI still works well when conditions change slightly.

Daily life example: A good driver can still drive safely in light rain, at night, or on unfamiliar roads. Similarly, good AI should work well even when conditions are slightly different from its training.

How it works:

  • Test the AI with slightly modified inputs
  • Add small amounts of "noise" or distraction to the data
  • See if the AI's performance drops significantly

Adversarial Testing​

What it means: We try to "fool" the AI to find its weaknesses.

Daily life example: Security guards might try to break into their own building to find weaknesses in their security system.

How it works:

  • Create inputs designed to confuse the AI
  • See if small, intentional changes can make the AI give wrong answers
  • Identify vulnerabilities that could be exploited

Verification Methods in Practice​

Audit Trails​

What it means: Keeping detailed records of how AI systems are built and tested.

Daily life example: Like keeping receipts for your purchases, audit trails help track what was done and when.

Human-in-the-Loop Testing​

What it means: Having humans involved in the testing process.

Daily life example: Even though we have spell-check software, we still have human editors review important documents.

Continuous Monitoring​

What it means: Watching AI systems continuously after they're deployed.

Daily life example: Like monitoring your car's performance with regular maintenance checks, AI systems need ongoing monitoring to ensure they keep working properly.

Discussion Questions​

  1. Should AI systems ever make decisions without human oversight? When might this be appropriate or inappropriate?
  2. How can we balance the benefits of AI (like improved efficiency) with privacy concerns?
  3. What role should the public play in deciding how AI is used in their communities?