Navigating the Ethical Landscape of AI, Explained

UPI Study › Courses › Introduction to Artificial Intelligence › Free lessons › Lesson 9.1 Free with a free account

Navigating the Ethical Landscape of AI

Introduction to Artificial Intelligence · Chapter 9: AI Applications, Ethics, and the FutureSection 9.1 · lesson 62 of 6915 min · free with a free account

← Designing and Programming Advanced Robotic PlatformsNext lesson: AI in Everyday Life: Current Applications →

Earn 3 college credits for Introduction to Artificial IntelligenceRecommended for credit by NCCRS · Computer Science 373 · self-paced · one payment, lifetime access
Enroll in Introduction to Artificial Intelligence · $250 →Download the syllabus (PDF)

In this lesson

Frameworks and tools for fairness, privacy, explainability and accountability in AI.

  • Compare utilitarian, deontological and virtue ethics as lenses for AI decisions
  • Identify sources of algorithmic bias and explain why fairness measures conflict when base rates differ
  • Explain privacy protections such as data minimization, differential privacy, randomized response and federated learning
  • Describe how explainability, accountability practices and defenses against adversarial attacks support trustworthy AI

Free study notes

Get this lesson's notes by email

The key points and 4 practice questions with answers, in your inbox. One email, free.

One email. Unsubscribe in one click.

Transcript

Everything said in the video, section by section (1,969 words). Read along, or search it.

Opening Question

Faster Routes, Slower Neighborhood

A delivery company tests a new route-planning AI. It cuts total delivery time by 12 percent. But in one neighborhood, service becomes 30 percent slower. Should the company switch it on? When an AI system touches thousands of lives, how do we decide whether it's fair, who answers when it fails, and what it may know about us?

Ethical Frameworks

Utilitarian, Deontological and Virtue Lenses

Ethical frameworks are structured ways of asking what the right thing to do is. Three traditions dominate discussions of AI. A utilitarian looks at consequences: add up the gains and losses. If the total benefit clearly outweighs the harm, deploy. Its weakness: it can accept serious harm to a minority. A deontologist looks at duties and rights. Every neighborhood is owed fair service, so a large, unexplained loss for one group is a problem. Its weakness: duties can conflict. Virtue ethics asks what a fair, honest and caring company would do: probably look for a version that doesn't sacrifice one area, and tell residents what changed. Its weakness: vaguer guidance.

Core Principles

Five Principles of AI Ethics

The three lenses often disagree, and that's useful: each shows something the others miss. In practice, teams combine them. They estimate overall benefits, check for violations of rights and fairness, and ask whether the decision reflects their values. Most AI ethics guidelines boil this down to five principles: fairness, privacy, transparency, accountability, and safety. The rest of this lesson takes them in turn.

Algorithmic Bias

Where Unfair Differences Come From

Algorithmic bias means systematic, unfair differences in how a system treats, or performs for, different groups. It needs no bad intentions: data records a world that may already be unequal. Historical bias: the data accurately records an unfair world, like a lending history shaped by past discrimination. Sampling bias: some groups are scarce in the data, like a face model trained mostly on light-skinned faces. Measurement bias: a convenient stand-in replaces what we care about, like past arrests standing in for crime, when arrests depend on where police patrol. Aggregation and evaluation bias: one model fitted to groups that differ, or one average accuracy hiding poor results for a small group. And deployment bias and feedback loops. A tool that sends more patrols to some areas finds more incidents there, and then sends even more.

Proxies

Why Deleting Race or Gender Fails

This is real. Audits of commercial face-analysis systems found error rates as high as about 35 percent for darker-skinned women, against under 1 percent for lighter-skinned men. A tempting fix is to delete the sensitive attribute, like race or gender. It rarely works, because other features act as proxies. A postal code can reveal a lot about race or income. And deleting the attribute removes the ability to check whether outcomes are fair across groups.

Fairness Measures

Parity, Opportunity, Odds and Precision

To judge fairness, compare how a model treats different groups. Four measures are common, and each takes a different view of fair. Demographic parity: equal approval rates across groups. Equal opportunity: among people who truly qualify, each group is approved equally often. That's equal true positive rates. Equalized odds adds a second condition: unqualified people are also wrongly approved equally often. Equal true and false positive rates. And predictive parity: among those approved, the same share of each group truly qualifies. That's equal precision.

Worked Example

One Threshold, Two Base Rates

A worked example. A loan model sees two groups of 200 applicants. In group A, 120 would repay. In group B, only 80 would. One threshold for everyone catches 80 percent of those who'd repay, and wrongly approves 20 percent of those who wouldn't, in both groups. Group A: 96 correctly approved, 16 wrongly. Group B: 64 correctly, 24 wrongly. So group A's approval rate is 0.56, and group B's is 0.44. Precision is 0.86 against 0.73. Equalized odds holds exactly. But demographic parity fails, and so does predictive parity. The cause: different base rates of repayment.

Conflicting Measures

Matching Approval Rates Breaks Equalized Odds

Now the lender lowers group B's threshold, until its approval rate matches group A's 0.56. Under the example's score assumption, other things move. Group B's true positive rate rises to 0.90, and its false positive rate to 0.34. Precision falls to 0.64. Equalized odds is now broken. This isn't a quirk. Mathematical results show that when base rates differ, a model that isn't perfect can't satisfy several of these measures at once. A well-known dispute over a criminal risk score turned on exactly this: critics pointed to unequal false positive rates, and its makers to equally accurate scores. So teams must choose the measure that fits the harm they most want to avoid, and explain it. For loans, maybe equal opportunity. For screening, false positive rates.

Quick Check

Which Fairness Measure Holds

Quick check. Group X has 100 applicants, and 60 would repay. The model approves 50, and 45 of them would repay. Group Y has 100 applicants, and 40 would repay. The model approves 50, and 36 of them would repay. Which is the same for X and Y? Pause, and decide. The fraction accepted overall: 50 percent in both groups. So demographic parity holds. Equal opportunity fails: X approves 45 of its 60 qualified people, 75 percent, but Y approves 36 of 40, 90 percent. And predictive parity fails: 45 of 50 approved in X would repay, 90 percent, against 36 of 50 in Y, 72 percent. Different base rates again.

Reducing Bias

Before, During and After Training

Bias can be reduced at several stages. Before training, improve the data: collect more examples from under-represented groups, reweight them, and fix unreliable labels. During training, add a fairness requirement: penalize the model when its error rates differ across groups. That usually costs a little overall accuracy. After training, adjust decisions, for instance with different thresholds per group. That raises legal and ethical questions of its own. And throughout: test results by group, review with the people affected, monitor after launch, and offer an appeal. That's ongoing work, not a one-time check.

Privacy

Inference, Re-Identification and Memorization

Privacy is a person's ability to control information about themselves. AI raises four special concerns. Inference: models can work out sensitive facts, like health or politics, from harmless-looking purchases or posts. Re-identification: removing names often isn't enough. One well-known study estimated that about 87 percent of people in the United States could be identified from just a postal code, birth date and sex. Aggregation: many small, separate pieces of data, combined, reveal a detailed picture. And memorization: models can sometimes reveal parts of their training data.

Consent and Protections

Minimization, Differential Privacy and Federated Learning

Consent means a person agrees to how their data will be used. Meaningful consent is informed, specific, freely given, and withdrawable. In practice it's often weak. Long terms go unread, agreement is bundled with the service, and data collected for one purpose may later train a model for another. Ownership is even less settled. Is data about you owned by you, by the company that collected it, or by the service storing it? And what if you withdraw it? Practical protections help. Data minimization: collect only what's needed, and keep it only as long as needed. Anonymization, remembering re-identification is possible. Differential privacy adds calibrated noise, so results barely change whether or not any one person is included. Federated learning trains on people's own devices, sharing only model updates. And access controls with audit logs limit who can see data, and record who did.

Randomized Response

A Coin Flip That Hides Each Answer

The idea behind differential privacy shows up in an old poll trick: randomized response. A company wants the share of employees who've done something embarrassing. Each person flips a coin in private. Heads: answer truthfully. Tails: flip again, and say yes on heads, no on tails. One person's yes proves nothing: it could be the coin. Yet the group can be measured. If the true share is p, a yes comes with probability p over two, plus a quarter. So p is the observed share, minus 0.25, over 0.5. If 400 of 1,000 say yes, p is 0.15 over 0.5: 0.30. The price of privacy is noise. In one simulation of 2,000 repeated polls, the estimates spread about 3 percentage points either way. More people, more accuracy.

Transparency and Explainability

A Counterfactual Explanation for a Loan

Transparency is openness about the whole system: that AI is used, for what, on what data, and with what limits. Explainability is narrower: why this particular output? Short decision trees and simple scores are interpretable by design. Complex models need explanations after the fact, like which inputs mattered, or which change would flip the answer. A toy lender scores two times income, minus three times debts, plus years employed, in tens of thousands of dollars, and approves at 20. An applicant earns 60,000, owes 20,000, and has worked 4 years. Twelve, minus six, plus four: 10. Denied. Twenty thousand more income gives 14. No debts gives 16. Both together give exactly 20: approved. So does 50,000 more income alone. That's a counterfactual explanation: something the applicant can act on. But for complex models, explanations are approximations, and a persuasive one isn't proof the decision was fair.

Accountability and Safety

Who Answers When AI Causes Harm

Accountability means that when an AI system causes harm, someone can be identified, questioned, and required to put things right. With data providers, developers, deployers and users all involved, that's hard. Five practices help. Documentation of models and data sets. Logs and audit trails. Impact assessments before launch. Human oversight. And a way to appeal and get errors corrected. Safety means the system behaves as intended, even in unusual conditions: testing, monitoring, safe fallbacks, and people overseeing high-stakes choices. One safety problem is specification gaming: told to maximize clicks, a content system may promote sensational material. Making a system's goals match what people intend is called alignment.

Attacks on Models

Adversarial Examples and Data Poisoning

AI also adds new ways to attack a system, aimed at the model itself. Adversarial examples are inputs altered in tiny, deliberate ways that cause a wrong output. They attack at test time. Why tiny changes work: take 100 inputs, each with a weight of 0.5, and a decision margin of 1.5. Nudge every input by 0.04, in the harmful direction. Each nudge moves the score by only 0.02, but a hundred of them add up to 2.0: past the margin. The more inputs, as in an image, the easier this gets. Data poisoning corrupts the training data instead, to degrade the model or plant a hidden trigger. In one toy test, random label errors did no harm, but crafted points at 14 percent reversed the classifier. Defenses: check where data comes from, screen it for outliers, limit who can add to it, and test against known attacks.

Quick Check

Do 400 Tiny Nudges Flip the Decision

Quick check. A model has 400 inputs, each with a weight of 0.2, and a decision margin of 3. An attacker nudges every input by just 0.05, the harmful way. Does the decision flip? Pause, and work it out. Yes. Each nudge shifts the score by 0.05 times 0.2: 0.01. Four hundred of them add up to 4.0, more than the margin of 3. Judging by one tiny shift is the trap. It's the sum over many inputs that does the damage.

Back to Routing

Applying Every Lens and Tool

Back to the routing AI. A utilitarian tallies the 12 percent gain against the 30 percent loss. A deontologist asks whether that neighborhood is owed better. A virtue lens looks for a fairer version. The toolkit adds the rest: measure the effect group by group, protect the data the system uses, explain the change to residents, and name who answers for it.

Key Takeaways

AI Ethics in Summary

To sum up. Utilitarian, deontological and virtue lenses often disagree, and each reveals something the others miss. Bias can enter at every stage, and deleting a sensitive attribute doesn't remove it. Fairness measures conflict when base rates differ, so teams must choose and explain. Privacy needs real consent and protections like minimization, differential privacy and federated learning. Explanations and accountability let people challenge decisions, while adversarial examples and data poisoning attack the model itself.

Think About It

Choosing One Fairness Measure for Loans

Here's something to think about. You must pick one fairness measure for a tool that screens loan applications. Which would you pick, and whose interests would it favor? Which ethical lens shaped your answer, and what would someone using another lens say?

About Introduction to Artificial Intelligence at UPI Study

Introduction to Artificial Intelligence (Computer Science 373) is a self-paced online course of 69 video lessons in 9 chapters, from intelligent agents and search to game strategy, neural networks, logic and reasoning, machine learning, language, robotics and AI ethics. It carries a NCCRS credit recommendation for 3 semester credits.

Every lesson transcript is free on pages like this one, and the videos are free with a free UPI Study account. Enrolling adds a practice quiz after each lesson, the graded work and final exam that earn the credit, lifetime access, and an official transcript on completion. UPI Study has helped over 130,000 students with college credits. ACE and NCCRS have over 2,100 cooperating colleges that will consider accepting credits; the final decision is with the universities.

Enroll now → Course details Download the syllabus All Computer Science courses Every course with PRO →

Compare: AI vs Current Trends in CS and IT

Why take it here

Why students take Introduction to Artificial Intelligence with UPI Study

How machines search, reason and learnIntroduction to Artificial Intelligence covers how AI systems are built: intelligent agents, search and heuristics, game strategy, logic, neural networks, machine learning, language and perception, and the ethics of AI. It is a common upper-level course in computer science degrees.
Credit at a fraction of the priceThe same 3 credits can cost $750 or more at a US college. Here it is one payment, self-paced, with lifetime access.
Finish your degree soonerTransfer credits reduce the courses you still owe, so you can graduate earlier or lighten a heavy semester.
Strengthen an applicationA completed college-level course on an official transcript shows admissions and scholarship committees you can do the work.
Study on your own scheduleNo start dates, no deadlines. Watch the lessons, practice, and take the final when you are ready.
130,000+students helped with college creditsACE and NCCRS have over 2,100 cooperating colleges that will consider accepting credits; the final decision is with the universities.

Where the credit goes

Universities that work with NCCRS

Introduction to Artificial Intelligence carries a NCCRS credit recommendation. These universities work with NCCRS; some are UPI Study's signed partners. ACE and NCCRS have over 2,100 cooperating colleges that will consider accepting credits; the final decision is with the universities.

Thomas Edison State UniversityThomas Edison State UniversitySigned partnerUniversity of the PeopleUniversity of the PeopleSigned partnerCharter Oak State CollegeCharter Oak State CollegeSigned partnerStrayer UniversityStrayer UniversitySigned partnerColumbia Southern UniversityColumbia Southern UniversitySigned partnerIndiana TechIndiana TechSigned partnerPenn StatePenn StateWorks with NCCRSSouthern New Hampshire UniversitySouthern New Hampshire UniversityWorks with NCCRSSUNYSUNYWorks with NCCRS +many more cooperating universitiesFind my college →

Find my college → Not listed? Ask Athena in the corner of this page, or book a 15-minute call.

FAQ

Questions students ask us

About Navigating the Ethical Landscape of AI

What causes algorithmic bias in AI systems?
Algorithmic bias usually comes from data and design choices rather than bad intentions. Historical bias records an already unfair world, sampling bias leaves some groups scarce in the data, and measurement bias uses a convenient stand-in, such as arrests for crime. Aggregation and evaluation bias fit one model or one average to groups that differ, and feedback loops after deployment can amplify early differences, as when more patrols find more incidents.
Why can't a model satisfy every fairness measure at once?
When groups have different base rates and the model is not perfect, measures such as demographic parity, equalized odds and predictive parity generally cannot all hold together. In one loan example with repayment rates of 60 and 40 percent, a shared threshold gave equal true and false positive rates, but approval rates of 0.56 versus 0.44 and precision of 0.86 versus 0.73. Teams must choose the measure that fits the harm they most want to avoid.
How does randomized response protect privacy?
Randomized response lets each person answer a sensitive question behind a private coin flip, so any single yes proves nothing. On heads they answer truthfully; on tails they flip again and say yes on heads and no on tails. A yes then occurs with probability p/2 plus 0.25, so the true share p equals the observed share minus 0.25, divided by 0.5. If 40 percent say yes, p is 0.30.
What is an adversarial example in machine learning?
An adversarial example is an input altered in tiny, deliberate ways so that a model gives a wrong output at test time. Small changes add up across many inputs: with 400 inputs weighted 0.2, nudging each by 0.05 shifts the score by 0.01 per input, or 4.0 in total, enough to cross a decision margin of 3. Data poisoning differs because it corrupts the training data instead.

About the Introduction to Artificial Intelligence course

How many college credits is Introduction to Artificial Intelligence?
Introduction to Artificial Intelligence (Computer Science 373) carries a NCCRS credit recommendation for 3 semester credits. ACE and NCCRS have over 2,100 cooperating colleges that will consider accepting credits; the final decision is with the universities.
How much does Introduction to Artificial Intelligence cost and what is included?
$250, paid once, with lifetime access: all 69 video lessons, a practice quiz after each, the graded work and final exam, and an official transcript when you finish. Students who want several courses choose UPI Study PRO instead.
Which universities accept UPI Study credits?
UPI Study has signed partner agreements with universities such as Thomas Edison State University, University of the People and Charter Oak State College, and works with universities such as Penn State, SNHU and SUNY. ACE and NCCRS have over 2,100 cooperating colleges that will consider accepting credits; the final decision is with the universities. Use "Find my college" on this page to check yours.
How long does Introduction to Artificial Intelligence take?
It is self-paced with no deadlines. The 69 lessons total about 16 hours of video; most students add reading and practice time and finish in a few weeks to a few months, on their own schedule.
How are the exams done?
Quizzes and assignments are online in the course. The final exam is taken online with proctoring; a person on the UPI Study team checks assignments, usually within two days.
How do my credits reach my college?
When the course shows 100% complete, request the official transcript from your dashboard. It is sent to the college you name, and you can send it to more than one. Check your college first with "Find my college" on this page, or ask Athena.

Ready for the full course? Earn 3 credits in Introduction to Artificial Intelligence.

69 video lessons, a quiz after each one, and the graded work that earns the credit. Recommended for credit by NCCRS.

Enroll in Introduction to Artificial Intelligence →Download the syllabusEvery course with PRO →
Introduction to Artificial Intelligence
3 credits · NCCRSEnroll →