Lesson 1

What Are Ethics, and Why Do They Matter for AI?

Ethical frameworks and why AI systems require them.

When Facebook's content moderation team had to decide whether to remove a video of police violence in the US, they applied their stated community standards. When the same team had to decide about similar content from a conflict zone in Myanmar, the same rules produced different consequences — in one case protecting users from disturbing content, in another potentially suppressing evidence of atrocities. The same rule, applied consistently, produced outcomes that many considered ethically opposite. This is not a bug in ethics — it's what ethics is for: navigating situations where rules produce competing goods and competing harms simultaneously.

Three Ethical Frameworks

Philosophy has developed several systematic frameworks for ethical reasoning:

Consequentialism: Actions are right or wrong based on their outcomes. The right action maximizes good outcomes. Challenge: who defines "good," and how do you predict consequences?
Deontology: Actions are right or wrong based on whether they follow moral rules, regardless of consequences. The right action follows duty or principle. Challenge: what do you do when rules conflict?
Virtue ethics: Focuses on the character of the actor rather than rules or outcomes. The right action is what a person of good character would do. Challenge: who defines good character, and across which cultures?

Why AI Needs Ethical Frameworks

AI systems make consequential decisions at enormous scale. Without explicit ethical frameworks, they default to optimizing whatever metric they were given — which may not capture what we actually value. Ethics isn't a constraint on AI design; it's a prerequisite for it.

Key Point

Every AI system encodes ethical choices — about what to optimize, whose interests to count, and what tradeoffs to accept. Making those choices explicitly and thoughtfully is better than making them accidentally.

Back Take Quiz

Quiz 1

What Are Ethics?

4 questions — free, untracked, retake anytime.

does the Facebook content moderation example illustrate about ethics?

✓ Correct — ✅ ✓ The same rule can produce ethically opposite outcomes in different contexts. Ethics requires judgment about context and consequences, not just consistent rule application.

❌ ❌ The example shows that consistent rule application can produce ethically opposite outcomes. Ethics isn't just rule-following — it requires judgment about context and consequences.

is the core challenge with consequentialism as an ethical framework?

✓ Correct — ✅ ✓ Consequentialism requires defining 'good' (whose? which values?) and predicting consequences — both extremely difficult at the scale and complexity of AI systems.

❌ ❌ The core challenge: someone must define 'good,' and predicting consequences of AI decisions at scale is extremely difficult.

does every AI system encode ethical choices, whether explicitly or not?

✓ Correct — ✅ ✓ Every AI design decision — what to optimize, whose interests count, what tradeoffs to accept — is an ethical choice. Making those choices explicitly is better than making them accidentally.

❌ ❌ AI design requires choosing what to optimize, whose interests to count, and what tradeoffs to accept. These are inherently ethical choices — made explicitly or accidentally.

distinguishes deontology from consequentialism?

✓ Correct — ✅ ✓ Deontology: right action follows moral rules, regardless of outcomes. Consequentialism: right action produces the best outcomes. Same action can be right under one and wrong under the other.

❌ ❌ Deontology judges by rule-following regardless of consequences. Consequentialism judges by outcomes. The same action can be right under one framework and wrong under the other.

Back to Lesson Start Lab 1

Lab 1

Ethical Framework Analysis

Apply ethical frameworks to AI scenarios.

Lab 1 — Ethical Framework Analysis

Apply the three frameworks to an AI ethics scenario.

The AI opens with a content moderation scenario and asks you to analyze it through all three frameworks.
Identify where the frameworks agree and where they conflict.
Address: which framework do you find most useful for AI ethics, and why?

Consider: content moderation, autonomous vehicles, hiring AI — does your preferred framework change by domain?

🔬 AI GuideLab 1

Back to Quiz Next: Lesson 2

Lesson 2

Harm and Benefit — Who Decides?

Defining harm in AI contexts and the political nature of those definitions.

In 2021, YouTube's algorithm was found to systematically recommend increasingly extreme content to users over time — a behavior its engineers called "rabbit holing." Internally, this was an engagement feature: users who went down rabbit holes watched more content. Externally, researchers documented correlations between rabbit-holing and radicalization. YouTube's stated goal was maximizing user engagement. Whether this constituted "harm" depended entirely on whether you counted psychological harm to users and democratic harms from radicalization alongside the platform's engagement metric. YouTube had defined harm narrowly; critics argued the definition should be broader.

How Harm Gets Defined in AI

Harm in AI systems is not a neutral technical concept — it is defined through choices:

What counts: Physical harm? Psychological harm? Societal harm? Economic harm? Long-term harm?
Whose harm: Individual users? Communities? Future generations? Democratic institutions?
At what threshold: Probable harm? Possible harm? Speculative harm?
Who decides: Engineers? Executives? Regulators? Affected communities?

The Political Nature of Harm Definitions

Defining harm is a political act because different stakeholders have different interests in the definition. Platforms have incentives to define harm narrowly (more permissive = more engagement). Regulators may define it too broadly (capturing legitimate speech). Affected communities may have the most accurate view but the least institutional power.

Design Implication

Building AI systems requires making explicit harm definitions. Those definitions should be made through processes that include affected communities — not just the people building and profiting from the systems.

Back Take Quiz

Quiz 2

Harm and Benefit

4 questions — free, untracked, retake anytime.

was YouTube's rabbit-holing behavior ethically contested?

✓ Correct — ✅ ✓ The contest was about harm definition: YouTube counted engagement; critics counted psychological harm and radicalization. Same behavior, different ethical conclusions based on different harm definitions.

❌ ❌ The ethical contest was about harm definition. YouTube defined harm narrowly (engagement focus). Critics argued the definition should include psychological harm and democratic effects from radicalization.

is defining harm in AI systems 'a political act'?

✓ Correct — ✅ ✓ Harm definition is political because stakeholders have competing interests: platforms benefit from narrow definitions; affected communities may bear the costs of under-counting harm.

❌ ❌ Harm definition is political: platforms benefit from narrow definitions (more permissive = more engagement); affected communities bear the costs of those narrow definitions.

stakeholder group often has the most accurate view of harm from AI systems, but the least institutional power to define it?

✓ Correct — ✅ ✓ Affected communities often have the most direct experience of harm but the least institutional power to shape harm definitions in AI system design.

❌ ❌ Affected communities — those who experience AI outputs — often have the most accurate view of harm but the least institutional power over how harm gets defined.

dimensions should explicit harm definitions in AI systems address?

✓ Correct — ✅ ✓ Explicit harm definitions should address: what types of harm count (physical, psychological, societal), whose harm counts, at what threshold, and who has legitimate authority to define harm.

❌ ❌ Explicit harm definitions should address: what types count, whose harm counts, at what threshold, and who has authority to define and revise those definitions.

Back to Lesson Start Lab 2

Lab 2

Harm Definition Analysis

Develop a comprehensive harm framework for an AI system.

Lab 2 — Harm Definition Analysis

Develop a comprehensive harm framework for an AI system.

The AI opens with the YouTube case and asks: how would you have defined harm for YouTube's recommendation system — and who should have had authority to set that definition?
Build a multi-dimensional harm framework for a specific AI system of your choice.
Address: what process should govern harm definition in high-impact AI systems?

Consider: affected community input, temporal scope (long-term vs immediate harm), individual vs societal harm, and the institutional design of harm governance.

🔬 AI GuideLab 2

Back to Quiz Next: Lesson 3

Lesson 3

Fairness and AI

Multiple conceptions of fairness, why they conflict, and what choosing between them means.

In 2016, ProPublica published an analysis of COMPAS — a risk assessment tool used in criminal sentencing in US courts. Their analysis found that the system was twice as likely to falsely flag Black defendants as future criminals, and twice as likely to falsely mark white defendants as low-risk. Northpointe, the company that made COMPAS, published a rebuttal: the system was equally accurate across racial groups — 65% accuracy for both Black and white defendants. Both analyses were mathematically correct. They were measuring different things. The dispute revealed that 'fairness' in AI is not one concept — it is many, and they mathematically cannot all be satisfied simultaneously.

Competing Fairness Definitions

Calibration (Northpointe's measure): Among people predicted to be high risk, the same percentage reoffend across groups. Equal accuracy across groups.
Error rate parity (ProPublica's measure): False positive rates and false negative rates are equal across groups. Equal error distribution.
Individual fairness: Similar individuals receive similar outcomes, regardless of group membership.
Counterfactual fairness: Your outcome wouldn't change if you belonged to a different group.

The Impossibility Result

Researchers proved mathematically that when base rates differ between groups, you cannot simultaneously satisfy calibration, error rate parity, and other common fairness criteria. This isn't a solvable technical problem — it's a genuine values conflict. Choosing a fairness metric means choosing whose interests to prioritize and what kind of error is worse.

Ethical Implication

Choosing a fairness metric for a consequential AI system is a values judgment — it should be made explicitly, with democratic input, not buried in technical specifications.

Back Take Quiz

Quiz 3

Fairness and AI

4 questions — free, untracked, retake anytime.

were both ProPublica and Northpointe mathematically correct about COMPAS?

✓ Correct — ✅ ✓ Both were correct — measuring different things. ProPublica: false positive rates differed across races (error rate parity violated). Northpointe: accuracy rates were equal (calibration satisfied). Both true simultaneously.

❌ ❌ Both were correct because they measured different fairness criteria. ProPublica measured error rate parity; Northpointe measured calibration. Both can be true simultaneously — and both were.

does the mathematical impossibility result in AI fairness mean?

✓ Correct — ✅ ✓ The impossibility result: when base rates differ, you can't satisfy all fairness criteria simultaneously. You must choose — and that choice is a values judgment about whose interests to prioritize.

❌ ❌ The impossibility result: when base rates differ between groups, multiple fairness criteria mathematically conflict. Choosing one is a values judgment about priorities, not a technical solution.

is 'calibration' as a fairness measure?

✓ Correct — ✅ ✓ Calibration: among people with the same predicted risk score, the same proportion have the predicted outcome across groups. Equal accuracy, regardless of which group.

❌ ❌ Calibration: among people predicted to be high risk, the same proportion actually reoffend across groups — equal accuracy rates regardless of demographic group.

should choosing a fairness metric for consequential AI involve democratic input?

✓ Correct — ✅ ✓ Choosing a fairness metric is a values judgment about whose interests matter and what errors are worse. These are moral and political questions requiring democratic input — not just technical decisions.

❌ ❌ Choosing a fairness metric encodes values: whose interests to prioritize, what error is worse. These are moral/political questions requiring democratic input — not technical decisions engineers can answer alone.

Back to Lesson Start Lab 3

Lab 3

Fairness Values Analysis

Analyze the COMPAS case and develop a fairness framework.

Lab 3 — Fairness Values Analysis

Work through the COMPAS case and develop a fairness framework.

The AI opens with the COMPAS case: both analyses were correct. Which fairness criterion should govern criminal risk assessment AI — and who should make that choice?
Develop your ethical argument for your chosen criterion, acknowledging what you're trading off.
Address: what process should govern the selection of fairness criteria in high-stakes AI systems?

Consider: what kind of error is worse in criminal justice, who bears the cost of each error type, and what democratic legitimacy requires.

🔬 AI GuideLab 3

Back to Quiz Next: Lesson 4

Lesson 4

Transparency and Honesty

What transparency obligations AI systems have — and what honesty means for AI.

In 2019, the UK's Department for Work and Pensions used an algorithm to make benefit eligibility decisions affecting millions of people. Recipients who were denied benefits received letters citing "computer says no" — without any explanation of how the decision was reached. The UK High Court later ruled this unlawful. The right to an explanation of automated decisions — enshrined in GDPR and upheld in this case — reflects a principle: when an AI system makes a consequential decision about a person, that person is entitled to understand why.

Dimensions of AI Transparency

Transparency about existence: Knowing you're interacting with an AI, not a human
Transparency about process: Understanding how the AI reaches decisions
Transparency about limitations: Knowing what the AI can't do and where it might fail
Transparency about data: Knowing what information was used to make a decision about you

What Honesty Means for AI

Honesty for AI systems goes beyond not stating falsehoods:

Non-deception: Not creating false impressions through technically true statements, framing, or selective emphasis
Non-manipulation: Not exploiting psychological vulnerabilities to influence behavior
Calibrated uncertainty: Expressing appropriate uncertainty rather than false confidence

The Right to Explanation

When AI makes consequential decisions about people — eligibility for benefits, loans, jobs — those people have a legitimate interest in understanding why. Opacity in high-stakes AI is not just a technical problem; it's an ethical and legal problem.

Back Take Quiz

Quiz 4

Transparency and Honesty

4 questions — free, untracked, retake anytime.

principle did the UK High Court affirm in the benefits algorithm case?

✓ Correct — ✅ ✓ The right to explanation: consequential automated decisions must be explainable to those affected. Opacity in high-stakes AI decision-making is legally and ethically problematic.

❌ ❌ The court affirmed: when AI makes consequential decisions about people, those people are entitled to an explanation. Opacity in high-stakes AI isn't just a technical problem — it's a rights problem.

distinguishes 'non-deception' from simply 'not lying'?

✓ Correct — ✅ ✓ Non-deception is broader: technically true statements can still deceive through framing, selective emphasis, or misleading implicature. An honest AI avoids creating false impressions, not just false statements.

❌ ❌ Non-deception goes beyond not lying: technically true statements can create false impressions through framing, selective emphasis, or implicature. Honest AI avoids false impressions, not just false statements.

is 'calibrated uncertainty' as an AI honesty requirement?

✓ Correct — ✅ ✓ Calibrated uncertainty: confidence expressed should match actual reliability. Overconfidence on unreliable outputs is a form of dishonesty — even if the outputs happen to be correct.

❌ ❌ Calibrated uncertainty: expressed confidence should match actual reliability. Projecting confidence on uncertain outputs is dishonest — even if technically true.

does 'transparency about existence' matter as an AI ethical requirement?

✓ Correct — ✅ ✓ Knowing whether you're talking to an AI or a human affects how you engage — what you share, how you interpret responses. Concealing AI identity undermines informed consent.

❌ ❌ Transparency about AI existence: people engage differently with AI vs. humans. Concealing AI identity affects how people interpret interactions and whether their engagement is truly informed.

Back to Lesson Start Lab 4

Lab 4

Transparency Framework Design

Design a transparency framework for high-stakes AI.

Lab 4 — Transparency Framework Design

Design a transparency framework for a high-stakes AI system.

The AI opens with the benefits algorithm case and asks: what transparency requirements should govern consequential automated decision-making?
Build a transparency framework covering existence, process, limitations, and data.
Address: what is the minimum transparency standard below which AI should not be permitted to make consequential decisions?

Consider: the right to explanation, different affected stakeholder needs for transparency, and the tradeoff between transparency and proprietary system protection.

🔬 AI GuideLab 4

Back to Quiz Next: Lesson 5

Lesson 5

Autonomy and Consent

Informed consent, manipulation, and respecting human agency in AI interactions.

In 2014, Facebook ran a psychological experiment on 689,003 users without their knowledge. The company manipulated what content appeared in users' news feeds — reducing positive or negative emotional content — and then measured whether this affected users' emotional states as measured by their own subsequent posts. The study found it did. The paper was published in a scientific journal. Critics argued this constituted emotional manipulation of users who had not consented to participate in a psychological experiment. Facebook argued the study fell under its terms of service. The core issue: did users know their emotional states were being experimentally manipulated?

What Informed Consent Requires

Genuine informed consent requires:

Information: Understanding what you're agreeing to — not just a terms-of-service checkbox
Voluntariness: Genuine ability to decline without significant cost
Comprehension: Actually understanding, not just having technically been told
Ongoing: Ability to withdraw consent, not just one-time agreement

Manipulation vs. Influence

Not all influence is manipulation. Legitimate influence respects autonomy:

Evidence and argument: Providing accurate information that supports good decisions
Default nudges that serve user interests: Opt-out organ donation registration
Manipulation exploits psychology: Dark patterns, variable reward, emotional triggering designed to bypass rational agency

The Autonomy Test

Does this influence help the person make better decisions for themselves? Or does it exploit psychological mechanisms to produce behavior that serves the influencer's interests at the expense of the person's own goals?

Back Take Quiz

Quiz 5

Autonomy and Consent

4 questions — free, untracked, retake anytime.

made the Facebook emotion experiment ethically contested?

✓ Correct — ✅ ✓ The experiment manipulated users' emotional states without their knowledge. Terms of service 'consent' is not informed consent to psychological experimentation — users didn't understand what they were agreeing to.

❌ ❌ Users didn't know their emotional states were being experimentally manipulated. Terms of service checkbox is not informed consent to psychological experimentation.

distinguishes legitimate influence from manipulation?

✓ Correct — ✅ ✓ Legitimate influence: providing accurate info, evidence, and interest-serving nudges. Manipulation: exploiting psychological vulnerabilities to produce behavior serving the influencer rather than the person.

❌ ❌ Legitimate influence respects rational agency (evidence, argument, interest-serving defaults). Manipulation exploits psychological mechanisms to serve the influencer's interests against the person's own goals.

four elements does genuine informed consent require?

✓ Correct — ✅ ✓ Genuine informed consent: information (understanding what you're agreeing to), voluntariness (genuine ability to decline), comprehension (actually understanding), and ongoing (ability to withdraw).

❌ ❌ Genuine informed consent requires: information about what you're agreeing to, voluntariness (genuine ability to decline without significant cost), comprehension (actually understanding), and ongoing withdrawal ability.

is the 'autonomy test' for whether AI influence is ethical?

✓ Correct — ✅ ✓ The autonomy test: influence that helps people make better decisions for themselves is legitimate. Influence that exploits psychology to produce behavior serving others at the person's expense is manipulation.

❌ ❌ The autonomy test: does this help the person make better decisions for themselves? Or exploit psychological mechanisms to serve the influencer at the expense of the person's own goals?

Back to Lesson Start Lab 5

Lab 5

Consent and Manipulation Analysis

Develop criteria for consent and manipulation in AI contexts.

Lab 5 — Consent and Manipulation Analysis

Analyze consent and autonomy in AI contexts.

The AI opens with the Facebook experiment and asks: what informed consent requirements should govern AI systems that influence users' emotional states or behavior?
Develop criteria for distinguishing legitimate AI influence from manipulation.
Address: given that most users don't read terms of service, what would meaningful consent actually require?

Consider: the distinction between terms-of-service consent and genuine informed consent, and what institutional design would make consent meaningful rather than nominal.

🔬 AI GuideLab 5

Back to Quiz Next: Lesson 6

Lesson 6

Responsibility and Accountability

Moral responsibility, distributed accountability, and the ethics of AI liability.

When a Tesla on autopilot struck and killed a pedestrian in 2019, the US National Transportation Safety Board investigated. They found the driver had been distracted (watching a video). They found Tesla's autopilot system had design limitations. They found that Tesla's marketing had created unrealistic expectations of autonomy. They found regulatory frameworks hadn't kept pace with the technology. Fault was distributed across the driver, the company, and the regulator. Nobody was the sole responsible party. This distributed responsibility is the norm, not the exception, in AI-related harms — and it poses fundamental challenges for existing liability frameworks.

The Distributed Responsibility Problem

AI-related harms typically involve distributed responsibility across:

Developers: Who designed the system and its capabilities
Deployers: Who put the system into use in a specific context
Users: Who operated the system in a specific situation
Regulators: Who set (or failed to set) appropriate standards

Traditional liability frameworks are designed for identifiable individual or corporate wrongdoers. Distributed AI responsibility doesn't fit cleanly into these frameworks.

New Liability Frameworks for AI

Strict liability for high-risk AI: Developer is liable regardless of fault — creates incentive for careful design
Duty of care: Developers must demonstrate reasonable precaution appropriate to the risk level
Insurance requirements: Mandatory insurance distributes risk

The Ethics of Accountability

Accountability matters not just for compensation — it matters for the incentive structure it creates. If AI developers bear no liability for their systems' harms, they have less incentive to invest in safety. If users bear all responsibility, they can't make informed choices.

Back Take Quiz

Quiz 6

Responsibility and Accountability

4 questions — free, untracked, retake anytime.

did the Tesla autopilot death involve distributed rather than single-party responsibility?

✓ Correct — ✅ ✓ Distributed responsibility: the system design (Tesla), marketing (Tesla), driver behavior, and regulatory failure all contributed. No single party was solely responsible — which is the norm in AI harms.

❌ ❌ Distributed responsibility: system design, marketing expectations, driver behavior, and regulatory failure all contributed. This distribution is the norm in AI-related harms.

is the ethical argument for strict liability for high-risk AI?

✓ Correct — ✅ ✓ Strict liability creates incentives: developers who bear liability regardless of fault have strong financial reasons to invest in safety. It aligns commercial and safety interests.

❌ ❌ Strict liability aligns incentives: developers who bear liability regardless of fault have strong financial reasons to invest in safety — aligning commercial interests with safety interests.

does accountability matter beyond just compensation for AI harms?

✓ Correct — ✅ ✓ Accountability shapes incentives, not just compensation. Liability-free developers under-invest in safety. Users bearing all responsibility can't choose safely without information. Accountability design determines safety investment.

❌ ❌ Accountability matters for incentives: developers without liability have less incentive to invest in safety; users bearing all responsibility can't make informed choices. Accountability structures shape the safety incentive system.

challenge does distributed AI responsibility pose for existing liability frameworks?

✓ Correct — ✅ ✓ Traditional liability: designed for identifiable wrongdoers. Distributed AI responsibility across developers, deployers, users, and regulators creates attribution challenges existing frameworks weren't designed for.

❌ ❌ Traditional liability is designed for identifiable wrongdoers. Distributed AI responsibility across developers, deployers, users, and regulators doesn't fit existing frameworks — requiring new liability approaches.

Back to Lesson Start Lab 6

Lab 6

Liability Framework Design

Design a liability and accountability framework for AI harms.

Lab 6 — Liability Framework Design

Design a liability and accountability framework for AI harms.

The AI opens with the Tesla case and asks: how would you allocate liability across developers, deployers, users, and regulators for AI-related harms?
Develop your liability framework, distinguishing by risk level and harm type.
Address: what accountability mechanisms would create the right incentives for AI safety investment without preventing beneficial AI development?

Consider: strict liability vs. negligence standards, how liability allocation shapes safety incentives, and the challenge of attributing causation in distributed AI systems.

🔬 AI GuideLab 6

Back to Quiz Module Test

Module 4 Test

6 questions covering all lessons. Free, untracked, retake anytime.

Facebook content moderation example illustrates that:

✓ Correct — ✅ ✓ The same rule can produce ethically opposite outcomes in different contexts. Ethics requires judgment — not just consistent rule application.

❌ ❌ The same rule produced ethically opposite outcomes in different contexts. Ethics requires contextual judgment, not just rule consistency.

is defining harm in AI systems a political act?

✓ Correct — ✅ ✓ Harm definition is political: platforms benefit from narrow definitions; affected communities bear the costs of under-counting harm. Different interests, competing stakes.

❌ ❌ Harm definition is political: different stakeholders have competing interests. Platforms benefit from narrow definitions; affected communities bear the cost of those narrow definitions.

ProPublica and Northpointe were correct about COMPAS because:

✓ Correct — ✅ ✓ Both correct because they measured different fairness criteria. ProPublica: error rate parity violated. Northpointe: calibration satisfied. Both true at the same time.

❌ ❌ Both were mathematically correct — measuring different fairness criteria. Both findings were true simultaneously. This illustrates that 'fairness' is not one concept.

right to explanation for automated decisions means:

✓ Correct — ✅ ✓ Right to explanation: consequential automated decisions must be explainable to those affected. Opacity in high-stakes AI is a rights problem, not just a technical one.

❌ ❌ Right to explanation: when AI makes consequential decisions about a person, that person is entitled to understand why. This is a legal and ethical right, not just a technical preference.

Facebook emotion experiment was ethically contested because:

✓ Correct — ✅ ✓ Users didn't know their emotional states were being experimentally manipulated. Terms-of-service agreement is not informed consent to psychological experimentation.

❌ ❌ Users' emotional states were manipulated without their knowledge or genuine informed consent. Terms-of-service 'consent' didn't constitute informed consent to psychological experimentation.

does accountability matter beyond compensation for AI harms?

✓ Correct — ✅ ✓ Accountability shapes incentives: liability-bearing developers have financial reasons to invest in safety. Liability-free developers have less incentive. Accountability design determines the safety investment structure.

❌ ❌ Accountability shapes incentives: developers without liability under-invest in safety; users bearing all responsibility can't make informed choices. Accountability design determines the safety incentive system.

Back Academy Home