AI Ethics Frameworks

This course is for anyone who designs, evaluates, or is affected by systems that make decisions on people's behalf — product managers shaping content moderation policy, engineers building autonomous systems, policymakers writing AI governance rules, or anyone who wants a sharper vocabulary than “that seems wrong” for evaluating an algorithm's choices.

No philosophy background required. Each framework is explained from its original source, in plain language, before being applied to concrete AI examples.

Utilitarianism, most associated with Jeremy Bentham and John Stuart Mill, is a form of consequentialism: an action is right if it produces the best overall outcome — typically framed as maximizing well-being and minimizing suffering, summed across everyone affected. Nothing is intrinsically forbidden; the ends genuinely can justify the means, if the math works out.

In AI systems, utilitarian logic shows up constantly and usually invisibly: a content moderation system tuned to “minimize total harm across the user base” is utilitarian. A resource allocation algorithm optimizing aggregate outcomes (most lives saved, most requests served) is utilitarian. The appeal is that it's measurable and optimizable — exactly what makes it natural for engineers to reach for by default, often without recognizing it as one ethical position among several rather than simply “the objective approach.”

Immanuel Kant's moral philosophy takes the opposite approach: some actions are wrong regardless of their consequences, because they violate a duty or a rule that should hold universally. Kant's famous test — the categorical imperative — asks whether you could will the rule behind your action to become a universal law everyone follows. Lying fails this test even when lying would produce a better outcome, because a world where everyone lies whenever it's convenient undermines the very possibility of trust.

In AI systems, deontological logic shows up as hard constraints that hold no matter what the optimizer “wants” — a content policy that never allows certain content regardless of engagement benefits, a safety system with rules an autonomous vehicle will never violate (don't intentionally target a pedestrian) even if simulation shows a rule violation would minimize total harm in some edge case. These are exactly the guardrails that keep pure outcome optimization from producing outcomes nobody actually wants.

Tracing back to Aristotle, virtue ethics asks a different question entirely: not “what outcome maximizes good?” or “what rule applies?”, but “what would a person of genuinely good character — honest, courageous, fair, wise — do here?” Right action flows from good character rather than from calculation or rule-following, and good character is cultivated through practice and habituation, not derived from a formula.

This is the hardest framework to encode directly into an AI system, precisely because it resists reduction to a formula or a rulebook — which is exactly why it maps unusually well onto how modern AI systems are actually trained. A model fine-tuned on examples of how a careful, fair, thoughtful human moderator or advisor handled thousands of specific, messy real situations is closer to learning “character by imitation” than to running a utility calculation or checking a rule list — even though the underlying mechanism is statistical pattern-matching, not virtue in the Aristotelian sense.

In 2018, a team led by Edmond Awad published “The Moral Machine Experiment” in Nature — an online platform presenting trolley-problem-style dilemmas for self-driving cars (spare passengers or pedestrians? more lives or fewer? young or old?) that collected roughly 40 million decisions from people in 233 countries, the largest empirical study of applied moral intuition ever conducted.

The findings map directly onto this course's three frameworks: most respondents leaned utilitarian in aggregate (favoring fewer deaths over more), but layered in strong, near-universal deontological exceptions (a strong reluctance to treat any decision as intentionally targeting a person) and showed clear virtue-adjacent cultural variation in what counted as a “fair” or “honorable” choice. Crucially, the study found real, systematic cross-cultural clusters — there was no single global consensus.

Every AI system that ranks, filters, moderates, allocates, or acts already encodes an ethical position — usually without anyone stating it out loud. Use this template to name it explicitly the next time you evaluate, design, or are affected by one.

I'm evaluating the ethical logic behind this AI system or decision: [describe the system, e.g. "a content moderation algorithm" or "an autonomous vehicle's collision-avoidance behavior"]

Walk me through it against these three lenses:
1. Utilitarian lens: What outcome is this system apparently optimizing to maximize or minimize? Whose well-being counts, and is anyone's excluded from the calculation?
2. Deontological lens: What hard rules or limits does this system never violate, regardless of the outcome? Are those limits stated explicitly, or only implied by what it refuses to do?
3. Virtue-ethics lens: Was this system trained by imitating examples of good human judgment in context, rather than by rule or by optimization? If so, whose judgment, and in what contexts, was it trained on?
4. Where these three lenses would give DIFFERENT answers for a real decision this system might face, which one appears to actually govern, and is that the right choice for this context?

Give me a one-paragraph assessment of which framework(s) this system is actually built on, and one concrete scenario where that choice would visibly matter.