Responsible AI
EMBA 8160 — Session 5
Core Principles
as established in the EU's Ethics Guidelines for Trustworthy AI
Beneficence
AI should actively benefit people and society — not merely avoid harm. The obligation is positive: to do good, not just to refrain from doing harm.Non-maleficence
AI must not cause harm, including through inaction, unintended side effects, or emergent behavior. This is the classical medical principle "do no harm" extended to autonomous systems.Autonomy
Respect for people's ability to understand, contest, and override AI decisions that affect them. This operationalizes as informed consent, explainability, and meaningful override mechanisms — the right not just to be informed but to actually exercise control.Justice
The benefits and burdens of AI should be distributed fairly across groups and geographies. This covers both procedural justice (fair process) and distributive justice (fair outcomes) — and directly motivates bias and fairness requirements in system design.Explicability
AI behavior must be understandable and auditable by the relevant stakeholders — at the appropriate level of abstraction for each audience. A patient, a clinician, and a regulator need different explanations of the same decision; all three are legitimate.The Regulatory Landscape
US State-Level AI Legislation
New York — Advanced AI Licensing Act (A 8195): Requires developers and operators of certain “high-risk advanced AI” to obtain a licence. Related bills, such as an amendment in the Illinois legislature (HB 1002), require state certification for the use of diagnostic algorithms.
Rhode Island (H 6286): Mandates a “distinctive watermark” to authenticate content generated by generative AI.
District of Columbia — Stop Discrimination by Algorithms Act (B 114): Requires notice about how AI is used in employment, including the source of information, and a separate notice to individuals affected by algorithmic eligibility determinations.
Massachusetts — Act Preventing a Dystopian Work Environment (HB 1873): Requires employers or vendors using AI to provide specific notices.
Washington (HB 1951): Requires developers of automated decision tools to disclose to deployers the types of data used to program or train the tool, and how the tool was evaluated for validity.
California (AB 331) and New York (AB 7859): Bills that would require AI deployers to allow individuals to opt out of decisions based on automated decision-making.
Bill status reflects the 2023–2024 legislative sessions and will have moved since — verify current status before relying on any single item.
Class Discussion
Discussion Question 1
Foundations of AI Ethics
An agentic AI system deployed in a hospital autonomously schedules and cancels patient appointments based on resource optimization. A patient misses a critical follow-up and suffers a worsening condition.
How do the five core ethical principles map onto this scenario, and which principle do you consider most violated?
Discussion Question 2
Bias & Fairness
Your team is building a loan-decision agentic system for a regional bank. You discover that historical lending data reflects decades of discriminatory lending practices.
You have three options: (A) train on the data as-is, (B) remove protected attributes from features, or (C) apply fairness-aware re-weighting. What is your recommendation and why? What fairness metric would you use to validate it?
Discussion Question 3
Accountability & Oversight
You are designing an agentic customer service system for a financial services firm. The system can autonomously resolve disputes, issue refunds up to $\$500$, and escalate to human agents.
A proposal is made to raise the autonomous refund threshold to $\$5,000$ to reduce human workload by 80%. What autonomy level would you recommend for this new threshold, and what safeguards would you require before approval?
Discussion Question 4
Governance
Your organization wants to deploy an agentic AI system that monitors employee communications (Slack, email, documents) to detect IP theft and policy violations. The system will flag suspicious behavior to HR.
What governance structures, consent mechanisms, and technical safeguards would you require before approving this deployment, and are there conditions under which you would refuse to build it?
Discussion Question 5
Synthesis
You have been hired as the AI Ethics Lead at a startup deploying a multi-agent system to autonomously manage a portfolio of real estate investments — identifying properties, negotiating purchase terms, managing tenants, and initiating legal proceedings when necessary.
Using any frameworks discussed today, identify the three most significant ethical risks and design one concrete mitigation for each. Would you take this job?
