Module 913 min read · AI and Society

AI Ethics and Human Values

Ethics is not a checkbox AI systems pass or fail. It is an ongoing process of identifying whose interests are at stake, what values are in tension, and how to make decisions under genuine uncertainty about consequences. The field of AI ethics has grown rapidly in response to real harms from deployed systems — but it has also attracted criticism for providing ethical cover without requiring real accountability. This module examines what AI ethics is, what it struggles with, and what it actually takes to build AI systems that respect human values.

The core ethical principles and their tensions

Several frameworks have emerged as anchors for AI ethics discussions. The most widely cited include fairness, accountability, transparency, and safety — often abbreviated as FATE. The EU's approach adds human agency and oversight, technical robustness, privacy, and non-discrimination. Virtually every major AI ethics statement emphasizes some combination of these principles.

The challenge is that these principles are genuinely in tension with each other and with commercial and operational realities. Transparency may compromise safety (adversaries can exploit known vulnerabilities). Fairness across different groups may require different treatment, which itself raises equity questions. Explainability and accuracy are often in inverse relationship — the most accurate models are often the least interpretable. Ethics involves navigating these tensions rather than resolving them once and for all.

Fairness is not singular

Mathematical definitions of fairness are not compatible with each other. Equal accuracy across groups (demographic parity) conflicts with equal error rates across groups (equalized odds), which conflicts with predictive accuracy for individuals. A system can satisfy one definition of fairness while violating another. Choosing among them is a values decision that cannot be made purely technically — it requires deciding whose interests to prioritize and what kind of errors matter most.

The value alignment problem

At its deepest level, AI ethics is concerned with the value alignment problem: ensuring that AI systems pursue goals that are genuinely aligned with human values rather than proxy measures that diverge from what we actually care about. A social media algorithm optimizing for "engagement" aligns with a narrow proxy metric and diverges from the actual human interests users have — in accurate information, meaningful connection, and wellbeing.

The alignment problem becomes more acute as AI systems become more capable. A narrow AI that maximizes click-through rates can cause social harm. A highly capable AI system pursuing subtly misspecified objectives could cause catastrophic harm. The field of AI safety is concerned with these longer-term alignment challenges, while near-term AI ethics focuses on the harms from systems deployed today.

From principles to practice

Ethics washing
Critics have identified "ethics washing" — using AI ethics principles and statements to maintain reputational cover while continuing harmful practices. When companies publish responsible AI principles while continuing to sell facial recognition to law enforcement, or conduct ethics reviews that never delay deployment, ethics serves public relations rather than harm prevention. Genuine ethics requires that principles have consequences for what gets built.
Structural over individual ethics
Much AI ethics discourse focuses on individual responsibility — the developer who makes an ethical choice, the company that establishes review processes. But the most significant AI harms emerge from structural factors: market incentives that reward deployment speed over safety, regulatory gaps that allow harmful systems, and power imbalances that prevent affected communities from having meaningful input. Individual ethics operates within structural constraints that must also be addressed.
Participatory approaches
Effective AI ethics includes the communities that AI systems affect as genuine participants in design and evaluation, not just as subjects of ethical consideration. Participatory design processes that involve marginalized communities in AI development — not as consultants but as decision-makers — produce systems with fewer harmful surprises and greater legitimacy.

Specific ethical challenges that matter now

  • Predictive systems and self-fulfilling prophecies: When risk scores predict criminal recidivism, housing instability, or child welfare risk, deploying those predictions can create the conditions they predict — denying resources to people flagged as high-risk reinforces the patterns that produced the score
  • Consent and data: The data underlying AI systems was collected in contexts where users did not meaningfully consent to its use in training AI — or could not have anticipated that use. Retroactive consent frameworks are inadequate
  • Autonomy and manipulation: Personalized AI systems that learn to exploit psychological vulnerabilities to achieve engagement undermine the autonomy that ethical frameworks are designed to protect
  • Responsibility gaps: When AI systems cause harm, attribution is diffuse — developers, deployers, users, regulators all bear some responsibility, and this diffusion can enable everyone to point elsewhere
The limits of technical solutions

Many AI ethics problems are framed as technical problems with technical solutions — better fairness metrics, more explainable models, more diverse training data. These improvements matter, but they don't resolve the underlying question of whether a system should be built and deployed at all. Some applications of AI — mass surveillance, autonomous lethal weapons, psychological manipulation for engagement — raise ethical objections that better technical implementation cannot address.

The practice of AI ethics

Genuine AI ethics is a practice, not a checklist. It involves asking hard questions before deployment: What could go wrong? Who is most vulnerable? What are the reversibility options if harm emerges? It involves building structures that make those questions matter — review processes with real authority to delay or stop deployment, external audits with genuine access, affected community input with actual influence. And it involves ongoing monitoring after deployment, because systems change and their contexts change in ways that produce new harms over time.