When AI Doesn’t Know What to Do: Designing Fallible Ethics for Intelligent Machines

Redacción HC
28/07/2023

As artificial intelligence (AI) systems take on more critical roles—from driving cars to diagnosing illnesses—they’re increasingly expected to make ethical decisions. But here lies a paradox: how can we program machines with moral principles when human ethics are deeply fragmented, culturally contingent, and sometimes even contradictory?

In their provocative article “Éticas falibles para máquinas (in)falibles” (Fallible Ethics for (Un)fallible Machines), philosophers Jordi Vallverdú and Sarah Boix of the Universitat Autònoma de Barcelona challenge the assumption that universal moral codes can—or should—guide AI behavior. Instead, they propose a model that embraces uncertainty, acknowledging that sometimes, machines might need to fail ethically in a controlled, transparent, and probabilistic way.

Published in the peer-reviewed journal Arbor (2021), this conceptual study offers a philosophical and normative roadmap for navigating the moral ambiguity that accompanies real-world decision-making in AI systems.

Why a Single Ethical Code for AI Is a Myth

A growing body of evidence—from the MIT Moral Machine experiment to public debates over autonomous weapons—demonstrates that moral choices are far from universal. What is considered “just” in one culture may be unacceptable in another. The study by Vallverdú and Boix begins by confronting this reality: there is no one-size-fits-all ethical framework suitable for global deployment in AI.

For instance, when asked to prioritize lives in a hypothetical crash scenario, respondents from different countries offered wildly divergent answers. What should an AI do when even human consensus is out of reach?

The Proposal: Embracing Ethical Fallibility in AI

Instead of attempting to force moral certainty into systems inherently constrained by limited data, time, and understanding, the authors argue for designing “fallible” ethics in machines. Here are the core ideas:

1. Ambiguity Is Inevitable

Ethical principles frequently clash. A system may be asked to maximize lives saved (utilitarianism) while also respecting equal rights (deontology). In many real-time scenarios, there’s no “correct” answer—only ethical dilemmas.

2. Machines Must Recognize Their Own Limits

AI should be designed to acknowledge when it lacks sufficient information, time, or clarity. The authors emphasize that ethical uncertainty is not an error—it’s a feature of moral life.

3. Randomization as Ethical Strategy

Here’s where the article breaks new ground: it proposes incorporating “regulated ethical randomness” into AI decision-making. When a machine encounters an irresolvable moral conflict, it may choose to invoke a controlled element of chance, selecting between competing values according to predefined probabilities.

This isn’t arbitrariness. Rather, it’s a mechanism of moral transparency: when no decision is ethically superior, the system clearly signals its fallback to probabilistic reasoning.

Contextual and Transparent Decision-Making

Beyond randomness, the authors advocate for adaptive, context-sensitive ethics. AI systems should be calibrated to operate within specific cultural, legal, and institutional environments rather than being governed by abstract universal codes.

Furthermore, machines must explain their reasoning. Transparency isn’t optional. If ethical randomization is invoked, the system should log the event, note the competing principles, and communicate this to users or oversight bodies.

“Machines must act with procedural clarity—acknowledging ambiguity, documenting choice, and justifying the invocation of probabilistic ethics.” (Vallverdú & Boix, 2021)

Practical Implications: Designing Ethically Resilient AI

While the article is theoretical, its implications reach deep into AI design and governance:

For Engineers:

• Build mechanisms that detect ethical conflict and trigger fallback protocols like weighted randomization.
• Develop explainable AI architectures that log and justify decisions in real-time.

For Policymakers:

• Recognize the limits of top-down ethical codes; legislate frameworks that allow for moral flexibility in AI.
• Require developers to report when and how ethical randomness is used.

For Society:

• Promote public education on how AI handles moral ambiguity to foster trust and informed discourse.
• Encourage interdisciplinary collaboration—between philosophers, developers, legislators, and civil society—to create socially legitimate AI systems.

Why This Matters in the Global South

The insights from this study are particularly relevant to regions like Latin America, where ethical norms and regulatory environments differ significantly across contexts. In sectors like digital health, justice, and urban mobility, AI systems must adapt to local values. Implementing fallible, context-aware ethics could offer a practical path forward in settings where standardized Western models of AI ethics fall short.

Conclusion: Toward an Honest AI That Knows Its Limits

The ambition to create “perfectly ethical” machines may be misplaced. What we need instead are honest machines—systems that know when they can’t decide perfectly, that admit their fallibility, and that act with openness, fairness, and procedural clarity.

“Fallible ethics” doesn’t mean giving up on responsibility. It means creating more realistic, context-sensitive, and trustworthy AI—and acknowledging that sometimes, even the smartest systems must flip a moral coin.

Topics of interest

Technology Academia