Something quietly unsettling happened in 2023. A man in New Orleans named Michael Williams spent nearly a year in jail — charged with murder based substantially on facial recognition evidence — before the case collapsed. The algorithm that flagged him was later found to have an error rate of up to 35% for darker-skinned faces. No one was prosecuted for that year of his life. No engineer was fired. No company was fined. The system, as they say, had simply made a mistake.

This is not a story about a rogue AI. It is a story about a structural vacuum at the heart of AI ethics — the absence of a coherent theory of punishment, accountability, and moral responsibility when the agent causing harm is not a person but a process.

We have spent enormous intellectual energy asking who should design AI systems ethically, what values they should encode, and how bias can be mitigated. These are important questions. But they all share a quietly optimistic assumption: that if we get the design right, the ethics will follow. What we have not seriously confronted is the harder, darker question — when AI causes harm despite our best intentions, who suffers the consequences? And if no one does, what does that mean for the entire project of AI ethics?

---

The Accountability Gap in AI Ethics: A Structural, Not Accidental, Problem

Let us begin with a historical observation. Every major technological revolution has eventually produced a corresponding legal and moral framework for assigning blame. The steam engine gave us industrial liability law. The automobile gave us traffic courts and insurance mandates. The pharmaceutical industry gave us the FDA and product liability torts. These frameworks were never perfect, but they shared a common architecture: a human agent, a traceable causal chain, and a mechanism for redress.

AI systems break all three of these pillars simultaneously.

Consider the causal chain problem. When a recidivism prediction algorithm — like COMPAS, which was controversially analyzed by ProPublica in 2016 — assigns a Black defendant a higher risk score than a white defendant with an identical criminal history, who caused that outcome? Was it the data scientists who trained the model? The prosecutors who chose to use it? The legislators who failed to regulate it? The company that sold it? Each actor can plausibly point to another, and the actual harm — a longer sentence, a denied parole — dissolves into a fog of distributed causation.

"We are at the very beginning of the time when machines will be able to store more information than any human, use it, and apply it better." — John von Neumann, The Computer and the Brain, 1958

Von Neumann was describing capability, not culpability. He could not have anticipated that the very sophistication he celebrated would become the primary obstacle to moral accountability.

This is what legal scholars are beginning to call the "responsibility gap" (Verantwortungslücke) — a term developed by philosopher Andreas Matthias to describe the structural impossibility of assigning blame when a machine acts in ways its designers did not specifically intend and could not fully predict. The gap is not a bug in our current system. It appears to be an inherent feature of how sufficiently complex AI systems operate.

---

Three Scenarios for How This Plays Out

Let us conduct a brief thought experiment. Imagine three near-future scenarios, each of which appears plausible given current trajectories.

Scenario One: The Liability Shell Game

In this scenario — which arguably describes the present — accountability continues to be diffused across corporate structures, regulatory bodies, and technical complexity until it effectively disappears. Companies deploy AI systems through layers of contractors and APIs. When harm occurs, legal teams argue that no single entity had sufficient control or foresight to be liable. Courts, lacking technical expertise, accept settlements that cost far less than the harm caused. The victims receive modest compensation, if any. The systems continue operating.

This is not speculation. It is, with some variation, what happened in the Uber self-driving car fatality in 2018, where Rafaela Vasquez — the safety operator — was charged, while Uber itself reached a settlement and continued its autonomous vehicle program. The corporation externalized the moral cost onto an individual worker.

Scenario Two: The Regulatory Capture of AI Ethics

In a second scenario, governments respond to public pressure by creating robust AI regulatory frameworks — but these frameworks are effectively written by the same technology companies they are meant to govern. The European Union's AI Act, for all its genuine ambition, has already shown signs of this dynamic: the final text was substantially softened on foundation model regulation following intense lobbying from major AI developers.

"The law is a system for the distribution of power. When power is sufficiently concentrated, it tends to write its own rules." — Roberto Mangabeira Unger, The Critical Legal Studies Movement, 1983

Under this scenario, AI ethics becomes a compliance theater — a series of audits, impact assessments, and ethics boards that create the appearance of accountability without its substance. Companies that are genuinely trying to do the right thing are burdened with paperwork. Companies that are not simply hire ethicists to manage their public image.

Scenario Three: The Emergence of Distributed Accountability

A third, more optimistic scenario involves the development of genuinely novel accountability architectures suited to the distributed nature of AI harm. These might include: mandatory algorithmic insurance pools (analogous to environmental liability funds), real-time audit trails embedded in AI systems at the infrastructure level, and participatory governance models that give affected communities standing to challenge algorithmic decisions before harm occurs rather than after.

This scenario is not utopian. Elements of it already exist in nascent form. The AI Tools Have a Governance Problem — And It's Getting Expensive analysis has documented how the cost of governance failures is beginning to create market incentives for genuine accountability structures. When AI ethics failures become expensive enough, corporations will eventually internalize them — though the question is whether that happens before or after irreversible social damage.

---

The Philosophical Dimension: Can a Process Be Guilty?

Here we must venture into territory that is genuinely philosophically unsettled. The entire Western tradition of moral responsibility — from Aristotle's concept of prohairesis (deliberate choice) to Kant's categorical imperative to contemporary consequentialism — assumes a moral agent capable of intention, reflection, and the experience of consequences.

AI systems, at their current stage of development, satisfy none of these conditions. They do not intend. They do not reflect in any philosophically meaningful sense. And crucially, they do not suffer. When we "punish" an AI system by taking it offline, we are not imposing a cost on the system — we are simply stopping a process. The punishment lands nowhere.

This creates what I would call the Moral Recipient Problem: for punishment and accountability to function as social mechanisms, there must be an entity capable of receiving and being changed by them. Without that entity, accountability becomes purely performative — a ritual we perform for the benefit of human observers rather than a genuine mechanism of correction.

P.F. Strawson's landmark 1962 essay "Freedom and Resentment" argued that our entire system of moral responsibility is grounded in what he called reactive attitudes — resentment, gratitude, indignation, love — that we direct toward other agents we perceive as genuinely capable of good will or ill will. We hold people responsible, Strawson argued, not because of some abstract metaphysical fact about free will, but because we are participants in a web of interpersonal relationships that requires these reactive attitudes to function.

AI systems, as they currently exist, stand outside this web. We cannot genuinely resent a language model. We cannot feel indignation toward an algorithm. And this means that our entire inherited vocabulary of moral accountability likely needs to be rebuilt from the ground up for the AI context — not merely extended or adapted.

---

What Accountability Without Punishment Actually Looks Like

Let me be direct about my own view here, offered with appropriate tentativeness: I believe the search for a single "responsible party" in AI harm cases is likely a category error — a misapplication of an individual-agent model to a systemic phenomenon.

The more productive framework may be what sociologist Charles Perrow called "normal accident theory" — the insight that in sufficiently complex, tightly coupled systems, catastrophic failures are not aberrations but statistical inevitabilities. Perrow developed this theory analyzing the Three Mile Island nuclear accident, but its application to AI systems appears remarkably apt.

If AI harms are, in part, normal accidents — emergent from system complexity rather than individual negligence — then the appropriate response is not punishment but systemic redesign. This means:

1. Pre-deployment harm modeling — not just bias testing, but adversarial scenario planning for the full distribution of possible failures, including low-probability, high-impact events.

2. Mandatory incident reporting — analogous to aviation's near-miss reporting system, which has made commercial flight extraordinarily safe precisely by treating every near-failure as a learning opportunity rather than a liability exposure.

3. Sunset clauses — automatic decommissioning requirements for AI systems operating in high-stakes domains unless they pass regular re-certification, shifting the burden of proof from harm to safety.

4. Affected community standing — legal mechanisms allowing communities disproportionately impacted by algorithmic systems to challenge their deployment before harm accumulates, rather than seeking redress afterward.

None of these are punishment in the traditional sense. But they are accountability in a structurally meaningful sense — they create genuine costs for negligence and genuine incentives for care.

---

The Deeper Question AI Ethics Must Confront

Marshall McLuhan famously argued that "the medium is the message" — that the form of a technology shapes social relations independently of its content. By extension, we might say that the accountability structure of a technology is its ethics. Not the principles written in its documentation, not the values its designers profess, but the actual mechanism by which harm is recognized, attributed, and corrected.

By that measure, AI ethics as currently practiced is less a genuine ethical system and more an elaborate set of aspirations — beautifully articulated, institutionally supported, and largely unenforceable. The AI Tools Have a Governance Problem — And It's Getting Expensive piece touches on precisely this tension: the gap between ethical aspiration and operational accountability is not merely philosophical — it has measurable economic and social costs.

The European Union's AI Act, for all its limitations, represents the most serious attempt to date to bridge this gap through legal architecture. According to the EU AI Act official documentation, high-risk AI systems in domains like biometric identification, critical infrastructure, and employment decisions are subject to mandatory conformity assessments, transparency requirements, and human oversight obligations. Whether this framework proves sufficient remains genuinely uncertain — but it represents a meaningful shift from voluntary ethics toward structural accountability.

What is still missing, even from the EU framework, is a coherent theory of what happens after harm occurs — not just how to prevent it, but how to recognize it, attribute it, and make it right. That is the intellectual work that AI ethics must now do.

---

A Question Worth Sitting With

Michael Williams spent 11 months in jail. The algorithm that contributed to his detention has likely been updated, retrained, or replaced. No one is in prison for that year of his life.

I do not raise this to generate outrage — outrage without structural analysis is merely emotion. I raise it because it illustrates, with brutal clarity, the distance between the ethics we profess and the accountability we practice.

As AI systems move deeper into criminal justice, healthcare, financial access, and employment — domains where errors are not inconveniences but life-altering — the question of what genuine accountability looks like becomes not a philosophical luxury but a practical urgency.

We have spent years asking what values AI should have. It is time to spend equal energy asking what consequences AI should face.

---

A question to consider: If we accept that AI systems will inevitably cause harm — not through malice but through complexity — what would a genuinely just accountability framework look like, and who would have the authority to enforce it?