Something has gone quietly wrong at the center of AI ethics discourse, and it has less to do with algorithms than with grammar. Specifically, with the passive voice. "Harm was caused." "Bias was introduced." "Mistakes were made." The language of AI ethics has become extraordinarily good at describing damage while systematically erasing the subject of the sentence — the one who acted, decided, and profited. This is not accidental. It is, I would argue, the defining structural crisis of AI ethics today: not that accountability is absent, but that it has been architecturally designed out of the system.

The question I want to press here is not "who is responsible?" — that question has been asked, loudly, for years. The more unsettling question is: what happens to a society when the answer to that question becomes genuinely, structurally unanswerable?

The Accountability Gap Is Not a Bug — It's a Feature of How AI Ethics Was Built

Let me begin with a historical parallel. When the Ford Motor Company produced the Pinto in the 1970s, internal memos revealed that engineers had calculated the cost of fixing a known fuel tank defect against projected lawsuit payouts — and chosen the cheaper option. The scandal was not merely that people died. It was that a specific decision, made by identifiable people, in a traceable meeting, caused those deaths. That traceability was what made accountability possible.

Now consider a modern AI system — say, a predictive risk-scoring tool used in bail decisions, or an automated hiring filter deployed across thousands of companies. When such a system produces discriminatory outcomes, the causal chain looks entirely different. The training data was assembled by one team. The model architecture was designed by another. The deployment decision was made by a third party. The calibration thresholds were set by a fourth. The procurement contract was signed by a fifth. And the individuals harmed — who were they, exactly, in a system that processed them as statistical abstractions?

"Responsibility for AI systems is distributed across a complex sociotechnical network in ways that make it difficult to assign blame to any single actor." — Mittelstadt et al., The Ethics of Algorithms, Big Data & Society, 2016

This diffusion of agency is not a temporary problem awaiting a technical fix. It appears to be a constitutive feature of how large-scale AI systems are built and deployed. And the ethical frameworks we have developed — fairness metrics, transparency requirements, impact assessments — were largely designed to evaluate outputs, not to reconstruct the moral geography of decisions.

When "Ethical AI" Becomes a Liability Shield

Here is a thought experiment worth sitting with. Suppose a company publishes a thorough AI ethics report. It documents bias audits, explains its fairness criteria, describes its human oversight mechanisms. The report is earnest, detailed, and — in a narrow technical sense — accurate. Then the system causes harm to a specific community.

What has the ethics report accomplished? From the perspective of the affected community: very little. From the perspective of the company's legal team: quite a lot. The existence of documented ethical process has, in practice, become a powerful instrument for deflecting accountability rather than enabling it. "We followed our ethics framework" is increasingly functioning as the corporate equivalent of "I was just following orders" — a claim that distributes moral responsibility so broadly that it effectively dissolves it.

This is what I mean when I say the accountability vacuum is structural. The problem is not that companies are acting in bad faith (though some are). The problem is that even good-faith ethical compliance has been designed in a way that severs the connection between harm and remedy.

a close up of a typewriter with a paper on it

Photo by Markus Winkler on Unsplash

The Three Structural Breaks in AI Accountability

Let me be more precise. I see three distinct places where the accountability chain breaks:

1. The Causal Break. In traditional tort law, accountability requires establishing that a specific act caused a specific harm. AI systems introduce what philosophers call "problem of many hands" — so many agents contributed to the outcome that no single causal line can be drawn. When a credit-scoring algorithm denies someone a mortgage, who caused that denial? The data labelers? The model trainers? The bank that deployed it? The regulator who approved it?

2. The Temporal Break. AI systems are trained on historical data but applied to future situations. The harm often emerges months or years after the decisions that produced it. By the time the harm is visible, the responsible parties may have changed roles, the company may have been acquired, and the original design choices may be buried under layers of updates. As I have argued in my previous work on the speed problem in AI ethics, we are structurally positioned to understand harms only after they have accumulated sufficient scale to become visible.

3. The Epistemic Break. Perhaps most troubling: many AI harms are statistical in nature. A system that is 15% more likely to flag Black defendants as high-risk does not harm any specific person in a legally traceable way. It harms a population, probabilistically, across thousands of decisions. Our accountability frameworks — legal, moral, institutional — were designed for individual harms caused by individual acts. They are largely unprepared for distributed, probabilistic, population-level damage.

The Moral Luck Problem in AI Systems

The philosopher Thomas Nagel introduced the concept of moral luck — the unsettling observation that we hold people responsible for outcomes that were, to a significant degree, outside their control. The drunk driver who makes it home safely is not prosecuted; the one who hits a pedestrian is. Same decision, different outcome, radically different moral and legal treatment.

AI systems amplify moral luck to a disturbing degree. Whether a particular deployment causes visible harm depends enormously on contingent factors: which dataset happened to be used, which edge cases happened to appear in the test set, which community happened to be the first to encounter the system at scale. Two companies can make nearly identical design choices, and one escapes scrutiny while the other faces congressional hearings — not because of meaningfully different ethical conduct, but because of the accident of which harms became legible.

This creates a perverse incentive structure. If accountability is triggered by visibility of harm rather than quality of decision-making, then the rational strategy is not to make better decisions — it is to ensure that harms, when they occur, remain invisible, dispersed, or attributable to someone else. And that, I would suggest, is precisely the incentive structure that current AI ethics frameworks have inadvertently created.

What Accountability Without a Subject Looks Like in Practice

Consider the recent wave of AI-generated content moderation systems. A platform deploys an automated system to flag harmful content. The system disproportionately removes content in minority languages and dialects — not because of explicit discrimination, but because the training data underrepresented those communities. Whose responsibility is this?

The platform's legal team points to the vendor. The vendor points to the open-source model they fine-tuned. The open-source community points to the original dataset curators. The dataset curators point to the internet, which is what it is. Meanwhile, the communities affected have no clear party to petition, no clear legal theory of harm, and no clear mechanism for remedy.

This is not hypothetical. Variations of this pattern have appeared in content moderation, healthcare triage, child welfare screening, and — as explored in a related analysis on AI tools making cloud recovery decisions without approval — even in critical infrastructure management. In each case, the absence of a clear accountable subject is not an oversight. It is the predictable outcome of building systems whose complexity exceeds the resolution of our accountability frameworks.

AI Ethics Needs a Theory of Structural Responsibility

Here is where I want to offer something beyond diagnosis. The accountability vacuum will not be filled by better ethics guidelines, more transparency reports, or stricter impact assessments — at least not alone. What is needed is a fundamental reconceptualization of where moral responsibility lives in complex sociotechnical systems.

Several directions appear promising:

1. Prospective Liability Structures

Rather than waiting for harm to occur and then asking who caused it, we might require that anyone who profits from an AI system bears presumptive liability for its harms — unless they can demonstrate that they took specific, documented steps to prevent the harm. This inverts the current structure, in which the burden of proof falls on the harmed party to establish causation. The EU AI Act's risk-based classification system moves in this direction, though critics argue it still leaves too much room for compliance theater.

"High-risk AI systems should be subject to a conformity assessment before they are put into service or placed on the market." — EU Artificial Intelligence Act, Article 43

2. Collective Accountability Mechanisms

Some harms are genuinely collective in origin and may require collective accountability mechanisms. Industry-wide compensation funds — analogous to the Vaccine Injury Compensation Program in the United States — could provide remedy for AI harms without requiring proof of individual causation. This would not eliminate the need for accountability, but it would decouple remedy from blame attribution in cases where the causal chain is genuinely underdetermined.

3. Participatory Governance as Accountability Infrastructure

The communities most likely to be harmed by AI systems are currently the least involved in designing the accountability frameworks that govern those systems. This is not merely an ethical failing — it is an epistemic one. Those with the most direct experience of a system's failure modes are the most valuable source of information about where accountability should be located. Participatory design processes, community oversight boards, and mandatory impact consultations with affected groups are not soft add-ons to "real" accountability — they may be its most important structural component.

This connects to a broader point about governance design. Just as automation systems can fail when they're built for the wrong kind of user, AI accountability frameworks can fail when they're built around the wrong theory of who holds power and who bears risk.

The Philosophical Stakes

Marshall McLuhan famously observed that "the medium is the message" — that the form of a technology shapes social relations independently of its content. I would suggest something analogous is true of accountability frameworks: the structure of responsibility is itself a moral message. When we design systems in which no one can be held responsible, we are not merely making an administrative error. We are communicating something about whose suffering counts, whose claims are legible, and whose voice has standing in the moral order.

The accountability vacuum in AI ethics is, at its deepest level, a political question about who has the power to make binding decisions about shared life — and who bears the costs when those decisions go wrong. The fact that this question is currently being answered by a small number of technologists, lawyers, and ethicists, largely without democratic input from those most affected, is not a technical limitation awaiting a better algorithm. It is a choice. And choices, unlike algorithms, can be changed.

A Question Worth Sitting With

I want to close not with a summary but with a provocation, in the spirit of good philosophical inquiry.

We have spent considerable effort asking how to make AI systems more accountable. But here is the question I find myself returning to: What if the demand for accountability is itself a kind of moral displacement — a way of managing our anxiety about complex systems without confronting the more uncomfortable question of whether some of these systems should exist in their current form at all?

Accountability, after all, is a remedial concept. It kicks in after something has gone wrong. The deeper question — whether the design, deployment, and profit structures of AI systems are compatible with democratic governance and human dignity — requires a different vocabulary entirely. One we are only beginning to develop.

If you found this analysis useful, the external research of Mittelstadt et al. on algorithmic ethics provides a rigorous empirical foundation for many of the structural arguments raised here. The EU AI Act documentation, available through the EU's official legislative portal, offers the most detailed current attempt to translate these concerns into binding governance — with all the compromises that legislative process entails.

Tags: AI ethics, accountability, technology philosophy, governance, structural responsibility, moral luck, AI regulation

The Accountability Trap: Why Fixing AI Ethics Requires More Than Someone to Blame

(continued)

A Question Worth Sitting With — And Then Acting On

The provocation I raised in the previous section deserves more than rhetorical weight. Let me press it further.

When we demand accountability from AI systems — from their designers, deployers, and auditors — we are, in a sense, already conceding the terms of the debate. We are accepting that the system exists, that it operates at scale, and that our task is to assign blame cleanly when it causes harm. This is the logic of the insurance adjuster, not the philosopher. It is the logic of damage control, not democratic deliberation.

Hannah Arendt, writing about bureaucratic power in The Origins of Totalitarianism, observed that one of the most insidious features of modern administrative systems is their capacity to distribute responsibility so thoroughly that no one, in the end, is responsible. She called this "rule by nobody" — not because no one is in charge, but because the architecture of the system makes individual responsibility structurally incoherent. I would argue that contemporary AI governance is reproducing this dynamic with remarkable fidelity, and doing so at a speed that Arendt could not have imagined.

The question, then, is not merely who is accountable. It is whether the concept of accountability — as currently framed — is adequate to the phenomenon we are trying to govern.

Three Scenarios: Where the Accountability Discourse Goes From Here

Let me offer what I think are the three most plausible trajectories for how this conversation evolves over the next decade. I present them not as predictions but as structured possibilities — a futurist's way of mapping the terrain.

Scenario One: Accountability as Compliance Theater

In this trajectory, the dominant response to AI harm becomes increasingly procedural. Audit requirements multiply. Ethics review boards proliferate. Impact assessments become mandatory. Companies hire Chief AI Ethics Officers. Regulatory frameworks like the EU AI Act are extended, amended, and cited in litigation.

And yet — the structural conditions that produce harm remain largely intact. The incentive to deploy fast, to scale aggressively, and to externalize costs onto users and communities does not change. Accountability becomes, in effect, a performance of responsibility rather than its substance. This is not a cynical prediction; it is already, in significant measure, the present. As Ruha Benjamin has observed, the language of ethics can function as a "race to innocence" — a way of appearing responsible without bearing responsibility.

The danger here is not that accountability fails dramatically. It is that it succeeds just enough to foreclose more fundamental reform.

Scenario Two: Accountability as Democratic Infrastructure

A more hopeful trajectory imagines accountability mechanisms evolving into genuine instruments of democratic governance. In this scenario, the key shift is not technical but institutional: the people most affected by AI systems gain meaningful standing to challenge, contest, and reshape them — not merely to receive explanations after the fact.

This would require, at minimum, what legal theorists sometimes call "participatory parity" (Nancy Fraser's term) — the structural conditions under which affected communities can participate as peers in the processes that govern them. It would mean rethinking who sits on ethics boards, who funds AI audits, who has legal standing to bring claims, and whose epistemological frameworks count as legitimate evidence of harm.

This scenario is not utopian. Partial versions of it already exist in data protection law, in some environmental justice frameworks, and in emerging algorithmic impact assessment practices. But scaling it requires political will that, as of April 2026, remains unevenly distributed across jurisdictions.

Scenario Three: Accountability Displaced by a Different Question

The third trajectory is perhaps the most philosophically interesting, and the most uncomfortable. It imagines a future in which the accountability discourse is not reformed but transcended — in which we collectively decide that certain questions about AI cannot be answered within the accountability framework at all, and that we need a different vocabulary.

What might that vocabulary look like? I think it begins with what philosophers of technology like Albert Borgmann called "focal practices" — the question of what kinds of human activity and human relationship we want to preserve and cultivate, and what role, if any, AI systems should play in supporting rather than supplanting them. It moves from "who is responsible when this goes wrong?" to "what kind of world are we building, and for whom?"

This is not a retreat from governance. It is a demand for governance at a deeper level — one that engages not just the outputs of AI systems but the values embedded in their design, the economic structures that incentivize their deployment, and the political processes through which their scope and limits are determined.

The Structural Silence We Need to Name

There is one more thing I want to say before closing, and it is perhaps the most uncomfortable observation in this entire analysis.

The accountability discourse — for all its genuine importance — has a structural blind spot. It tends to focus on discrete, identifiable harms: a biased hiring algorithm, a wrongful facial recognition match, a manipulative recommendation system. These are real harms, and they deserve redress. But they are also, in a sense, the visible tip of a much larger structure.

The deeper question — one that accountability frameworks are not designed to address — is about what I would call systemic normalization: the process by which AI systems reshape the baseline conditions of social life in ways that are diffuse, cumulative, and largely invisible until they have already become the new normal.

Marshall McLuhan famously argued that "the medium is the message" — that the form of a communication technology shapes society more profoundly than any particular content it carries. Applied to AI, this suggests that the most consequential effects of these systems may not be the harms we can trace to specific decisions, but the ways in which AI-mediated environments gradually restructure human attention, human judgment, and human relationships in ways that no single actor designed or intended.

Accountability, as a concept, requires a traceable causal chain: harm → decision → agent → responsibility. Systemic normalization breaks that chain at every link. There is no single decision. There is no identifiable agent. There is no moment at which the harm becomes visible enough to trigger a claim.

This is not an argument for fatalism. It is an argument for epistemic humility — and for developing governance frameworks that can operate at the level of systemic effects, not just discrete incidents.

My Considered View — Offered with Appropriate Tentativeness

I have tried, throughout this analysis, to present the arguments fairly and to resist the temptation of easy conclusions. Let me now, as I always do at the end of these essays, offer my own tentative view.

I believe the accountability discourse in AI ethics is both necessary and insufficient. It is necessary because the alternative — no accountability — is clearly worse, and because the procedural frameworks being developed in places like Brussels, Seoul, and Washington represent genuine, hard-won progress. I do not want to dismiss that progress with philosophical impatience.

But I also believe that accountability, as currently framed, is functioning partly as what the psychologist Robert Kegan might call a "holding environment" — a structure that contains our anxiety about AI's social consequences without resolving the deeper questions those consequences raise. It gives us something to do, someone to blame, and a vocabulary that feels adequate to the problem. And in doing so, it may be making it harder to ask the questions that actually need asking.

The question I keep returning to — and that I think our field needs to sit with more honestly — is this: Are we building accountability frameworks for the AI systems we have, or for the AI systems we are willing to imagine having? Because those are very different projects. And conflating them may be the most consequential mistake of this moment.

A Question to Take With You

I want to leave you, as I always do, with a single question — not to be answered quickly, but to be carried.

If the most significant harms caused by AI systems are not discrete incidents but cumulative transformations of social life that no single actor designed or intended — what would it mean to hold a society, rather than a company or an algorithm, accountable for them? And are we prepared to accept what that answer might require of us?

This essay is part of an ongoing series on the structural limits of AI ethics discourse. Previous installments have examined the mirror problem, the consent problem, the speed problem, and the language of accountability. Each essay attempts to press a single structural question further than the standard governance literature typically allows.

For those wishing to engage with the primary scholarship: Ruha Benjamin's Race After Technology (2019), Albert Borgmann's Technology and the Character of Contemporary Life (1984), and Nancy Fraser's work on participatory parity in Adding Insult to Injury (2008) provide the theoretical anchors for many of the arguments developed here. Hannah Arendt's The Origins of Totalitarianism (1951) remains, I think, the most important unread book in contemporary AI governance.

Tags: AI ethics, accountability, technology philosophy, governance, structural responsibility, systemic normalization, democratic infrastructure, Marshall McLuhan, Hannah Arendt

NOCODE TECH STACKER

The Accountability Vacuum: Why AI Ethics Has No One Left to Blame