Who Is Responsible When AI Gets It Wrong?

By Dyuti Khulbe/Public Praxis Forum

Artificial intelligence has quickly become part of everyday life, even for people who rarely think about technology. It recommends the videos we watch, filters job applications, helps doctors interpret medical images, detects fraudulent financial transactions, translates languages, and increasingly assists governments in delivering public services. Much of this happens quietly, often without us even realising that an algorithm has influenced a decision that affects our lives.

For all its sophistication, however, artificial intelligence has a habit of reminding us that it is far from infallible. An autonomous vehicle is involved in an accident. A facial recognition system wrongly identifies an innocent person. An AI recruitment tool systematically disadvantages women applicants. A chatbot confidently provides dangerous medical advice. Each incident sparks public debate, attracts media attention, and prompts investigations. Yet once the immediate questions about what happened begin to settle, another question inevitably takes centre stage.

Who is responsible?

At first glance, this appears to be a legal question. Someone made a mistake, someone suffered harm, and someone should be held accountable. But artificial intelligence complicates a relationship that societies have long taken for granted. When a human decision goes wrong, responsibility can usually be traced to an individual or an institution. AI systems, by contrast, are rarely the product of a single actor. They are designed by engineers, trained on vast datasets, deployed by organisations, purchased by governments or businesses, and ultimately used by people exercising different degrees of judgment. Responsibility, therefore, does not disappear. It becomes distributed.

That is what makes artificial intelligence so interesting from the perspective of governance. Beneath the discussions about algorithms and machine learning lies a much older question about how societies organise responsibility when decisions are no longer made in familiar ways. In many respects, the challenge posed by AI is not simply technological. It is institutional.

A Very Old Question in a Very New Context

Although artificial intelligence feels like one of the defining technologies of the twenty-first century, the questions it raises are surprisingly familiar. Throughout history, societies have repeatedly had to rethink responsibility whenever new technologies altered the way people worked, communicated, or exercised power. The printing press transformed who could produce knowledge. Industrial machinery reshaped labour and working conditions. The internet changed how information travelled across the world. Each innovation created opportunities that previous generations could scarcely imagine, but each also forced societies to reconsider where responsibility should lie when something went wrong.

Artificial intelligence belongs to that same tradition, although its influence reaches further. Unlike earlier technologies, AI is increasingly involved not only in performing tasks but also in shaping decisions. It can recommend prison sentences, assist doctors with diagnoses, identify welfare fraud, prioritise emergency calls, and help employers decide which applicants deserve an interview. In many cases, humans remain involved. Yet the role of the machine is no longer simply mechanical. It has become advisory, predictive, and, in some situations, deeply influential.

This distinction matters because decision-making has always carried moral and political weight. Deciding who receives a loan, who is hired for a job, or who qualifies for public assistance is rarely a purely technical exercise. These decisions affect people’s opportunities, livelihoods, and rights. Once algorithms begin participating in these processes, questions about technology quickly become questions about fairness, accountability, and public trust.

The computer scientist Joseph Weizenbaum, writing decades before today’s AI boom, warned against allowing computers to make decisions that require human judgment. His concern was not that machines would become evil, but that societies might gradually surrender moral responsibility by treating technical systems as though they were capable of making ethical choices. Some decisions, he argued, demand human accountability precisely because they involve values rather than calculations.

More than forty years later, that observation feels remarkably contemporary.

Why We Wanted Machines to Decide

One reason artificial intelligence has been embraced so enthusiastically is that it appears to offer something institutions have pursued for centuries: more objective decision-making. Human beings are inconsistent. We become tired, overlook information, carry unconscious biases, and sometimes make decisions influenced by emotion, habit, or prejudice. Organisations have therefore long searched for ways to make decision-making more predictable, standardised, and evidence-based.

Artificial intelligence seems to promise exactly that. Unlike people, algorithms can analyse enormous quantities of data within seconds. They apply the same rules repeatedly and do not become distracted or fatigued. It is easy to understand why governments, hospitals, banks, and businesses have become interested in systems that appear capable of producing faster and more consistent decisions than humans alone.

Yet this promise rests on an assumption that deserves closer examination. It assumes that objectivity is something technology can simply deliver.

As the science and technology scholar Sheila Jasanoff has argued, knowledge is never produced independently of the societies that create it. Scientific and technological systems do not emerge outside politics, culture, or institutions; they are shaped by them. Artificial intelligence is no exception. Long before an algorithm reaches a conclusion, countless human choices have already determined what data it will learn from, which outcomes it will optimise, what risks are considered acceptable, and what counts as a successful decision.

In other words, AI systems do not replace human judgment as much as they reorganise it. The technology may automate parts of the decision-making process, but the values embedded within that process remain unmistakably human.

The Problem Isn’t Only Bias. It’s That Fairness Has No Single Definition

Much of the public discussion around artificial intelligence eventually arrives at the same conclusion: AI is biased.

That statement is often true.

But it also hides a more interesting problem.

It assumes that everyone already agrees on what an unbiased decision would look like.

In practice, they often do not.

Imagine a university using an AI system to help select applicants. Should the system prioritise grades above everything else? Should it take socio-economic disadvantage into account? Should it ignore names to reduce discrimination? Should it consider extracurricular activities? What if some schools offer far fewer opportunities than others?

Each of these approaches reflects a different understanding of fairness. None is purely technical. Before an algorithm evaluates a single application, people must first decide what values the system should pursue.

The same dilemma appears across almost every sector where AI is being adopted. Banks want lending decisions to be fair, but disagree about which financial indicators matter most. Hospitals seek accurate diagnoses while also trying to ensure that different communities receive equitable care. Governments hope automated systems will improve efficiency, yet they must also consider transparency, accountability, and public trust.

Artificial intelligence does not resolve these disagreements.

It inherits them.

Computer scientist Cathy O’Neil, in Weapons of Math Destruction, argues that algorithms often give the appearance of objectivity while quietly reproducing the assumptions built into the systems that created them. Their authority comes partly from mathematics, which can make decisions appear neutral even when they reflect contested human choices. The issue, she suggests, is not that algorithms are incapable of processing information accurately, but that they often optimise for goals that societies themselves have never fully agreed upon.

Perhaps this explains why debates about AI rarely remain technical for very long. They quickly become debates about values.

Can Responsibility Be Delegated?

If an AI system recommends a harmful decision, it is tempting to treat the technology as though it were an independent actor. Headlines often reinforce this impression. We read that “AI denied a loan,” “AI misidentified a suspect,” or “AI recommended a medical treatment.” The language is convenient, but it can also be misleading.

Artificial intelligence does not decide to enter a hospital, a courtroom, or a recruitment process. Institutions choose to place it there. They decide which problems should be automated, which decisions should remain with humans, and how much weight should be given to an algorithm’s recommendation.

This distinction matters because responsibility cannot simply be outsourced alongside technology.

Legal scholars have increasingly pointed out that responsibility in AI systems is distributed across a network of actors. Developers design the models. Companies provide the infrastructure. Organisations decide how the systems will be used. Regulators establish legal boundaries. Professionals determine whether to rely on an algorithm’s advice. Even users shape outcomes through the way they interact with these systems.

The result is something philosophers sometimes describe as the “problem of many hands.” When many people contribute to a decision, responsibility becomes more difficult to trace. No single actor appears fully responsible, yet the consequences remain very real.

Artificial intelligence has made this challenge far more visible, but it did not invent it. Large organisations have long struggled with questions of shared responsibility. AI simply exposes how complicated accountability becomes when decision-making is distributed across technical systems, institutions, and human actors.

This Is Why Governments Are Trying to Govern AI

For much of its history, artificial intelligence was discussed primarily as a technological achievement. Today, it has become a governance challenge.

Governments around the world are asking questions that would have seemed unusual only a decade ago. Should every AI system be regulated in the same way? Should an algorithm recommending music be treated differently from one helping determine prison sentences or welfare eligibility? How much transparency should companies be required to provide? And if an AI system causes harm, what obligations should fall upon developers, companies, or public institutions?

These questions are beginning to shape new laws and international frameworks. UNESCO’s Recommendation on the Ethics of Artificial Intelligence calls for AI systems that respect human rights, dignity, and fundamental freedoms. More recently, the European Union’s AI Act adopted a risk-based approach, recognising that not all AI systems pose the same level of societal risk. Rather than regulating artificial intelligence as a single technology, it distinguishes between applications according to the potential consequences they may have for people’s lives.

This shift is significant.

It suggests that governments are no longer asking only what can artificial intelligence do?

They are increasingly asking what should artificial intelligence be allowed to do?

That is not a technological question.

It is a political one.

The Responsibility Was Never Artificial

The more societies rely on artificial intelligence, the more tempting it becomes to think of machines as independent decision-makers. We speak of AI approving loans, recommending prison sentences, identifying fraud, or selecting job applicants as though responsibility somehow moves with the technology. In reality, artificial intelligence has not replaced human judgment. It has simply made the chain of responsibility longer and, at times, harder to see.

This is why discussions about AI so often become discussions about governance. Questions that initially appear technical- how an algorithm was trained, what data it learned from, or how accurate it is- quickly give way to broader institutional questions. Who decided that this process should be automated? Who determined which risks were acceptable? Who monitors the system once it is deployed? And when harm occurs, who has both the authority and the obligation to respond?

These questions are unlikely to disappear as AI becomes more sophisticated. If anything, they will become more pressing. Governments are already experimenting with AI in areas such as taxation, immigration, healthcare, policing, education, and public administration. Businesses increasingly rely on algorithms to shape hiring, pricing, customer service, and financial decisions. In each case, the technology may differ, but the underlying challenge remains remarkably similar: institutions are delegating parts of their decision-making while remaining responsible for the consequences.

Perhaps this is where the public conversation sometimes becomes misleading. We often ask whether artificial intelligence can be trusted, as though trust were a property of technology itself. But trust is rarely built through technology alone. It is built through institutions that are transparent about how technologies are used, accountable when they fail, and willing to explain the reasoning behind decisions that affect people’s lives.

Political philosopher Onora O’Neill, whose work on trust and public accountability has influenced debates far beyond artificial intelligence, argues that trust cannot simply be demanded; it must be earned through institutions that are demonstrably trustworthy. That insight feels particularly relevant today. Public confidence in AI will depend not only on whether algorithms become more accurate, but on whether the organisations deploying them remain open to scrutiny, provide meaningful oversight, and accept responsibility when things go wrong.

Seen from this perspective, artificial intelligence is not creating an entirely new problem. It is exposing an old one. Societies have always struggled with how to organise responsibility when decisions are made collectively rather than individually. AI simply makes that challenge more visible because it sits at the intersection of technology, law, business, and public governance.

The title of this essay therefore turns out to have a more complicated answer than it first appears.

Who is responsible when artificial intelligence gets it wrong?

There is rarely a single person, company, or institution that carries the entire burden. Responsibility is shared across designers, developers, organisations, regulators, and those who ultimately choose to rely on these systems. That complexity should not be mistaken for an absence of accountability. If anything, it makes accountability more important.

Perhaps, then, the more interesting question is not whether artificial intelligence can make decisions.

It is whether our institutions are prepared to remain accountable for the decisions they increasingly ask artificial intelligence to help make.