The Ethics of Intelligent Machines: Who Is Responsible When AI Gets It Wrong?

In 2016, a software tool called COMPAS was used by courts across the United States to predict the likelihood that a defendant would reoffend. An investigation by ProPublica found that the algorithm was nearly twice as likely to falsely flag Black defendants as future criminals compared to white defendants. The company that built COMPAS disputed the findings. The defendants whose sentences were shaped by its scores had no way to challenge a decision made by a system whose inner workings were a trade secret. Nobody went to prison for the algorithm's failures. Nobody was held accountable at all.

This case sits at the heart of one of the most pressing questions in contemporary technology: when an artificial intelligence system causes harm, who bears responsibility? The question is not merely philosophical. As AI systems are deployed in healthcare diagnostics, credit scoring, university admissions, hiring decisions, and criminal sentencing, the gap between the scale of their impact and the clarity of their accountability is growing into a chasm.

The challenge begins with the nature of modern machine learning itself. Unlike traditional software, where a programmer writes explicit rules that produce predictable outputs, a deep learning model learns its behaviour from data. The model's decisions emerge from billions of numerical parameters adjusted through training — a process that even its creators cannot fully explain. When a neural network denies someone a loan or flags a medical scan as abnormal, there is often no clean causal chain from input to output that a human can inspect and verify. This opacity is not a bug; it is frequently a feature, the price of the extraordinary pattern-recognition capabilities that make these systems useful.

But opacity is incompatible with accountability. Legal systems built on the principle that decisions affecting individuals must be explainable and contestable are poorly equipped to handle systems that cannot explain themselves. The European Union's General Data Protection Regulation introduced a right to explanation for automated decisions, but implementing this right in practice — when the system itself cannot provide a meaningful explanation — remains an unsolved problem. Courts, regulators, and ethicists are still working out what accountability should even mean in this context.

One approach places responsibility on the developers who build AI systems. If a model produces discriminatory outcomes, the argument goes, the team that designed it, chose its training data, and deployed it should be held liable. This is a reasonable starting point, but it quickly runs into difficulties. Training data reflects the world as it is, not as it should be. A model trained on historical hiring decisions will learn to replicate historical biases, because those biases are encoded in the data. Is the developer responsible for the biases of the society whose data they used? And if the model performs well on average but fails systematically for a specific subgroup, at what point does acceptable imperfection become actionable harm?

Another approach focuses on the organisations that deploy AI systems. A hospital that uses an AI diagnostic tool, a bank that uses an algorithmic credit scorer, a court that uses a recidivism predictor — these institutions make the decision to deploy the technology and bear responsibility for its consequences. This framing has the advantage of placing accountability with entities that have legal standing and can be sued, regulated, and sanctioned. But it creates perverse incentives: if deployers bear all the risk, they may be reluctant to adopt AI systems even when those systems would, on balance, produce better outcomes than human decision-makers.

The question of bias deserves particular attention. AI systems do not create bias from nothing — they amplify and systematise biases that already exist in human institutions and historical data. A facial recognition system that performs poorly on darker skin tones does so because the training datasets were dominated by lighter-skinned faces, reflecting decades of underrepresentation in the technology industry. A natural language processing model that associates certain professions with certain genders does so because the text it was trained on reflects a world where those associations were real. Fixing these biases requires not just better algorithms but better data — and better data requires confronting the historical inequities that produced the skewed datasets in the first place.

The Islamic intellectual tradition offers a relevant framework here. The concept of maslaha — public interest or welfare — has long been used in Islamic jurisprudence to evaluate the permissibility of actions based on their consequences for the community. An AI system that produces net benefit for society while causing disproportionate harm to a specific group fails the maslaha test, regardless of its technical performance metrics. This consequentialist lens, grounded in a tradition of careful ethical reasoning, is a valuable complement to the procedural frameworks that dominate Western AI ethics discourse.

What is clear is that the current situation — where AI systems make consequential decisions affecting millions of people, with accountability diffused across developers, deployers, regulators, and the systems themselves — is not sustainable. The technology is advancing faster than the ethical and legal frameworks needed to govern it. Closing this gap requires collaboration between technologists, ethicists, legal scholars, policymakers, and the communities most affected by AI decisions. It requires transparency about how systems work and what their failure modes are. And it requires the humility to acknowledge that building a machine that can predict the future does not exempt us from responsibility for the future we are building.