Artificial Intelligence, Machine Learning and Robotics
What Are AI, Machine Learning and Robotics?
Artificial intelligence (AI) refers to computer systems that perform tasks requiring human-like reasoning. Machine learning is a way of building AI: instead of programming a system with explicit rules, you feed it large amounts of training data and it learns to identify patterns. Robotics combines mechanical engineering with computing to create machines that act in the physical world — everything from factory assembly arms to autonomous vehicles. These three technologies are increasingly intertwined: modern robots use AI and machine learning to make decisions in real time.
The benefits of these technologies are considerable. AI systems can diagnose diseases from scan images with accuracy comparable to specialists. Machine learning can detect fraud in banking transactions in milliseconds. Robots can carry out dangerous or repetitive tasks that would harm or exhaust human workers. However, the same capabilities raise four distinct ethical issues that your exam focuses on: accountability, safety, algorithmic bias, and legal liability.
There are four named ethical issues for AI, machine learning and robotics: accountability (who is responsible for decisions made by AI?), safety (can AI systems fail in ways that harm people?), algorithmic bias (can AI systems discriminate unfairly?), and legal liability (who is legally at fault when AI causes harm?).
Issue 1: Accountability
Accountability in AI refers to the question of who is answerable when an AI system makes a decision that affects someone’s life. When a doctor makes a clinical decision, that doctor can be asked to explain their reasoning and held to professional standards. When an AI system makes the same decision, the reasoning may be opaque — even to the developers who built it. This opacity is sometimes called the “black box” problem: the system produces outputs without making the reasoning visible or auditable.
Consider an AI system used to decide whether to grant a loan application or whether a prisoner should be considered for early release. If the system rejects an application, the affected person may have no way to find out why, no clear process for challenging the decision, and no human decision-maker to appeal to. This raises serious concerns about fairness, transparency, and the ability of individuals to contest decisions that affect them.
Many powerful machine learning models — particularly deep learning systems trained on millions of data points — make predictions through thousands of weighted calculations that cannot be easily summarised in human-readable rules. Developers can test inputs and observe outputs but may not be able to explain why a specific prediction was made for a specific individual. This creates accountability gaps: if a patient is denied treatment or a job candidate is rejected, there may be no clear human accountable for that outcome and no mechanism to review whether the decision was fair.
Issue 2: Safety
Safety concerns arise when AI systems operate in high-stakes environments where failure has physical consequences. A machine learning model trained in the lab may encounter real-world situations that differ from its training data and behave unexpectedly. In medical diagnosis, a misclassification could lead to incorrect treatment. In an autonomous vehicle, an unexpected failure could cause a road collision. In industrial robotics, a malfunctioning arm could injure a worker.
Safety concerns are also raised by adversarial attacks: techniques that deliberately feed AI systems misleading input to cause errors. For example, small, imperceptible changes to an image can cause a well-trained image classifier to confidently produce a completely wrong output. If such vulnerabilities exist in safety-critical systems — such as those used for border control, medical imaging or autonomous navigation — the consequences could be serious.
Issue 3: Algorithmic Bias
Algorithmic bias occurs when a machine learning system produces outcomes that are systematically unfair to particular groups of people. This usually happens because the training data reflects historical patterns of human prejudice. If a hiring algorithm is trained on historical employment decisions made in an era of gender or racial discrimination, it may learn to replicate those same patterns — flagging candidates from certain groups as less suitable, not because they are, but because the historical data it learned from was biased.
Facial recognition technology provides another clear example. Studies have found that some commercially deployed facial recognition systems perform significantly less accurately on darker-skinned faces and on women, compared with lighter-skinned men. If such systems are used for identity verification or law enforcement purposes, they are more likely to produce errors that disadvantage the groups they perform worst on. The bias is not intentional — it emerges from training data that was not representative of the full population.
Algorithmic bias is caused by biased training data — if the historical data used to train an AI reflects past discrimination, the model learns to replicate it. Bias can also arise from unrepresentative data — if certain groups are underrepresented in training data, the model performs less accurately for those groups. Neither form of bias requires any intent to discriminate.
Issue 4: Legal Liability
Legal liability is the question of who bears legal responsibility when an AI system causes harm. Existing legal frameworks generally assume a human decision-maker: if a car crashes due to driver negligence, the driver is liable; if a product is defective, the manufacturer is liable. When an autonomous vehicle causes a collision while operating without a human driver, or when an AI system misdiagnoses a patient, the existing framework struggles to assign responsibility cleanly.
Potential parties include: the company that developed the AI software, the organisation that deployed the system, the engineer who designed the training process, or the dataset provider. In practice, contracts and insurance arrangements may allocate liability, but this creates inconsistency and leaves affected individuals uncertain about whether they can seek redress. Legislators and courts in the UK and internationally are actively working to develop legal frameworks that can address AI liability, but this remains an open and contested area.
An autonomous vehicle illustrates all four issues simultaneously. Accountability: if the car makes a decision in a near-miss scenario, who decided how that decision would be made, and can it be reviewed? Safety: what happens when the car encounters weather conditions, road layouts or unexpected obstacles not well-represented in its training data? Algorithmic bias: if pedestrian detection systems were trained predominantly on data from one type of setting, they may perform less reliably in other environments. Legal liability: if a collision occurs in fully autonomous mode, is the car manufacturer, the software developer or the vehicle owner responsible? The self-driving car is not the only AI system raising these questions — it is simply an unusually visible example of issues that apply across healthcare AI, predictive policing, recruitment software and financial systems.