Table of Contents
Type: EPQ Coursework | Subject: Health and Social Care | Level: A-Level | Word Count: ~5,000 words
This model EPQ coursework was produced by an Essays UK specialist as reference material for learning purposes only. For support in this field, see our health and social care writing specialists.
Overview: This A-Level Extended Project Qualification (EPQ) coursework evaluates the use of artificial intelligence in the United Kingdom’s National Health Service, weighing its potential to improve diagnosis, ease workforce pressure and reduce cost against risks of bias, data misuse and eroded accountability. It is structured as an evidence-based research report with a methodology, themed analysis, an overall evaluation and a reflection on the research process, and is offered purely as a structural and referencing exemplar with full Harvard (Cite Them Right) referencing.
Few technologies have attracted as much attention from governments, clinicians and the public as artificial intelligence (AI), and few sectors stand to be reshaped by it as profoundly as healthcare. In the United Kingdom, the National Health Service (NHS) faces a well-documented combination of rising demand, an ageing population, constrained budgets and persistent workforce shortages. Against this backdrop, policymakers have increasingly presented AI as part of the solution: a set of tools that might read scans faster than a radiologist, flag deteriorating patients before a crisis, automate routine administration and free clinicians to spend more time with the people in front of them. Yet the same technology raises difficult questions about fairness, safety, privacy and the proper role of human judgement in medicine.
This Extended Project Qualification (EPQ) sets out to move beyond the headlines and examine the evidence with a critical eye. It is motivated by a personal interest in both computing and the health professions, and by a conviction that technologies introduced into clinical settings deserve close scrutiny. Enthusiasm for innovation is understandable, but healthcare is a domain in which errors have serious consequences, and in which the people affected are often among the most vulnerable in society. An evaluation of AI in UK healthcare therefore has to hold two ideas together at once: that the potential benefits are real and significant, and that realising them responsibly is far from straightforward.
For the purposes of this project, “artificial intelligence” is understood broadly as computer systems that perform tasks normally requiring human intelligence, with a particular emphasis on machine learning, the branch of AI in which systems learn patterns from data rather than following explicitly programmed rules. Where specific figures or scenarios are used to illustrate an argument, they are presented as clearly hypothetical devices to demonstrate reasoning; they are not drawn from any named hospital, patient or commercial product.
The central research question guiding this EPQ is: To what extent does artificial intelligence offer a net benefit to the United Kingdom’s National Health Service? The phrasing matters. The question does not ask whether AI “works”, which would invite a simplistic yes-or-no answer, nor does it assume that benefit is the only relevant consideration. Instead it asks about the balance of advantages and disadvantages, and in doing so it commits the project to an evaluative rather than merely descriptive approach.
Three subsidiary aims support this question. The first is to establish what AI in healthcare actually consists of, cutting through marketing language to identify the categories of application that are genuinely in use or in advanced development. The second is to analyse the arguments and evidence on both sides, grouping them into coherent themes rather than treating each claim in isolation. The third is to reach a reasoned overall judgement that acknowledges uncertainty and specifies the conditions under which AI is most likely to help rather than harm. A final, reflective aim, characteristic of the EPQ as a qualification, is to evaluate the research process itself and to consider what a more extended investigation would require.
This project is a secondary research study: it synthesises and critically evaluates existing published sources rather than generating new primary data. This approach is appropriate for an A-Level EPQ for several reasons. Primary research involving patients or NHS data would raise ethical and practical barriers well beyond the scope of a school project. A literature-based study, by contrast, allows the depth and breadth of an evaluative argument to be developed responsibly from the substantial body of material already in the public domain.
Sources were selected with attention to authority, relevance and currency. Priority was given to three categories: publications from official and regulatory bodies such as NHS England, the National Institute for Health and Care Excellence (NICE), the Medicines and Healthcare products Regulatory Agency (MHRA) and the World Health Organization (WHO); peer-reviewed academic literature and recognised specialist texts; and reputable reports commissioned by government or professional organisations. Sources were appraised using a simplified version of the CRAAP test (currency, relevance, authority, accuracy and purpose). Promotional material from technology vendors was treated with particular caution, given its clear commercial purpose and its tendency to overstate benefits while understating risks.
The analytical method has three stages. First, applications of AI were mapped into categories to clarify the scope of the subject. Second, the arguments for and against were gathered and organised thematically, so that recurring considerations, such as bias or accountability, could be examined systematically rather than piecemeal. Third, the competing considerations were weighed against one another to reach an overall evaluation. Throughout, an effort was made to distinguish between what the evidence currently supports, what is plausible but unproven, and what remains speculative.
Two limitations of the methodology should be stated at the outset. Because the field moves quickly, some sources risk becoming dated, and claims about the “current” state of deployment are best read as broad characterisations rather than precise snapshots. In addition, a secondary study depends on the quality and balance of the available literature; a known feature of research in this area is publication bias towards positive results, which an evaluation must consciously counterbalance by seeking out critical perspectives.
Before the merits of AI can be evaluated, it is necessary to be clear about what is being discussed, because “AI” is used loosely in public debate to cover everything from simple automation to speculative visions of machines that reason like clinicians. Most systems relevant to healthcare today fall under machine learning, and in particular supervised learning, in which an algorithm is trained on large numbers of labelled examples, such as images already classified by experts as showing disease or not, and learns to predict the label for new, unseen cases. A subset of these methods, deep learning using artificial neural networks, has driven much of the recent progress in interpreting images and other complex data.
It is important to recognise what these systems do and do not do. A well-trained model can be extremely good at pattern recognition within the kind of data it was trained on, often matching or exceeding human performance on narrow, well-defined tasks. It does not, however, “understand” medicine in any meaningful sense, and its performance can degrade sharply when it encounters data that differ from its training set, a problem known as distribution shift. This distinction between narrow competence and general clinical reasoning runs through the whole evaluation, because many of the benefits of AI arise from the former while many of the risks arise from mistaking it for the latter.
A further clarification concerns the difference between AI as a decision-support tool and AI as an autonomous decision-maker. In almost all current UK healthcare settings, AI operates in a support role: it offers a prediction, a prioritisation or a flag that a human clinician then reviews and acts upon. Fully autonomous clinical decision-making, in which a system acts without human oversight, is rare and tightly constrained. This “human in the loop” model is central to how the technology is regulated and to how its risks are managed, and the evaluation assumes it as the realistic baseline rather than the science-fiction alternative of machines practising medicine unsupervised.
Applications of AI in UK healthcare can be grouped into several broad categories, each with a different balance of maturity and risk. Mapping them in this way helps to avoid the common error of treating “AI in healthcare” as a single thing to be judged as a whole.
Two observations follow from this mapping. First, the categories differ enormously in how directly they touch patient safety: an error in appointment scheduling is inconvenient, whereas an error in interpreting a cancer scan may be life-threatening, so an evaluation that lumps them together will reach muddled conclusions. Second, the most mature and lowest-risk categories, notably administrative automation, often receive the least public attention, while the headline-grabbing diagnostic applications are precisely those where the stakes are highest.
The arguments in favour of AI in the NHS can be organised into three themes: improving the accuracy and speed of diagnosis, easing pressure on an overstretched workforce, and improving the efficiency and sustainability of a financially constrained system.
The most compelling case for AI concerns diagnosis, and in particular the interpretation of medical images. The underlying logic is straightforward: certain diagnostic tasks are essentially pattern-recognition problems at which suitably trained machine-learning systems can excel, and they can perform them consistently, rapidly and without fatigue. Where a human radiologist’s concentration wanes over a long list, an algorithm processes each image with the same attentiveness, around the clock.
Consider a simplified, illustrative scenario in a breast-screening programme, used here only to demonstrate the reasoning rather than to report real figures. Suppose that out of every 1,000 scans reviewed, a small number contain subtle early-stage cancers, and that an AI second-reader flags cases for closer human inspection. If the system reliably draws attention to borderline cases that a tired human reader might otherwise pass over, the overall sensitivity of the programme could improve, catching cancers earlier when they are more treatable, while a well-calibrated tool can also reduce unnecessary recalls of healthy patients. The illustrative table below summarises how such a tool might be positioned.
| AI Application | Primary Benefit | Principal Risk | Relative Maturity |
|---|---|---|---|
| Imaging second-reader | Earlier, more consistent detection | Missed or false alarms if miscalibrated | High |
| Deterioration risk alerts | Faster intervention for at-risk patients | Alert fatigue; over-reliance | Medium |
| Administrative automation | Reclaimed clinician time | Errors propagated through records | High |
| Symptom-checker triage | Better signposting of care | Unsafe advice; inequitable access | Medium |
Speed matters as much as accuracy. In conditions such as stroke, where the window for effective treatment is measured in minutes and hours, a system that rapidly analyses a brain scan and alerts the relevant team can shorten the time to treatment and improve outcomes. Topol (2019), in an influential review of the digital future of the healthcare workforce, argues that the central promise of such technologies is not to replace clinicians but to give them back time and attention by handling tasks that machines do well. The appeal, on this view, is augmentation rather than substitution: the machine does the tireless screening, and the human applies judgement where it matters.
The second theme concerns the workforce. The NHS operates under chronic staff shortages across many professions, and the gap between demand and capacity shows little sign of closing through recruitment alone. In this context, the argument for AI is one of leverage: if technology can absorb a meaningful share of routine, repetitive work, each member of staff can focus their scarce time on the activities that most require human skill and compassion.
Administrative automation is the clearest example. A great deal of clinical time is consumed not by seeing patients but by documentation, correspondence, coding and the search for information scattered across systems. Tools that transcribe and summarise consultations, draft routine letters or surface the relevant parts of a patient’s history can, in principle, return hours to each clinician’s week. Because these tasks are lower-risk than diagnosis, the benefits can often be captured with less danger to patients, which is why some commentators argue that the NHS should prioritise this “unglamorous” category.
There is also a quality-of-working-life dimension. Burnout among healthcare staff is widely linked to administrative burden, so if AI reduces that burden it may improve retention as well as productivity. The claim should not be overstated: technology that is poorly designed or imposed without consultation can add to workload rather than reduce it, a point the critical sections below take seriously. But the underlying argument, that the greatest near-term value of AI may lie in giving time back to people, is a strong one.
The third theme is economic. A publicly funded health service operating under tight financial constraints has a duty to use its resources efficiently, and AI is frequently presented as a way to do more with the same or fewer inputs. The mechanisms are several: earlier diagnosis can mean cheaper treatment of disease caught at an earlier stage; better prediction of which patients are at risk of deterioration or readmission can target preventive effort where it yields most benefit; and optimisation of scheduling, bed management and theatre lists can reduce the waste of expensive capacity sitting idle.
Crucially, efficiency gains in healthcare are not simply about saving money for its own sake; in a system where resources are finite, money saved in one area can be redirected to meet unmet need elsewhere. A tool that reduces missed appointments, for instance, increases the effective capacity of a clinic without any new staff, which in a rationed system is equivalent to treating more people. The economic case therefore connects directly to the ethical goal of maximising the health benefit obtained from a fixed budget.
Caution is nonetheless required. Claims of cost savings often rest on assumptions about scale, uptake and sustained performance that may not hold in practice, and they frequently ignore the substantial costs of procurement, integration, staff training, ongoing monitoring and maintenance. A tool that performs well in a research setting may deliver far smaller real-world savings once these costs are included, or may simply shift effort from one part of the system to another. The economic promise of AI is real but conditional, and an honest evaluation must treat optimistic projections as hypotheses to be tested rather than results to be assumed.
If the benefits are organised into three themes, so too are the risks, though they cut across several more. This section examines four of the most significant: algorithmic bias and its effect on health inequality; data privacy, security and consent; accountability and the opacity of complex models; and the longer-term risk of deskilling and the loss of the human element in care. Taken together, these are not reasons to reject AI, but they define the conditions that must be met for its use to be responsible.
The most serious risk, because it can harm the people least able to challenge it, is algorithmic bias. Machine-learning systems learn from historical data, and if that data reflects existing inequalities or is unrepresentative of the population the system will serve, the model can reproduce or even amplify those inequalities. A diagnostic tool trained predominantly on data from one demographic group may perform less well for others; a system built on images of lighter skin, for example, may be less reliable at detecting conditions on darker skin, with the result that those patients receive a poorer standard of automated support.
This risk is particularly insidious because it can be invisible. An overall accuracy figure that looks impressive may conceal large disparities between subgroups, so that a system performing well “on average” may be systematically failing a minority until the pattern is actively looked for. The ethical stakes are high: a technology introduced in the name of better care could, if deployed carelessly, widen rather than narrow the health inequalities that already divide outcomes along lines of ethnicity, socioeconomic status and geography.
The mitigation is demanding but clear in principle. It requires diverse and representative training data, rigorous evaluation of performance across subgroups rather than in aggregate, transparency about the populations on which a tool was validated, and ongoing monitoring once a system is in use. The WHO (2021), in its guidance on the ethics and governance of AI for health, places equity and the avoidance of bias among its core principles, reflecting a broad consensus that fairness cannot be an afterthought. The difficulty is that meeting these requirements is costly and slows deployment, creating a tension between the drive to innovate quickly and the duty to innovate safely.
Machine learning depends on data, and in healthcare that data is among the most sensitive that exists. Training and running AI systems typically requires access to large volumes of patient records, which raises acute questions about privacy, security and the basis on which such data may be used. In the UK, the processing of health data is governed by data-protection law, principally the UK General Data Protection Regulation and the Data Protection Act 2018, overseen by the Information Commissioner’s Office, and by the common-law duty of confidentiality that has long governed the clinician-patient relationship (Information Commissioner’s Office, 2023).
Several distinct concerns arise. The first is consent and legitimate use: patients provide their data to receive care, and it is not obvious that they have agreed to its use for training commercial algorithms, even where the ultimate purpose is benign. The second is security: concentrating large datasets creates a valuable target, and a breach of health records is far more damaging than a breach of, say, shopping habits, because the information cannot be changed. The third concerns private technology companies, which often supply the expertise the NHS lacks but whose commercial interests may not align neatly with the public interest in a system built on taxpayer-funded data.
These concerns are not arguments against using health data at all, which would be to forgo its considerable potential to improve care, but they do demand robust safeguards: clear legal bases for processing, strong technical security, transparency with the public about how their data is used and by whom, and governance arrangements that keep control of the data firmly in public hands. Public confidence, once lost, is very hard to regain, and is itself a precondition for the whole enterprise.
A third cluster of risks concerns responsibility. Medicine operates within a framework of professional accountability through which, when something goes wrong, responsibility can be established and lessons learned. The introduction of AI complicates this framework in two ways.
The first is the problem of opacity, often called the “black box” problem. Many of the most powerful machine-learning models, particularly deep neural networks, arrive at their outputs through processes that are not readily interpretable, even by their designers. A model may correctly identify a tumour without being able to explain, in terms a clinician can scrutinise, why it reached that conclusion. This matters because medicine values not only correct answers but the ability to reason about them. A recommendation that cannot be interrogated is difficult to trust appropriately: a clinician may either over-rely on it, deferring to the machine against their own judgement, or dismiss it, losing its value.
The second is the question of liability. If a clinician follows an AI recommendation that turns out to be wrong, or overrides a correct one, where does responsibility lie, with the clinician, the hospital that deployed the system, or the company that built it? Existing lines of accountability assume a human decision-maker; distributing a decision across human and machine blurs them. The prevailing answer, that the human clinician remains responsible and the AI is a tool that informs but does not make the decision, preserves accountability but places a heavy burden on individual practitioners, who are expected to supervise systems whose inner workings they may not be able to inspect. Resolving this tension is partly a regulatory and legal task and partly a matter of designing systems that are interpretable enough to be supervised meaningfully.
The final risk is slower-acting and harder to quantify, but potentially profound. If clinicians come to rely on AI for tasks they once performed themselves, there is a danger that the relevant skills atrophy, a process sometimes called deskilling or automation complacency. A radiologist who rarely reads a scan without an algorithm’s prompt may, over time, become less able to detect what the algorithm misses, which is precisely the situation in which human expertise is most needed. The concern is not hypothetical; analogous effects have been observed in other safety-critical fields where automation has reduced hands-on practice.
Beyond individual skills lies a broader concern about the character of care. Medicine is not only a technical activity but a relational one; much of its value to patients lies in being listened to, understood and reassured by another human being. There is a risk that an efficiency-driven enthusiasm for automation could, if pursued without care, erode the time and attention devoted to this human element, treating it as a residual rather than the core of what healthcare is for. The appropriate response is to be deliberate about which tasks are automated and why: automation that removes drudgery and returns time to the relationship is to be welcomed, whereas automation that squeezes the relationship itself to extract further efficiency is self-defeating.
The risks described above do not arise in a vacuum; they are the subject of an evolving ethical and regulatory framework whose adequacy is itself part of the evaluation. In the UK, AI tools that meet the definition of a medical device are regulated by the MHRA, which assesses their safety and performance before they may be placed on the market (Medicines and Healthcare products Regulatory Agency, 2023). NICE has developed evidence-standards frameworks setting out what digital health technologies must demonstrate to justify adoption, providing a route by which claims of clinical and economic benefit can be tested rather than taken on trust (National Institute for Health and Care Excellence, 2022). At the level of principle, the WHO’s guidance articulates values, spanning safety, transparency, accountability and equity, that command wide international agreement.
Four classical principles of biomedical ethics offer a useful lens through which to view these arrangements. Beneficence and non-maleficence, the duties to do good and to avoid harm, require that benefits be demonstrated and risks minimised before and during deployment, which is the purpose of regulatory approval and ongoing monitoring. Autonomy bears on consent and on patients’ right to understand and, within limits, to influence how technology is used in their care. Justice speaks directly to the problem of bias and to ensuring that the benefits of AI are shared fairly rather than accruing only to those already well served. The value of this framework is that it reframes the regulatory questions as ethical ones: the point of requiring representative validation data is not mere compliance but the pursuit of justice, and the point of insisting on human oversight is to preserve accountability and respect for autonomy.
A recurring difficulty, however, is that regulation tends to lag behind technology. AI systems can be updated continuously, so that a tool approved in one form may behave differently after later changes, which strains a regulatory model designed for static products such as drugs or fixed devices. Keeping oversight meaningful in the face of systems that learn and change is an unsolved problem, and one that an honest evaluation must flag rather than wish away: the existence of a framework is reassuring, but its sufficiency, given the pace of change, remains an open question.
Returning to the research question, the analysis supports a qualified and conditional affirmative. The benefits are genuine and, in some areas, substantial: AI can improve the consistency and speed of certain diagnoses, relieve staff of routine burdens, and help a constrained system use its resources more efficiently. These are not trivial gains in a service under the pressures the NHS faces, and to reject them outright on the grounds of risk would itself carry an ethical cost, since failing to adopt a safe, beneficial technology means tolerating worse outcomes than are achievable.
Yet the evaluation also shows that the benefits are neither automatic nor uniform, and that they are shadowed by real risks that fall hardest on the most vulnerable. The value of AI depends heavily on how it is implemented: on whether training data are representative, on whether performance is monitored across subgroups and over time, on whether data are handled lawfully and securely, on whether accountability is preserved through meaningful human oversight, and on whether automation is directed at drudgery rather than at the human relationship at the heart of care. The same technology, deployed carelessly, could entrench inequality, erode privacy and obscure responsibility.
The most defensible conclusion, therefore, is that the question “does AI benefit the NHS?” is less useful than the question “under what conditions does AI benefit the NHS?” The evidence suggests that the net benefit is positive where three conditions hold: where the application is matched to an appropriate level of risk, so that the highest-stakes uses face the most stringent safeguards; where human clinicians remain responsible decision-makers supported rather than supplanted by the technology; and where robust, independent evaluation and monitoring continue throughout a system’s life rather than ending at the point of approval. Where these conditions are absent, the balance can easily tip the other way. It is also notable that the applications offering the clearest, lowest-risk gains, particularly administrative automation, are not those that dominate public discussion, which suggests that a sober strategy would invest first where the risk-adjusted benefit is highest rather than where the technology is most eye-catching. The choice is not between innovation and safety but about sequencing them.
This EPQ set out to evaluate the extent to which artificial intelligence offers a net benefit to the UK’s National Health Service. Having mapped the field, examined the arguments for and against across clearly defined themes, and considered the ethical and regulatory context, it concludes that AI offers a real but conditional net benefit. The technology can improve diagnosis, relieve workforce pressure and enhance efficiency, but these gains are contingent on responsible implementation that addresses bias, protects data, preserves accountability and safeguards the human dimension of care. The honest answer to the research question is therefore not a simple “yes” or “no” but “yes, under specifiable conditions, and with greatest confidence in the lower-risk applications.”
The wider significance of this conclusion is that the debate about AI in healthcare is often framed as a contest between enthusiasts and sceptics, when the more productive question concerns governance: how to capture the benefits while managing the risks. Framed this way, the task before the NHS is not to decide whether to use AI but to build the data, regulatory and professional infrastructure that allows it to be used well. The determinants of success are less the sophistication of the algorithms than the quality of the systems, safeguards and judgement that surround them.
A distinctive feature of the EPQ is the requirement to reflect critically on the project itself, and several lessons stand out. The first concerns the management of scope. AI in healthcare is an enormous subject, and an early draft of the research question was far too broad to answer in the available length. Narrowing the focus to a single health system and committing to an explicitly evaluative question made the project tractable and gave the analysis a clear spine. In hindsight, this narrowing should have happened sooner, sparing a good deal of loosely relevant early reading.
The second lesson concerns the evaluation of sources. The literature on AI in healthcare is uneven in quality and often shaped by the interests of those producing it, with a marked tendency towards optimistic reporting. Applying a consistent appraisal framework and deliberately seeking out critical and regulatory perspectives was essential to producing a balanced account rather than an uncritical summary of promotional claims. If the project were extended, a systematic approach to searching and recording sources, closer to a formal literature review, would strengthen the rigour and reduce the risk of selection bias in the material considered.
A third reflection concerns the limits of a secondary study. This project can weigh published arguments but cannot test them against new evidence, and some of the most important questions, such as how a given tool performs in a particular NHS setting or how patients feel about the use of their data, can only be answered through primary research under proper ethical oversight. A natural extension would be a small primary study, for example a survey of public attitudes to AI in the NHS, which would complement the literature-based analysis with direct evidence and would require careful attention to ethics and sampling. Recognising these boundaries is itself a finding: it clarifies what this project establishes, what it only suggests, and what remains for further investigation.
Finally, the project has reshaped my own thinking. I began with a broadly optimistic view shaped by media coverage of diagnostic breakthroughs and ended with a more measured position that locates the real challenge not in the technology but in its governance. That shift, from asking whether AI works to asking under what conditions it helps, is perhaps the most valuable outcome of the research, and a habit of thought that extends well beyond this subject.
You May Also Like
All work is written by human writers. 100% AI free, guaranteed.
100% money back guarantee if you find plagiarism in our work.
COMPANY DETAILS
