No equations, lots of pictures.
Designed for toddlers, ideal for the general public but interesting enough for the medically inclined.
Everyone is talking about artificial intelligence. The hype has seeped into every corner of industry; Collins Dictionary’s 2025 word of the year was ‘vibe coding’, meaning the use of AI by non-experts to write code to produce functional programs. Unsurprisingly, these technological shifts are already bubbling into medicine – both its scientific foundations and its clinical realities.
As AI becomes more prevalent in medical practice, it is important for clinicians to understand what AI is, and critically what it is not. Where will AI fly and falter? Which tools have already slipped quietly into daily practice, and which remain closer to science fiction? Who may benefit the most from these innovations?
Broadly, we will cover two types of AI: generative AI, which creates text, images, or other outputs (think medical chatbots or tools that summarise records), and analytical AI, which interprets data (e.g., blood tests, scans, vital signs) to predict outcomes or support diagnoses.
At the Gazette, we are acutely aware of how relevant these issues are to current medical students. Our aim is to equip you with confidence, literacy, and a sense of informed curiosity.
AI is here to stay. Time to get acquainted.
The dream of electronic brains
Turns out stuff happened before ChatGPT came out in 2022.
Artificial intelligence may feel like a phenomenon of the 2020s, but its history is long and winding. Computer scientists have dreamed of creating machines capable of reasoning, learning, and problem-solving for three-quarters of a century. Before we dive into the nuts and bolts of modern AI, it helps to start with a few foundations.
To create an ‘electronic brain’, it is essential to distill what makes human intelligence intelligent. Historically, two major schools of thought emerged. Firstly, if reasoning is the essence of intelligence, a logic-inspired approach is the best approximation of human cognition. This gave rise to symbolic AI, which relies on explicitly coded logical rules, which dominated early AI research. These rules, however, are inherently less flexible and limit the versatility of algorithms.
The second approach took biological inspiration. What if learning -not logic – formed the foundation of intelligence? From this theory, neural networks composed of artificial neurons were born.
Neurons, artificial or otherwise, are simple processing units. Individually they are simple; collectively, when arranged in networks with weighted connections, they can model complex patterns. Artificial, like biological neurons, vary in their intrinsic properties (e.g. sensory vs motor neurons) that influence their role in a network. Within the network, the stronger the weight of the connection, the more attention a neuron will pay to a given input.
Artificial neurons and neural networks were first described in a landmark paper by Walter Pitts and Warren McCulloch in 1943. Their work was a revolutionary crossover between neuroscience and artificial intelligence, and laid the foundation for modern artificial intelligence.
Neural nets, as these networks are commonly known, may come in many different sizes and shapes depending on their architecture. Like the brain, artificial neurons are organised into layers with an input layer, one (or more) processing layers and an output layer. If a network has more than one processing, or hidden, layer it is called deep learning – in essence, the more hidden layers, the more processing steps.
Think of it like a police investigation. A single detective might spot obvious clues, but a team of officers working behind the scenes can each focus on different aspects of the case: motive, opportunity, forensics, and alibis. By combining these partial insights, the team is far more likely to uncover the truth.
Similarly, deeper neural networks distribute the task of understanding data across many layers, enabling them to solve more complex problems with greater accuracy.
To train these neural networks, there are three main training regimes: supervised, unsupervised and reinforcement learning. Different approaches are best suited to different types of problems.
- Supervised learning is great for classification problems, such as recognising whether an image contains a fire hydrant for a CAPTCHA test – think about teaching through a set of labelled flashcards.
- Unsupervised learning is where the system uncovers structure on its own, useful for detecting anomalies – this is why your bank questions your 3 a.m. Instagram purchases.
- Reinforcement learning is like playing football and learning through reward – win more matches, refine your strategy.
Despite promising beginnings, the field slumped in the 70s. Partially due to limited funding and computing power, symbolic AI plateaued and machine learning slumbered. The tide turned in 1986 when Geoffrey Hinton and colleagues popularised a concept known as backpropagation. Backpropagation is a method for efficient calculation of error and adjustment of network weights. This enables the model to learn from its mistakes by tracing them back to the decisions that caused them. For those interested, the relevant mathematics involves heavy matrix operations, and the challenge of optimising them has shaped modern AI hardware design.
With new methods, faster computers, and abundant data, AI research accelerated. To drive innovation, competitions such as the ImageNet challenge pushed researchers to develop models capable of recognising thousands of images. The best performing architectures were ‘convolutional neural nets’ (CNNs), where information flows in a single direction through multiple different layers or filters. Much like the visual cortex, these layers include lower-order (e.g., lines) and higher-order (e.g., whole eyes) features combining to detect whole objects (e.g., faces).
Yet the revolution that propelled AI into mainstream consciousness was the creation of large language models (LLMs). In 2017, researchers introduced ‘transformer architecture’. This builds on recurrent neural nets, where the most current output loops over itself to hold an internal memory, in order to predict the next word (known as a token). Transformers could analyse text more efficiently and capture long-range relationships between words, dramatically improving language modelling.
Context is critical to language, as the meaning of a word can shift dramatically depending on its surroundings. Consider the difference between “the patient is in septic shock” and “the patient received an electric shock during resuscitation.” Although the same word is used, the intended meanings are entirely different. The ability of transformer-based models to retain and weigh contextual information is essential for making such distinctions accurately.
These developments catapulted AI into public awareness as companies like OpenAI published free-for-use generative AI model ChatGPT.
Artificial intelligence continues to evolve at breathtaking speed. By the time you read this, new architectures, training methods, and clinical applications will already be emerging. But hopefully, with some basic understanding as anchor and compass, you may feel less lost at sea in these uncertain waters.
by the time you read this, new architectures, new training methods, and new clinical applications will already be emerging.
Deus Ex Machina?
Clinical implementation of AI is already here, and we are likely to see it expand substantially across the next few decades. The Gazette interviewed Dr Hutan Ashrafian, the Lead of AI at the Global Health Institute at Imperial College London and Chief Scientific Officer at the venture capital firm Flagship Pioneering, for his perspective on the future role of AI. We have included some of his insights in this piece, and you can read the full length interview on the OMSG website.
Medical imaging is an area where AI is likely to triumph.
Dr Ashrafian led a project in collaboration with Google DeepMind on breast cancer screening, where an AI-assisted mammography analysis was non-inferior to two radiologists, and superior to one. His work is already being rolled out and several trusts have already begun to adopt AI-assisted imaging. Wider uptake seems inevitable.
There will be challenges. Jurassic-era IT systems may have to be dragged kicking and screaming, but the advantages of these technologies will prevail. In terms of how it affects the modern day medical student, Hutan argues that ‘[some] suggested AI would replace radiologists, but the need for radiologists has actually increased. AI replaces tasks, but our reliance on imaging and non-radiation tools is growing.’
One of the most significant breakthroughs of the next decade may be the development of agentic AI: systems capable of executing complex tasks with partial autonomy, including creating and pursuing sub-goals. In medicine, these systems could autonomously triage referrals, monitor and maintain patient status or function as virtual multi-disciplinary teams under broad human oversight. Health data will expand in tandem, ‘your watch, clothes, jewellery, and even the glasses you drink from will be the diagnostic tools of the future.’ A digital Dr House.
By contrast, large-language models, including the medical, may be reaching their technological plateau. Whilst undeniably powerful, LLMs lack true reasoning capacity. Therefore, their expanding role is likely to entail broader application rather than a fundamental paradigm shift. For example, real-time note taking in a GP consult could massively improve the efficiency of this under-resourced service. No more doctors typing away at a screen in a consultation, ‘AI should be used not just for pathology but to minimize bureaucracy, giving doctors more time with patients.’
As AI becomes increasingly embedded in medical practice, governance will be one of the most critical areas of development. The law has historically lagged behind technological progress but the potential ramifications of poor regulation in healthcare could be profound. To tackle this, Dr Ashrafian developed TURBO-aligned industry standards; ’it must be Testable, Understandable, Reliable, free from Bias, and Operable.’
Regulatory frameworks must strike a careful balance: excessive restriction may stifle innovation, while insufficient oversight risks patient harm and erosion of trust.
Furthermore, we must examine the wider societal implications of AI. Can AI be used to minimise existing health disparities instead of widening socioeconomic health inequality? What are the environmental costs of large-scale model training that may disproportionately hurt lower income countries? Will we train models with implicit biases to favour privileged affluent patients?
And the everlasting question, will AI replace doctors? Well, according to ChatGPT, “Unlikely—AI will transform and support medical work, but doctors’ judgment, empathy, and decision-making remain essential.”
And a final piece of advice from Hutan, “Don’t only follow orthodoxy; follow innovation. The application of AI will make the most difference in healthcare and life. You are in the right profession at a world-changing inflection point.”
Illustrated by Suzan Mozak.
