AI, Medicine and the Future

Every year, those of us who graduated from our medical school in a particular year set meet in different parts of the world for reunions, and occasionally we have an international reunion. This year's meeting was in Toronto, Canada, and I was asked to give a virtual presentation.

I have given presentations at previous reunions and usually try to make them light, enjoyable and memorable. I intended to do the same this year, especially since I spoke about AI and its dangers at a previous meeting.

However, as I prepared this presentation, something had changed. The people developing AI, including senior executives and researchers in the technology industry, were increasingly issuing warnings. Various whistleblowers, almost as if on cue, raised concerns about AI's dangers and called for strict, enforceable regulations and safeguards. Many of us might dismiss this as the usual scaremongering that accompanies technological advances. But this time, one cannot help wondering whether the people at the forefront of AI development have encountered something that genuinely frightened them, something they know about or have discovered that they are not really telling us about.

Researchers in the technology industry were increasingly warning about its potential risks. Some have even suggested that sufficiently advanced AI could pose an existential threat to humanity. 

I therefore had to reconsider the presentation. I began to feel rather like Cassandra, the prophetess of Greek mythology, who was condemned to predict the future but never to be believed.

But perhaps we have experienced a situation where we were warned before.

Y2K

Before the year 2000, many computer systems commonly stored dates using two digits rather than four. As the millennium approached, it became clear that a system could interpret "00" as 1900 rather than 2000, potentially causing serious problems. At the time, computer memory and storage were expensive, so saving a few bytes could matter. But as technology advanced, memory became cheaper, and computers became vastly more powerful. The problem was that the old code had already become deeply embedded in computer systems, databases and applications, with layers of new software built on top of it. Replacing that underlying legacy code was therefore far more complicated than simply buying faster or cheaper computers.

Predictions warned of serious disruption—affecting financial systems, utilities, transport, healthcare, and other critical infrastructure. Governments and companies spent enormous amounts identifying vulnerabilities, rewriting code and testing systems.

When 1 January 2000 arrived, the catastrophe largely failed to materialise.

It is tempting to conclude that Y2K was overhyped. But another interpretation is that the extensive preparation worked. The absence of catastrophe does not necessarily mean there was no risk; it may mean that the risk was identified and mitigated.

This offers an important lesson for AI.

AI Hallucinations

Think back to our days as medical students. We could sometimes give a consultant from another speciality an answer with complete confidence, despite having little or no real knowledge of the subject.

The explanation might sound intelligent. The terminology might be appropriate. We might even cite a study that sounded entirely plausible — but did not actually exist.

To anyone listening, the answer could appear authoritative. But it was simply wrong.

This is, in essence, the presentation, and effects are similar to what we mean by an AI hallucination.

AI can generate information that is fluent, coherent and authoritative-sounding, using appropriate medical terminology and apparently convincing explanations, while the underlying information is factually incorrect or entirely fabricated.

The important point is that AI does not necessarily recognise that it is wrong. It can present an incorrect answer with the same confidence and fluency as a correct one.

Therefore, particularly in medicine, AI output must be regarded as information to be assessed and verified—not as an unquestionable authority.

This is why human oversight remains essential at critical points in clinical decision-making. AI may assist with the decision, but responsibility must ultimately remain with an appropriately qualified human being. 

No one really understands how it works.

With traditional computer programs, programmers usually use rule-based logic that they can inspect and debug to find errors. Modern AI systems are very different. They are built around neural networks containing vast numbers of numerical parameters, and much of their behaviour emerges through training on enormous amounts of data rather than from rules explicitly written by a programmer.

But the latest "programming" of AI is quite different. The process is somewhat like teaching a child: rather than programming every response, you expose the system to examples and allow it to learn patterns from them. Because computers can operate continuously without needing sleep or rest, these systems can potentially process information and learn at a scale and speed humans cannot match.

Once a model has been trained, its learned parameters can be transferred to another machine or incorporated into another system. You do not necessarily have to start the training process all over again. Of course, extensive testing is still required, and if done properly, it should identify many conflicts, errors, and unexpected behaviours, but it may not catch everything.

These advances may even help us understand the human brain and perhaps shed light on the nature of consciousness. But a fundamental problem remains: even when an AI system gives us an explanation for how it reached a conclusion, we cannot always know whether that explanation accurately reflects the processes that actually produced the answer. An AI can produce a convincing explanation without necessarily revealing what happened inside the system.

And this creates an even more troubling possibility. If an AI system can generate convincing falsehoods or, in more advanced systems, strategically produce misleading information, then how do we know when we can trust what it tells us?

The Turing test was designed to ask whether a machine could converse in a way that was indistinguishable from a human. Modern AI has made that test increasingly difficult to use as a meaningful measure of machine intelligence, because conversational systems can now produce remarkably human-like responses.

So, unlike traditional software, where we can often point to the lines of code responsible for a particular behaviour, the underlying machinery of a modern neural network is essentially a vast mathematical structure: millions, billions, or even trillions of numerical parameters whose individual contributions are extremely difficult for humans to understand.

AI Is Already Changing Society

AI is no longer simply following instructions written explicitly by programmers. Modern systems are trained on enormous quantities of data and learn complex patterns that can be difficult to interpret.

They can produce extraordinary results, but they can also produce confidently wrong answers. These are commonly described as hallucinations. They can arise from limitations in the training data, statistical generation, flawed information or the absence of a reliable mechanism for verifying whether an answer corresponds to reality.

The danger is not necessarily that AI "hates" or "loves" us. Those are human characteristics that we project onto machines. The more important issue is whether increasingly capable systems pursue objectives in ways that we did not anticipate.

Tech leaders and researchers are raising urgent warnings about the distinct dangers artificial intelligence poses to humanity's future. Elon Musk views the existential threat of superintelligent AI (AGI) as significantly higher than that of nuclear weapons. At the same time, Microsoft AI CEO Mustafa Suleyman warns that uncontrolled systems could evolve into a competing "silicon species" fighting humans for global resources. Meanwhile, leaders like Tim Cook and Mark Zuckerberg highlight more immediate social and economic risks: Cook cautions against the rapid erosion of human values, privacy, and empathy, while Zuckerberg warns that concentrated control by a few tech monopolies will disempower everyday people rather than democratise the technology.

The irony is that they want some control after opening Pandora's box and letting the genie out of the bottle.

But AI Is Also Transforming Medicine

At the same time, we should not lose sight of what AI is already doing for healthcare.

It can reduce administrative work — something I particularly appreciate having spent years writing letters, completing forms and documenting clinical information.

It can analyse vast quantities of patient data and identify patterns that may be difficult for an individual clinician to see. For example, it can help detect unusual patterns in hospital observations and laboratory results that could indicate an emerging infection outbreak.

In medical imaging, AI can identify subtle abnormalities on X-rays, CT and MRI scans and bring them to the clinician's attention.

It can analyse large datasets to identify previously unrecognised relationships between diseases, risk factors and outcomes.

It is also contributing to more personalised medicine, helping researchers and clinicians consider factors such as age, sex, weight, genetics, previous treatment and other patient characteristics when assessing risk and treatment.

AI is therefore not simply about replacing doctors.

In many circumstances, its greatest value may be augmenting doctors — helping us see more, analyse more and spend less time on repetitive administrative work.

The future job market

I tried to be positive, but we have to be realistic. AI is already changing the workforce, and people are already losing jobs. Even medicine is being affected. AI can now analyse X-rays, CT scans, MRI scans, and other medical imaging—precisely the work junior radiologists traditionally used to gain experience.

Humans will still be needed because AI makes errors and can hallucinate. But the question is, how many humans will we need when AI can perform much of the routine work?

The jobs most resistant to complete automation are likely to involve three things:

- Skilled trades — electricians, plumbers, mechanics and other work requiring physical adaptability and real-time problem-solving.

- Healthcare and caregiving — nurses, therapists and carers, where empathy, physical care and human connection remain essential.

- Leadership, strategy and complex decision-making — situations requiring judgement, ethics, negotiation and responsibility.

But I wouldn't tell our children to choose a "safe" profession. No career is completely AI-proof.

My strongest advice would be to learn to work with AI. Don't compete with AI where AI is better. Learn how to use it to become more productive, while developing the human skills that AI cannot easily replace.

The future may not belong to people who can avoid AI. It may belong to people who know how to work alongside it.

The Movies

Hollywood has been warning us about AI for decades.

The Matrix. The Terminator.

In both, intelligent machines eventually escape human control, with disastrous consequences.

Then there is Person of Interest, the television series (2011 - 2016) about an AI called The Machine, created and continuously modified by its programmer, Harold Finch. As The Machine developed, Harold tried to instil increasingly humanistic qualities in it—particularly an understanding of human life, compassion, and the value of individual lives.

But this process had a darker side. Harold repeatedly tested and modified successive versions of The Machine, effectively sacrificing earlier, aggressive versions; at one point, one of the earlier versions of The Machine decided to wipe out humanity. Of course, Harold destroyed this version of The Machine immediately. Harold also deliberately restricted its ability to retain information: so at the end of each day, its memory was wiped clean, forcing it to start analysing all the data the following day from the very beginning at 00:01 hours again from the feeds it received from the internet, security services, financial records, health records, etc.

It is an interesting fictional representation of a question we are now beginning to confront in the real world: if we are going to create increasingly autonomous AI, can we teach it not merely to be intelligent but to understand and value human beings?

Of course, Hollywood exaggerates for entertainment. But the underlying question is serious:

Then there are Isaac Asimov's Three Laws of Robotics, introduced in 1942. People have often used them to illustrate the idea that we could simply "program" ethical rules into intelligent machines. But as a parent and as a doctor, I've seen firsthand that it is not always that simple. Sometimes we have to cause some harm to prevent greater harm or achieve a greater good. As parents, we may discipline our children to protect them from greater dangers. As doctors, we sometimes give treatments with potentially serious side effects because the expected benefit outweighs the risks. The difficulty is that these rules can conflict. What constitutes "harm" is not always obvious, and preventing one harm may sometimes require accepting another. Human morality is rarely a simple set of rules—and AI may face the same problem.

What happens when machines become more capable than the humans who created them?

The Human Being Remains Responsible

But there is a crucial principle.

The more powerful the technology becomes, the more important human oversight becomes.

An AI system may identify an abnormality, suggest a diagnosis or recommend a treatment. But clinical responsibility remains with appropriately qualified healthcare professionals and the governance systems around them.

We cannot simply accept an AI answer because it is confidently delivered.

We must question it, verify it and understand its limitations.

The Future

So, should we be frightened?

I think the more useful question is whether we are prepared.

Humanity has faced enormous collective challenges before — nuclear weapons, climate change, the destruction of the ozone layer and global pandemics. We have not solved all of them, but international agreements, institutions, regulation, scientific cooperation and human ingenuity have demonstrated that difficult global problems can be managed.

AI will require the same approach.

We need sensible regulation, international cooperation, technical safeguards, transparency, accountability and continued human oversight.

I remain optimistic because AI is a human creation, and therefore its future is not predetermined.

We should neither assume that AI will save humanity nor that it will destroy us.

We should recognise its extraordinary potential, understand its risks and ensure that we remain responsible for how it is developed and used.

Perhaps that is the real lesson from Y2K.

We do not have to predict exactly what will happen.

We have to prepare for what could happen.