An
AI system at Moorfields Eye Hospital in London can diagnose over 50
sight-threatening eye conditions from retinal scans with accuracy matching
the best consultant ophthalmologists, and can do so in seconds from a scan
that previously required a specialist appointment. The system, developed in
partnership with Google DeepMind and validated in peer-reviewed research
published in Nature Medicine, is not a research prototype. It is in clinical
use, diagnosing real patients, and its deployment has demonstrably reduced
waiting times for conditions including macular degeneration and diabetic
retinopathy that cause irreversible vision loss if not treated promptly. This
is AI in healthcare at its most impressive: validated, deployed, and making a
measurable difference to patient outcomes in a high-stakes clinical
domain.

The Moorfields example is real and significant. It is also not
representative of most AI in healthcare, which encompasses a much wider range
of applications with much more variable evidence quality, clinical utility,
and governance maturity. Understanding the difference between the Moorfields
model of rigorously validated, clinically integrated AI and the broader
landscape of AI products marketed to health systems without equivalent
evidence requires engagement with the healthcare AI ecosystem that the
enthusiastic headline coverage of individual successes often does not
encourage. The revolution in healthcare that AI promises is partly already underway,
partly years from clinical realisation, and partly in danger of causing harm
if deployed without the rigour that clinical medicine
demands.

Diagnostics: Where the Evidence Is Strongest

AI diagnostic tools have the strongest evidence base of any
healthcare AI application category, and several have achieved the unusual
distinction of being validated in independent prospective clinical trials
rather than retrospective analyses of archived data. Diabetic retinopathy
screening AI approved by the FDA, CE marked in the EU, and assessed by NICE
in the UK has been shown to perform at specialist level across diverse
patient populations, enabling screening programmes to operate at volumes that
the specialist ophthalmology workforce cannot sustain. AI mammography reading
tools are under evaluation for UK national breast cancer screening, with
trial results suggesting they can safely reduce radiologist workload while
maintaining or improving sensitivity. Dermatology AI that identifies skin
cancer from clinical photographs has been validated at consultant
dermatologist level in multiple independent studies.

The common characteristics of validated diagnostic AI are worth
noting: large, diverse training datasets; prospective rather than
retrospective validation; testing across different clinical settings and
patient demographics; and peer-reviewed publication of methods and results.
These are the same standards that pharmaceutical treatments must meet before
clinical deployment, and the AI diagnostic tools meeting them represent
genuine advances in medical capability. The NICE Evidence Standards Framework
for digital health technologies provides a tiered evidence requirement that
has become the de facto regulatory standard for AI diagnostics in the NHS,
providing a benchmark against which AI product claims can be
evaluated.

Clinical Decision Support and Its Limits

Beyond diagnostics, AI clinical decision support tools are being
deployed across a wide range of applications including sepsis prediction,
medication safety checking, surgical planning, and mental health risk
assessment. The evidence base for these applications is more heterogeneous
than for diagnostics, and the implementation challenges are often greater. A
diagnostic AI that produces a binary output, disease present or absent, with
a confidence score that clinicians can interpret, is simpler to integrate
than a clinical decision support system that generates recommendations across
multiple clinical domains, requires integration with complex electronic
health record systems, and needs to fit into clinical workflows designed
around human rather than AI decision-making.

The risk of automation bias, in which clinicians over-rely on AI
recommendations even when their own clinical judgement would have led to a
better decision, is documented in multiple healthcare settings and represents
one of the most significant safety concerns with clinical AI deployment.
Research from King’s College London found that clinicians shown AI risk
scores for patient deterioration adjusted their own assessments toward the AI
output even when the AI was demonstrably wrong, a finding with significant
implications for how AI clinical support should be designed and governed. The
answer is not necessarily to avoid AI in clinical support, but to design
human-AI interaction in ways that preserve appropriate clinical scepticism
rather than creating uncritical deference to algorithmic
outputs.

Drug Discovery and Development

AlphaFold2’s prediction of protein structures, which earned its
developers the Nobel Prize in Chemistry in 2024, represents perhaps the most
significant scientific contribution of AI to medicine. The ability to predict
the three-dimensional structure of proteins from their amino acid sequences
at accuracy approaching experimental methods has transformed structural
biology and is accelerating drug discovery across multiple therapeutic areas.
Pharmaceutical companies including Pfizer, AstraZeneca, and dozens of biotech
startups are using AlphaFold-derived structural insights to identify drug
targets, design drug molecules, and predict drug-protein interactions with a
speed that was not previously possible. The translation from structural
insight to approved medicine takes years, but the AlphaFold pipeline is
already generating lead compounds in clinical trials that would not have been
identified without it.

What This Means for You

As a patient, the most important implication of AI in healthcare
is that the quality and availability of diagnosis and treatment you receive
is increasingly shaped by AI systems whose performance and governance you
cannot directly assess. Asking your healthcare provider about the role AI
plays in your care, requesting human review of AI-generated diagnoses for
serious conditions, and supporting patient advocacy organisations that hold
NHS procurement of AI tools to rigorous evidence standards are all meaningful
responses to an AI healthcare landscape that is changing faster than patient
awareness of it. The revolution AI promises in healthcare is most real in
diagnostics and drug discovery, most promising but less proven in clinical
decision support, and most uncertain and most risky in the many healthcare AI
products marketed to health systems without evidence standards comparable to
those applied to pharmaceutical treatments. The international dimensions of
AI in healthcare deserve acknowledgment alongside the NHS-centred analysis
that dominates UK coverage. Low- and middle-income countries face
dramatically different AI healthcare challenges and opportunities than
well-resourced health systems. AI diagnostic tools that reduce the specialist
burden for common conditions could be genuinely transformative in health
systems where specialist access is severely limited; the same tools deployed
in well-resourced systems primarily speed up processes that were already
reasonably functional. Investment in healthcare AI that specifically
addresses the needs of lower-resource health systems, including ensuring that
training data includes diverse populations and that deployment models do not
require infrastructure that low-income settings lack, is a global health
priority that the World Health
Organization has specifically identified in its guidance on AI in
health. For related coverage, see our analysis of AI
in mental health detection and NHS
AI deployment.

 The regulatory pathway for AI
medical devices in the UK, managed through the MHRA with NICE health
technology assessment for NHS commissioning, is more rigorous than the
pathway for AI software tools marketed as wellness or productivity products
rather than medical devices. Ensuring that the distinction between these
categories is clear and consistently enforced is a priority for the MHRA that its current
AI strategy explicitly addresses, though the volume of AI health products
reaching the market is creating enforcement pressure that existing resources
may not adequately meet.

About the Author

Stuart Kerr is a technology correspondent at LiveAIWire, covering
artificial intelligence, digital innovation, and the social impact of
emerging technologies. Follow LiveAIWire for daily analysis at liveaiwire.com.