That is the version of AI in healthcare that gets celebrated at conferences and in press releases. It is real, it is significant, and it understates both the breadth of what AI is already doing in medicine and the genuine complexity of the questions it raises.
This
post covers the full picture: what AI is genuinely doing in healthcare right
now, where it is working most clearly, where it is failing or falling short,
the ethical questions it raises that deserve honest engagement rather than
dismissal, and what the realistic trajectory looks like over the next several
years.
Where AI Is Already Working in Healthcare
Medical
Imaging and Radiology
Radiology is the area where AI has produced the most consistent and clinically validated results. AI systems trained on millions of medical images can identify patterns in chest X-rays, CT scans, MRIs, and retinal photographs that correlate with specific diagnoses, often with sensitivity and specificity that matches or exceeds experienced radiologists on specific tasks.
Google's AMSCAN system and DeepMind's work on mammography detection have both demonstrated accuracy comparable to specialist radiologists in controlled studies. Zebra Medical Vision and similar companies have built FDA-cleared AI tools for detecting specific conditions from imaging including liver disease, cardiovascular risk markers, and bone density issues. In contexts where specialist radiologists are in short supply, which describes most of the world outside wealthy urban centres, these tools have the potential to extend diagnostic capability significantly.
The important caveat that clinical researchers consistently emphasise: performance in controlled studies does not always translate directly to performance in real clinical environments, where imaging quality varies, patient populations differ from training data, and the integration with clinical workflow adds complexity that benchmark testing does not capture.
Drug
Discovery and Development
Drug discovery is one of the most expensive and time-consuming processes in science. Bringing a new drug from initial discovery to market takes an average of twelve to fifteen years and costs over a billion dollars, with a failure rate exceeding ninety percent at various stages of clinical trials. AI is being applied at multiple points in this pipeline with genuinely promising results.
AlphaFold's protein structure prediction has accelerated the identification of potential drug targets by giving researchers an unprecedented view of how proteins are shaped and how molecules might interact with them. Insilico Medicine used AI to identify a novel drug candidate for idiopathic pulmonary fibrosis in 18 months at a fraction of the traditional cost, and that candidate has entered clinical trials. Recursion Pharmaceuticals is running AI-driven drug discovery across multiple disease areas simultaneously in a way that traditional laboratory methods cannot match for scale.
The honest assessment at this stage: AI has demonstrably accelerated early-stage drug discovery. Whether AI-discovered drug candidates will prove clinically effective at rates that justify the shift in approach is a question that clinical trial results over the next five years will begin to answer definitively.
Clinical
Documentation and Administrative Burden
This is less glamorous than drug discovery but arguably more immediately impactful for the daily experience of healthcare professionals. Clinical documentation, the recording of patient encounters, diagnoses, treatment plans, and follow-up notes, consumes an enormous share of physician time. Studies in the United States have found that physicians spend roughly two hours on administrative and documentation tasks for every one hour of direct patient care.
AI-powered ambient documentation tools, including Microsoft's DAX Copilot and similar systems, listen to a patient-physician conversation and automatically generate a structured clinical note. In early deployments, physicians using these tools reported saving an average of 45 minutes per day on documentation, time that could be redirected to patients. The quality of generated notes requires physician review, but the starting point it provides is significantly better than a blank screen.
Early
Warning and Predictive Analytics
AI models trained on electronic health record data can identify patients at elevated risk of deterioration before clinical signs become obvious. Systems predicting sepsis onset, acute kidney injury, and post-operative complications have been deployed in hospitals across multiple countries, with studies showing earlier interventions and improved outcomes in some contexts.
A
community health programme in rural Kenya is using an AI model trained on local
patient data to identify children at risk of severe malnutrition before acute
symptoms present, enabling preventive interventions rather than emergency
responses. This application of predictive AI in a low-resource healthcare
context is less visible in global coverage than headline drug discovery stories
but represents one of the most practically significant uses of AI in health.
Where AI in Healthcare Is Falling Short
The honest account of AI in healthcare requires engaging with the places where it is not working as well as the places where it is, because the two are inseparable from a policy and deployment perspective.
The
Training Data Problem
AI systems learn from the data they are trained on. Medical AI systems have predominantly been trained on data from wealthy, urban, predominantly Western healthcare settings. The result is systems that perform well on populations similar to their training data and less well on populations that differ from it. A skin cancer detection AI trained predominantly on images from lighter-skinned patients performs measurably worse on darker-skinned patients. A sepsis prediction model trained on data from a large urban hospital may not generalise to a rural community hospital with a different patient demographic and care pathway.
This is not a marginal technical issue. It is a fundamental challenge that, if not addressed through deliberate training data diversification and rigorous testing across populations, could result in AI systems that systematically provide lower-quality care to already underserved communities.
The
Validation Gap
The number of peer-reviewed papers demonstrating AI's performance on clinical tasks in controlled settings vastly exceeds the number of rigorous studies demonstrating real-world clinical benefit in deployed healthcare systems. Performance on a retrospective dataset and performance in a live clinical environment are different things, and the gap between them has been larger than initially anticipated in several high-profile deployments.
A widely discussed 2019 Nature paper showed that an AI system for detecting diabetic retinopathy from retinal images performed excellently in laboratory conditions but underperformed in a real deployment partly because the image quality in actual clinical settings was lower than in the training dataset. This pattern has appeared in multiple subsequent studies and has made regulators and clinical implementers appropriately cautious about moving from impressive benchmark results to clinical deployment.
Liability
and Accountability
When
an AI system contributes to a clinical decision that results in patient harm,
who is responsible? The physician who used the tool? The hospital that deployed
it? The company that built it? These questions do not yet have clear legal
answers in most jurisdictions, and their absence is one of the genuine barriers
to broader clinical adoption. Physicians are understandably reluctant to rely
on tools whose error modes they do not fully understand and whose failures they
may be held personally accountable for.
The Ethical Questions That Deserve Honest Engagement
AI in healthcare raises ethical questions that cannot be resolved by better engineering alone. They require policy choices, community consultation, and honest engagement with competing values.
Consent
and Data Privacy
Training medical AI requires access to patient data at scale. The patients whose data trains a diagnostic AI typically did not consent to that specific use when they received care. Regulatory frameworks for medical data use in AI training vary significantly across jurisdictions and are actively evolving. The tension between the public health benefit of better-trained AI systems and the individual privacy rights of patients whose data enables those systems is genuine and unresolved.
The
Automation of Diagnostic Judgment
As AI diagnostic tools become more accurate on specific tasks, there is a genuine risk that clinicians defer to AI recommendations in ways that erode their own diagnostic skills over time. This automation-induced skill degradation has been documented in aviation and other high-stakes domains where automated systems became so reliable that human operators lost the manual competence to intervene effectively when automation failed. Medicine faces the same risk if AI adoption is not managed with deliberate attention to maintaining human clinical judgment.
Access
and the Risk of Deepening Inequality
Advanced
AI diagnostic tools deployed at scale in wealthy healthcare systems could
improve care for patients who already have the best access to medicine. If the
same tools are not accessible or are not adapted to function effectively in
lower-resource healthcare settings, the net effect of AI in global health could
be to widen rather than narrow existing health disparities. This is not
inevitable but it requires deliberate policy choices rather than market-driven
deployment patterns to prevent.
The AI Vanguard Take:
AI in
healthcare has moved from proof-of-concept to genuine clinical deployment in a
remarkably short time. The honest assessment is that it is making meaningful
differences in specific, well-defined tasks while raising serious questions in
areas that require policy engagement rather than engineering solutions. The
appropriate response is neither uncritical enthusiasm for every new application
nor reflexive scepticism about a technology with demonstrated clinical utility.
It is rigorous evaluation, honest reporting of both successes and failures, and
deliberate policy work on the distribution of benefits and risks.
Frequently Asked Questions
Can AI
diagnose diseases better than doctors?
On specific, well-defined diagnostic tasks using medical imaging, AI systems have demonstrated performance that matches or exceeds specialist physicians in controlled study conditions. In broader clinical contexts requiring integration of multiple information sources, patient history, physical examination findings, and clinical judgment, AI is a tool that assists rather than replaces physician decision-making. The framing of AI versus doctors is less useful than asking which specific tasks AI can perform accurately and reliably enough to be safely integrated into clinical workflows.
Should I
trust AI-assisted medical advice?
Consumer AI chatbots including ChatGPT, Claude, and Gemini are not medical devices and should not be used as a substitute for professional medical advice. They can be useful for understanding medical terminology, researching conditions, and formulating questions for a healthcare provider. Any specific medical decision should involve a qualified healthcare professional. The AI tools described in this post that are having genuine clinical impact are purpose-built, validated systems deployed within clinical workflows, not general-purpose chatbots.
How is
AI being used in African healthcare?
AI healthcare applications in Africa span a wide range of contexts. Zipline uses AI-optimised drone delivery networks to transport blood products and medicines to rural facilities across Rwanda, Ghana, and Nigeria. Babylon Health has deployed AI-assisted symptom checking and triage in Rwanda. The community health predictive analytics programme in Kenya described in this post represents one of the more locally adapted applications. The most promising approaches are those built on locally collected data and adapted to local healthcare infrastructure rather than tools developed for wealthy-country contexts and transplanted without adaptation.
Deep Dives on AI and Society: The AI Vanguard covers the broader implications of AI on
healthcare, employment, education, and governance every week. Subscribe below.
