Should Doctors Use AI? My Honest Opinion
I have been practicing and teaching pediatrics for years. I have watched medicine change with every decade — but nothing has moved as fast, or stirred as much debate in our corridors, as artificial intelligence. This article is not a technology glossary or a vendor brochure. It is my honest attempt to tell you, as one clinician to another, what the data shows, what I have personally experienced, and how I think about AI as a working tool — not a replacement — for doctors in 2025 and beyond.
The Numbers: Where Physician AI Adoption Stands Right Now
Let us start with the most striking statistic. When Doximity — with over 3 million member physicians on its platform — surveyed doctors in late 2025 through early 2026 and asked whether they use or are interested in using AI, the answer was 94%. That is not a futurist projection. That is where we are today.
To understand how rapid this shift has been: in 2023, only 38% of physicians reported using AI. By the end of 2024 that had risen to 66% — a 78% jump in a single year, according to the American Medical Association. By early 2026, 63% report active regular use. This is a pace of adoption comparable to CTs and MRIs in the 1980s — technologies that are now simply infrastructure.
Neurologists led AI adoption at 64%, followed by gastroenterologists (61%) and internists (60%). But the most frequent users? Family medicine — 88% of adopters in that specialty use AI every single day.
The "Pajama Time" Problem — And Why It Is Driving Adoption
If you want to understand why AI adoption in medicine has accelerated so dramatically, you have to understand burnout. In 2025, more than half of all US physicians reported feeling burned out. And the single biggest driver is not the clinical work — it is the documentation.
Doctors were spending two to three hours on EHR entry and paperwork for every single hour of direct patient care. The medical community has a term for this: pajama time — the charting you finish at home, late at night, when you should be resting. CMS estimates that 77% of healthcare workers waste at least 45 minutes per day on inefficient administrative workflows — the equivalent of four weeks lost annually per physician.
The Doximity 2026 survey found that 90% of physicians believe AI can reduce pajama time, and current AI users estimate it could cut their after-hours charting nearly in half. Among AI users, 75% report improved job satisfaction and 69% say it has contributed to better patient outcomes.
Where AI Is Being Used Right Now — By Clinical Domain
Physicians are not using AI for one thing. They are integrating it across clinical and administrative workflows. Explore each area below:
AI in Clinical Practice — Explore by Domain
Ambient documentation is the single fastest-growing AI use case in medicine. A JAMIA survey of 43 major US health systems found that AI note-writing was the only tool where all 43 systems had adopted it to at least some degree — making it uniquely universal. Tools like Abridge, Suki, and Epic's built-in scribe listen to the patient encounter and generate a structured clinical note automatically. Voice-based documentation use jumped from 20% to 29% of physicians in less than a year. You review the output; the machine handles the typing. For a clinician seeing 30–40 patients a day, this is transformative.
Radiology remains the dominant area for FDA-approved AI, accounting for approximately 76% of all cleared AI medical devices. These tools assist radiologists in detecting abnormalities in X-rays, CT scans, MRIs, and pathology slides — often flagging urgent findings that need immediate action and reducing reporting turnaround times. For paediatricians ordering chest X-rays for pneumonia or abdominal scans for suspected appendicitis, AI triage flags can serve as a genuine safety net, especially in high-volume emergency settings.
AI as a diagnostic second opinion is increasingly studied and deployed. A JAMA study from October 2024 showed that AI alone scored 90% accuracy on certain diagnostic tasks, while physicians using AI improved from 74% to 76%. This does not mean AI replaces clinical judgement — it means we are learning to collaborate. The top diagnostic use case today is literature search: 35% of physicians now use AI to rapidly scan and synthesise clinical evidence, up from 22% just months earlier. For rare presentations or unfamiliar syndromes, this is genuinely valuable — having a tool that can surface relevant research in seconds rather than hours changes how you approach an uncertain case.
About 10% of FDA-approved AI devices target cardiovascular medicine. These include ECG interpretation tools, AI-driven echocardiogram analysis to detect heart failure patterns, and models that predict cardiac events from EHR data. In paediatric practice, AI-assisted ECG interpretation is particularly useful — subtle rhythm abnormalities in children can be easy to overlook during a busy OPD, and an AI flag can prompt a second careful look before a child is discharged. It does not replace your read; it prompts you to look again.
One of the most compelling recent case studies comes from Cleveland Clinic. In September 2025, the clinic expanded its AI-driven sepsis detection system following a pilot that showed it generated 10 times fewer false alarms than the previous approach, identified 46% more sepsis cases, and provided advance warnings before antibiotics were needed seven times more often. This is significant because alert fatigue — where clinicians start ignoring alarms because so many are false positives — is a genuine patient safety hazard. Cutting false alerts by a factor of 10 means that when this AI flags a patient, staff take it seriously. That is the difference between a tool that gets used and one that gets ignored.
Paediatric-specific AI applications are growing, but the validation gap is real. A 2025 analysis of FDA-approved AI devices found that fewer than one-quarter of clinical evaluations provided age-specific performance data. Many tools validated in adults may not perform equivalently in children — and as paediatricians, we must ask vendors directly: "What is the performance data for paediatric populations?" AI applications for growth monitoring, developmental screening support, and neonatal imaging are emerging, but still require more robust paediatric validation before we should place significant clinical weight on their outputs.
How Doctors Are Actually Using AI: 2026 Data
The Doximity 2026 survey of 3,151 US physicians across 15 specialties gives us the clearest picture yet of real-world use. These are not theoretical applications — this is what doctors are doing today:
Source: Doximity 2026 State of AI in Medicine Report
A Brief History of a Very Fast Decade
Context matters. AI in medicine did not arrive overnight, but the pace of the last three years has been without precedent:
How to Start: A Practical Guide for Clinicians
Step 1: Begin with documentation
Documentation AI is the lowest-risk, highest-reward entry point. These tools are non-diagnostic — they assist with a process you already do, and you remain the reviewer. Start with whatever is already embedded in your EHR (Epic, Athenahealth, and Oracle Cerner all have options), or explore standalone tools like Abridge or Suki. The learning curve is minimal; the time savings begin almost immediately.
Step 2: Use AI for literature search
The most common AI use among physicians today is literature search — not diagnosis. Before your next complex case discussion, try asking an AI tool to summarise the latest guidance on a specific clinical question or surface key studies. Treat it as a research assistant, not an oracle. Always verify sources. This is a safe, bounded, genuinely useful workflow to build your AI confidence.
Step 3: Use AI for patient communication
Generating discharge summaries, care plan explanations, or referral letters in plain language is tedious work that AI does well. You provide the clinical content; AI helps with language, structure, and readability. Some tools now offer real-time translation, which is particularly valuable for communicating with caregivers from diverse backgrounds.
Step 4: Build your AI literacy
Only 28% of physicians feel adequately prepared to use AI — despite 57% already using it. That gap matters. The AMA offers seven free CME modules on AI in healthcare on its Ed Hub platform (worth 3.5 AMA PRA Category 1 Credits). Harvard Medical School's AI in Clinical Medicine programme is also available. A few hours of structured learning will make your everyday use of these tools significantly more effective and safe.
Quick-Start Checklist for Busy Clinicians
- Identify your single biggest daily time drain — charting, letters, or literature review
- Check whether your EHR already has embedded AI features (most do as of 2025)
- Confirm any tool you use is HIPAA-compliant and audited for data security
- Complete at least one AMA or institution-based AI CME module before deploying tools
- Always review AI-generated notes — never copy-paste without reading them
- Ask vendors: "What is the validated performance data for paediatric populations?"
The Limitations We Cannot Ignore
I would be doing you a disservice if I stopped at the benefits. Any honest discussion of clinical AI must grapple with its current limitations. These are not hypothetical concerns — they are documented in peer-reviewed literature and real-world deployments.
What Is Coming Next: AI in Medicine 2026 and Beyond
EHR Integration at Scale
Epic — used by roughly 40% of US hospitals — announced a major AI overhaul in mid-2025 embedding ambient scribes powered by Microsoft, a "doctor co-pilot" that suggests order sets in real time, a scheduling agent, and a patient-facing chatbot directly into its existing platform. This is the moment AI stops being a separate add-on and becomes clinical infrastructure. For the majority of hospital-based physicians, AI tools will increasingly simply appear in their existing workflow.
Predictive Analytics and Early Warning Systems
The Cleveland Clinic sepsis AI story is the template for what is coming across many disease domains. Early warning systems tuned for specific populations, specific ward contexts, and specific disease trajectories — with dramatically improved false-positive management — are being developed and deployed rapidly. The goal: a warning system clinicians actually trust enough to act on, every time.
Personalised Treatment in Paediatrics
One of the most exciting frontiers for paediatric medicine is AI-assisted personalisation — using a child's growth trajectory, pharmacogenomic profile, and prior treatment response to optimise therapy in ways that rigid weight-based tables cannot. This is still early-stage, but the investment and research activity are accelerating. Within five years, AI-guided dosing adjustments in paediatric oncology and chronic disease management may become standard of care.
AI Literacy as a Core Medical Competency
The students I teach today will practice in a world where AI is routine clinical infrastructure. Incorporating AI literacy — understanding what these models are built to do, where they fail, and how to remain in command of the clinical decision — into undergraduate and postgraduate training is no longer optional. Harvard Medical School has introduced formal AI in clinical medicine courses. Every medical college will need to follow.
Healthcare AI is projected to generate between $100 billion and $600 billion in global savings by 2050, primarily through improved drug discovery timelines and administrative efficiency. But the near-term gains — reduced burnout, better documentation, faster literature synthesis, fewer missed sepsis cases — are already measurable, right now, in the clinics and hospitals already using these tools.
My Personal Approach — What Has Actually Changed in My Practice
I started using an AI documentation assistant about a year ago, initially with significant scepticism. My instinct was that reviewing AI-generated notes would add time, not save it. I was wrong. Once I learned how the system structured notes — and retrained myself to review efficiently rather than rewrite — I recovered roughly 45 minutes each clinical day. That time now goes to complex cases and teaching.
For diagnostic support, my approach is deliberate and bounded. I use AI for literature synthesis when I encounter unusual presentations. I use AI to cross-check drug dosing and interaction queries. I do not use AI as a primary diagnostic tool, and I never accept an AI suggestion without verifying it against my own clinical assessment and, where relevant, specialist consultation.
What I tell my residents is this: the doctors who will flourish in the next decade are not those who distrust AI entirely, nor those who defer to it uncritically. They are the ones who develop genuine AI literacy — understanding what these tools are built to do, where they fail, and how to stay in command of the clinical decision.
Final Thoughts
The 2025–2026 data is unambiguous: AI has moved from the periphery to the centre of clinical practice faster than almost any technology in modern medicine. Ninety-four percent of physicians are using or actively interested in AI. Ambient documentation is now universal in major health systems. Clinically validated AI tools are reducing sepsis deaths, improving radiologist accuracy, and giving doctors back hours they had lost to paperwork.
And yet the core of what we do — listening to a frightened parent describe their child's symptoms, making a judgement call in a complex case, holding a patient's trust — is not something AI can replicate or should attempt to replace. Our task is to use these tools to protect and expand that human core, not to feel threatened by machines that assist us.
Start small. Stay curious. Get trained. Ask hard questions about validation data. And always remember: the final accountability — the final clinical decision — rests with you.
About the Author: This article reflects personal clinical experience in paediatric medicine and current peer-reviewed evidence. All statistics are sourced from major medical associations, FDA databases, and peer-reviewed journals published between 2024–2026.
References & Sources:
- Doximity — 2026 State of AI in Medicine Report (March 2026)
- American Medical Association — Augmented Intelligence in Medicine Survey (October 2024)
- FDA AI-Enabled Medical Devices Database (August 2024 – July 2025)
- JAMIA — Survey of 43 US Health Systems on AI Adoption (2025)
- JAMA Network Open — LLM Diagnostic Accuracy Study (October 2024)
- JAMA Network Open — FDA-Approved AI Device Generalizability Study (April 2025)
- Cleveland Clinic — AI Sepsis Detection System Rollout (September 2025)
- Harvard Medical School — AI in Clinical Medicine Programme (2025)
- Epic Systems — AI Platform Overhaul Announcement (August 2025)
- Vention Teams / CNBC — Healthcare AI Statistics Report (2025)