Prompting as Clinical Reasoning: What We Learned from the First DR. INFO x ZAKA Webinar

On 29 June, DR. INFO and ZAKA hosted their first English-language webinar on clinical AI, bringing together healthcare professionals to discuss a question many clinicians are already asking: how do you use AI without compromising clinical judgment?

Across one hour, Dr. Mohamad Diab, a medical doctor and AI instructor at ZAKA, and Dr. Nader Absi, a physician working on clinical AI deployment and governance, explored where AI is already changing clinical workflows, where it still falls short, and why better prompting is less about learning AI tricks and more about thinking like a clinician. Here are the biggest takeaways from the session.

AI's greatest strength is also its greatest risk

The webinar opened with a simple but important observation. Large language models produce answers that are fluent, well-structured, and confident. That makes them incredibly useful, but it also makes them easy to trust when they are wrong.

An incorrect recommendation rarely looks incorrect. It often uses the right medical terminology, sounds evidence-based, and may even include citations. The problem is that confidence is not evidence. As Dr. Mohamad Diab put it, the question is no longer whether clinicians will use AI. That transition is already happening. The real question is how to use it safely.

AI is becoming part of everyday clinical work

One reason adoption has accelerated is simple: clinicians are overwhelmed. Physicians spend less than a third of their working day with patients, while documentation, administrative work, and electronic health records consume far more time. AI is increasingly used to reduce that burden, not by replacing clinical reasoning, but by handling repetitive cognitive tasks.

Survey data discussed during the webinar showed physician AI adoption rising quickly over the last few years. Most current use cases involve documentation, summaries, literature retrieval, and clinical support rather than diagnosis itself. That is an important distinction. Despite the headlines, today's clinical AI is primarily helping clinicians work more efficiently, not making decisions independently.

Hallucinations have not disappeared, but they are becoming rarer

Understanding how language models work also explains why hallucinations occur. Rather than retrieving facts from memory the way humans do, large language models generate text by predicting what comes next based on patterns learned during training. Most of the time this produces useful responses. Sometimes it produces information that sounds convincing without being grounded in evidence.

In medicine, that can appear as fabricated citations, incorrect guideline recommendations, inaccurate patient summaries, or omitted medications and risk factors. Another concern is automation bias. People naturally place more trust in recommendations produced by technology, even when those recommendations are incorrect. The speakers highlighted research showing that biased AI recommendations can reduce clinician accuracy, even when the system explains its reasoning.

The encouraging news is that modern models are improving quickly. Compared with earlier generations, hallucination rates have fallen substantially, benchmark scores continue to improve, and evidence-grounded systems perform far better than models relying only on internal knowledge. Even so, the conclusion stayed the same: every clinical recommendation still requires verification.

Prompting is not prompt engineering. It is clinical reasoning.

The second half of the webinar shifted from understanding AI to working with it. Dr. Nader Absi challenged one of the most common assumptions about prompting: more context does not necessarily produce better answers. In medicine, patient records are full of information. Copying entire charts into an AI system often introduces noise instead of improving reasoning.

Experienced clinicians do not present every laboratory value when requesting a consult. They identify the details that change management. AI works the same way. The goal is not writing longer prompts. It is writing better clinical handoffs.

Three cases showed the same lesson

Rather than discussing prompting in theory, the webinar focused on three practical examples.

Case 1: The metformin patient

A 60-year-old woman with type 2 diabetes had an eGFR of 36 while taking metformin. Viewed in isolation, the recommendation was straightforward: monitor kidney function and consider a dose adjustment. Then one additional detail was introduced. Six weeks earlier, her eGFR had been 75. She had experienced two days of vomiting, dehydration, and poor oral intake.

The picture immediately changed. Instead of chronic kidney disease, this became a possible acute kidney injury during an acute illness, and the safer recommendation became temporarily holding metformin until renal function recovered. The medication had not changed. The prompt had.

Case 2: Atrial fibrillation, and one missing sentence

The live DR. INFO demonstration followed the same pattern. A 67-year-old man with newly diagnosed atrial fibrillation appeared to be an appropriate candidate for a direct oral anticoagulant, and the recommendation looked entirely reasonable. Then one missing detail was added: the patient had a mechanical mitral valve.

That single sentence completely changed management. Warfarin became the recommended therapy, while direct oral anticoagulants became contraindicated. The AI was not inconsistent. It simply answered the clinical picture it had been given.

Case 3: When the culture is not enough

The final example came from primary care. A patient with a urinary tract infection had a culture showing susceptibility to trimethoprim-sulfamethoxazole. Based only on the culture, the recommendation seemed obvious. Then another piece of information appeared: the patient was taking weekly methotrexate.

Suddenly the safest antibiotic was no longer the most appropriate choice, because of the risk of severe myelosuppression. The culture identified what worked. The medication list identified what was safe. Both were necessary to reach the correct decision.

Four practical rules for prompting

Across all three cases, the same principles emerged.

• Treat the prompt like a clinical consult. Frame the patient the way you would brief a colleague, not by copying the chart, but by highlighting what matters.
• Remember that AI answers the information you provide. The model cannot reason about details that never appear in the prompt.
• Focus on high-impact variables. Changes in trajectory, important comorbidities, medication interactions, pregnancy, red flags, or previous diagnoses often determine management far more than dozens of routine details.
• Prioritize signal over volume. Five pages of chart notes rarely improve an answer if the one critical fact is buried inside them.

Better prompts do not replace clinical judgment

The webinar closed with an important reminder. Even excellent prompts produce only a starting point. AI does not have bedside intuition. It does not understand local prescribing policies, formulary restrictions, or institutional practice unless that information is explicitly available. Most importantly, it carries no clinical responsibility. The recommendation belongs to the model. The final decision belongs to the clinician.

A practical habit suggested during the session was to review every AI-generated recommendation through three questions: does it fit this patient, not just the diagnosis; can you trace it back to a guideline or evidence source; and does the response acknowledge uncertainty where uncertainty exists. If the answer to any of these is no, the recommendation deserves closer scrutiny.

Questions from the audience

The live Q&A focused on a concern shared by many clinicians: if AI can hallucinate, how much should we trust it? The speakers emphasized that there is no shortcut for identifying hallucinations. The safest approach is to verify the underlying evidence, make sure a recommendation fits the patient's clinical context, and treat AI as a decision-support tool rather than a decision-maker.

Another thread centered on prompting itself. Would providing the entire patient chart produce better answers? Not necessarily. As Dr. Absi explained, more information often introduces more noise. The goal is to include the details that change management, not every detail available.

Finally, attendees asked whether prompt engineering will remain an essential skill. The consensus was that prompting will become more intuitive as models improve, but clinical reasoning and critical appraisal will remain irreplaceable.

Looking ahead

This webinar focused on improving the input: asking better clinical questions. The next session will focus on the output: learning to identify responses that sound polished and authoritative but contain subtle clinical errors. Because in medicine, the challenge is not simply getting an answer. It is knowing whether that answer deserves your trust.