AI for Medical Education and Research

How to use language models without outsourcing critical thinking

AI can explain a difficult concept, generate teaching cases, summarise papers and screen hundreds of abstracts before lunch. It can also invent a reference, flatten uncertainty and produce a polished explanation that is confidently wrong.

The important question is not whether clinicians should use AI but rather how to use these tools without confusing fluent language with reliable evidence.

This article distils the practical themes from the Medmastery webinar AI for Teaching and Research

What does a language model actually do?

A large language model (LLM) does not ordinarily retrieve a single stored answer from a database. It generates a sequence of tokens based on patterns learned from large collections of text and on the instructions and context supplied by the user.

This allows it to explain, transform and combine information with fluency. However there is one important limitation as the model may generate an answer that sounds plausible even when the underlying claim is unsupported.

The prose cannot validate itself. Confidence, detail and technical language are features of the response but not evidence that it is correct.

Some AI systems can also search the web, retrieve documents or use specialised tools. These capabilities may improve the evidence available to the model, but they do not remove the need to verify how the evidence has been selected and interpreted.

Prompting is specification, not magic

Good prompting is less about finding a secret phrase and more about defining the task clearly.

A useful prompt should state:

• what you want done;
• who the output is for;
• what source material may be used;
• the required format and level of detail;
• what should be excluded;
• how uncertainty and disagreement should be presented.

For example:

Using only the attached guideline, explain the indications and contraindications for thrombolysis in acute ischaemic stroke to emergency medicine registrars. Distinguish strong recommendations from areas of uncertainty and cite the relevant guideline section after each statement.

This is more dependable than simply asking the model to “act as an expert”. A role can help shape tone, but it cannot create expertise or compensate for missing evidence.

Agreement is not verification

Asking several AI systems the same question may reveal areas of disagreement. It does not confirm that a shared answer is correct.

Models can reproduce the same widely repeated misconception, rely on overlapping source material or generate similar citation errors. Consensus among models is a reason to inspect the evidence, not a substitute for doing so.

Medical claims should be checked against the original source:

• the guideline rather than a summary of the guideline;
• the paper rather than an AI-generated description of the paper;
• the dataset rather than a claim about the dataset;
• the regulatory document rather than marketing material.

References should be opened and read. A citation may exist but fail to support the sentence attached to it. Or the citation may not even exist…

Literature searching with AI assistance

Research-focused AI tools can help clinicians discover papers, identify related studies, map citation networks and produce summaries of a field. Tools linked to indexed scientific literature may provide a better starting point than a general web search.

However, the AI output still requires scrutiny.

An inline citation does not guarantee that the paper has been represented accurately. The tool may misunderstand the population, outcome, comparator or effect. It may also highlight easily indexed or frequently cited work while missing older, negative or non-English studies.

AI usually accelerates literature discovery but does not, by itself, perform a systematic review. The literature review still requires critical appraisal and synthesis.

AI as a teaching assistant

Used carefully, AI can expand what a medical educator can produce and can be used to:

• generate alternative versions of a case;
• simulate a patient or examiner;
• convert material into questions or flashcards;
• explain a concept at multiple levels of complexity;
• compare diagnostic approaches;
• provide formative feedback;
• identify possible gaps in an argument.

It is the educator who remains responsible for the accuracy of the material. Generated cases may contain contradictions, inappropriate assumptions or implausible clinical details. User feedback may result in a more fluent but still unsafe or inaccurate answer.

Avoiding cognitive debt

Clinical expertise develops through repeated acts of interpretation, pattern recognition, constructing differential diagnoses and recognising when the available evidence does not fit.

If AI performs those tasks before the clinician has attempted them then efficiency may come at the cost of deskilling the practitioner. The user is less able to recognise when the AI is wrong because the reasoning required to detect the error has been removed.

Probably the best safeguard is to reason first and consult AI second. In this way AI can extend reasoning rather than replace it. Best education practice would be to ask the clinician to:

1. commit to an initial interpretation;
2. identify the uncertainty;
3. review the AI response;
4. compare the two;
5. verify disputed claims against an authoritative source.

AI in research

AI can already assist with repetitive research tasks such as deduplication, abstract screening, data extraction and language editing. This allows researchers to spend more time on study design and interpretation.

However this enhanced efficiency does not remove overall accountability. Researchers need to record:

• what the system was asked to do;
• which model and version were used;
• how outputs were checked;
• whether confidential data were exposed;
• whether errors could be reproduced or audited;
• disclosure of AI assistance.

The task of the human researcher is not only to produce questions, but then to decide which questions are worth answering and what evidence would count as a meaningful answer.

Privacy, attribution and governance

Patient-identifiable information, unpublished manuscripts and confidential datasets should not be entered into an AI system unless its use has been approved and the AI data-handling arrangements are understood.

Clinicians and researchers should also follow local policies on attribution, authorship and disclosure. AI cannot accept responsibility for an error, defend a methodological choice or meet the criteria for authorship. Those obligations remain human.

Bottom line

AI can make medical teaching and research faster but can also make weak reasoning look polished.

Ai is best used to generate possibilities, reorganise information and reduce repetitive work. AI should not be used as an unexamined source of truth. Researchers and educators need to specify the task clearly, preserve independent reasoning, and check claims against original sources.

---

References

• Wiesbauer F. AI for Teaching and Research  Medmastery Free Webinar
• Wiesbauer F. Trustworthy AI in Cardiology LITFL
• Elshahati N. AI Scribes for Clinical Practice. LITFL
• Wiesbauer F. ChatGPT Essentials for Clinicians. Medmastery
• Guilleminot S. AI in Healthcare. LITFL