Guide
AI in medicine: what it can and can't do for medical research
AI in medicine covers regulated tools that help read scans or signals, and general tools that find, summarise and explain research. For medical research it can speed up finding and reading papers, but it can still invent or misquote citations and overstate what a study shows. It does not replace clinical judgement, and research tools are not diagnostic devices.
This is information about reading evidence, not medical advice. Talk to your doctor or pharmacist before changing anything.
Last checked: 21 October 2026. We make MIA. Other tools are described from their public sites and university library guides; check each before relying on this page.
Two different kinds of "AI in medicine"
It helps to separate them, because the rules and the risks differ.
- AI as a medical device. Software that analyses images, signals or patient data to support diagnosis or treatment. In the US, the FDA keeps a public list of AI-enabled medical devices "authorized for marketing in the United States," updated periodically. These tools go through regulatory review for a defined use.
- AI for information and research. Chatbots, literature search engines and summarisers. These help people find and read evidence. Most are not marketed as medical devices, and their outputs need checking.
Eric Topol's 2019 review in Nature Medicine described AI's impact at three levels: for clinicians, "predominantly via rapid, accurate image interpretation"; for health systems, through workflow and fewer errors; and for patients, by helping them use their own data. He listed the limits as "bias, privacy and security, and lack of transparency."
What AI can do well for medical research
- Find papers faster. Semantic search can surface studies that keyword search misses.
- Screen and extract. Tools built for evidence synthesis can help sort large numbers of abstracts and pull data into tables.
- Summarise and explain. A good summary can make a dense paper approachable, as long as it points back to the passage it relies on.
- Answer exam-style questions. In 2023, Singhal and colleagues (Nature) reported that Flan-PaLM reached 67.6% accuracy on MedQA, a set of US licensing exam-style questions. They also reported that their tuned model, Med-PaLM, "remains inferior to clinicians" on human evaluation. Exam scores are not clinical performance.
Where it fails
- Invented and inaccurate citations. Bhattacharyya and colleagues (Cureus, 2023) asked ChatGPT-3.5 to write 30 short medical papers. Of 115 references, 47% were fabricated, 46% were real but inaccurate, and only 7% were real and accurate. That study tested one model version in 2023, so the rate will differ for other tools. The lesson holds: every citation needs checking.
- Overreach. An association read as a cause, a subgroup read as everyone, a conference abstract read as a result.
- Flattening the evidence. A phase I signal and a guideline recommendation can come out sounding equally confident.
- Privacy. Pasting patient details into a general chatbot can breach privacy obligations. Check your organisation's policy first.
The World Health Organization has published guidance on the ethics and governance of large multi-modal models in health, covering these risks and the responsibilities of developers, providers and governments.
The main types of research tool (2026)
| Job | Example tools | What they are built for | Watch for |
|---|---|---|---|
| Searching and scoping a question | Consensus | An academic search engine with cited summaries; a "Consensus Meter" on yes/no questions | Summaries of abstracts are not the full paper |
| Screening and extracting for a review | Elicit | Evidence synthesis: literature screening and structured extraction into tables | Extraction still needs a human check |
| Point-of-care clinical questions | OpenEvidence | Cited clinical answers for verified healthcare professionals (per university library guides; site not checked directly) | Eligibility; it is a clinician's tool |
| General questions | General chatbots (e.g. ChatGPT) | Broad explanation and drafting | Fabricated or inaccurate references |
| Reading one paper closely | MIA | Testing what a paper or exhibit you bring will bear: finding, inference and gap kept apart | It does not search the literature |
These tools are not rivals for every job. A common pattern in library guides is to scope with one tool, review with another, then read the key papers closely.
Can AI replace doctors?
Not on the evidence available. Benchmarks measure narrow tasks under test conditions. Clinical care involves history, examination, uncertainty, values and responsibility for the decision. The research cited above reports useful performance on defined tasks alongside clear gaps against clinicians. The realistic near-term role is support: faster reading, better-organised evidence, and regulated tools for specific, validated uses.
AI diagnosis: why MIA doesn't do it
"AI diagnosis" is a common search, and some regulated devices do support diagnosis for specific, approved uses. MIA is not one of them, and it is built not to act like one.
- MIA does not diagnose. It is an educational literature desk, not a clinician.
- It does not recommend treatment or give dosing, schedules or implementation parameters.
- Imaging and pathology are read only as supplied, in educational pattern-language. That is not a radiology report, and a finding is not a cause.
We draw the line there because a confident wrong answer is the most dangerous output a medical tool can give. If you are worried about a diagnosis, talk to your doctor, and ask about a second opinion if you want one.
Where MIA fits
MIA (Medical Intel) is for the paper in front of you. Put a paper, a scan or an exhibit on the desk, or connect Google Drive or Dropbox, and ask what it will bear. The working separates what the file shows, what may reasonably be drawn from it, and what it does not establish. Clinical Conservative is the default: standard of care is kept apart from investigational claims and thin evidence is named as thin. Research Expansion, for hypothesis work, always opens with a banner saying it is not standard of care. Deep Review runs two independent workings and settles them: agree, contest, or still a gap.
Try MIA on a paper you rely on: 150 free credits, no card.
FAQ
How is AI used in medicine? Two main ways: regulated AI medical devices that help analyse images, signals or patient data for specific uses, and information tools that help find, summarise and explain research. The second kind needs its outputs checked.
Can ChatGPT be used by doctors for research? It can help explain and draft, but studies have found high rates of fabricated or inaccurate references in its medical content. Check every citation against the source, and do not enter identifiable patient information unless your organisation allows it.
Can AI replace doctors? Not on current evidence. AI performs well on some narrow tasks and exam-style questions, but studies still report gaps compared with clinicians, and care involves judgement and responsibility that tools do not carry.
Does MIA diagnose? No. MIA is an educational literature desk. It does not diagnose, recommend treatment or give dosing. It helps you read the papers and files you bring.
This is information about reading evidence, not medical advice. Talk to your doctor or pharmacist before changing anything.
Sources
- Topol EJ. High-performance medicine: the convergence of human and artificial intelligence. Nat Med. 2019;25(1):44-56. doi:10.1038/s41591-018-0300-7 - Singhal K, Azizi S, Tu T, et al. Large language models encode clinical knowledge. Nature. 2023;620(7972):172-180. doi:10.1038/s41586-023-06291-2 - Bhattacharyya M, Miller VM, Bhattacharyya D, Miller LE. High rates of fabricated and inaccurate references in ChatGPT-generated medical content. Cureus. 2023;15(5):e39238. doi:10.7759/cureus.39238 - US Food and Drug Administration. Artificial Intelligence-Enabled Medical Devices. https://www.fda.gov/medical-devices/software-medical-device-samd/artificial-intelligence-enabled-medical-devices - World Health Organization. Ethics and governance of artificial intelligence for health: Guidance on large multi-modal models. https://www.who.int/publications/i/item/9789240084759 - Vendor sites: elicit.com, consensus.app (re-checked 20 Oct 2026); openevidence.com (blocked automated access; description from university library guides).