AI in healthcare and medicine: uses, examples and risks

AI in healthcare is the use of models that learn from data to support clinical and administrative work: finding evidence, summarizing charts, reading images, suggesting codes and prioritizing cases. It saves time and adds consistency, but it can be wrong, inherit bias and expose patient data, so it needs validation and clinician oversight.

Editorial update: October 10, 2026. Published by Arkangel AI, a company that builds AI software for healthcare. This article is educational and is not medical advice; clinical decisions remain with licensed professionals.

What is AI in medicine?

"Artificial intelligence" covers several different techniques. In medicine it helps to separate them, because each one fails in its own way:

Type	What it does	Healthcare example
Rule-based systems	Apply rules written by experts	Drug-drug interaction alert
Predictive machine learning	Estimates risk from historical data	Readmission or deterioration risk score
Computer vision	Analyzes images and signals	Findings on chest X-rays, retinal photos or digital pathology
Generative AI and large language models (LLMs)	Reads and writes natural language	Summarizing a chart or answering a question with citations

Generative AI has grown fastest since 2023 and raises the most questions, because it produces fluent text even when the content is wrong. That is why, in clinical use, it matters so much that every claim can be traced to a source you can open.

How is AI used in healthcare?

AI is used today in six main areas. None of them removes the human decision; they change how quickly and how a clinician gets there.

1. Evidence search. A clinician asks a clinical question in plain language and gets a synthesis with cited studies and guidelines. See medical search AI.
2. Clinical decision support. Reminders, alerts, differential diagnoses or guideline summaries at the point of care. We explain it in what is clinical decision support (CDSS).
3. Imaging and signals. Models that flag possible findings on X-rays, CT scans or ECGs for a specialist to review.
4. Documentation. Chart summaries, draft discharge notes and structured data extracted from free text.
5. Coding, billing and audit. ICD-10 code suggestions with supporting text, and inconsistency checks before claims go out. See ICD-10 coding with AI.
6. Research and public health. Cohort identification, pharmacovigilance and real-world data analysis.

Examples of AI in medicine

• A cited clinical answer: "What is first-line management of hypertension in a patient with chronic kidney disease under current guidelines?" A specialized tool returns a summary linked to the guidelines and trials it used; the clinician reads the source before acting on it.
• Extracting data from notes: in work presented at IEEE CAI 2025, Arkangel AI's PANDORA system pulled variables from unstructured clinical notes to compute the PUMA COPD risk score with 94% accuracy (DOI 10.1109/CAI64502.2025.00280).
• Cleared devices: the FDA keeps a public list of AI-enabled medical devices, most of them in radiology.
• Payer audit: in an anonymized Colombian EPS case, Arkangel AI cut the average audit cycle from 94 to 29 days across 18 batches (42,300 line items), with the human auditor making the final decision (read the case).

Benefits: what the evidence shows

Eric Topol's review in Nature Medicine describes AI's potential at three levels: clinicians (faster interpretation of data), health systems (more efficient workflows) and patients (Topol, 2019). It also warns that much of the evidence is retrospective and that prospective trials are needed.

For evidence search, results are promising but should be read with their design in mind:

• In a preliminary randomized pilot (25 participants, conference poster, 2025), clinicians using Arkangel AI answered questions 79% faster with 34% fewer searches. A later peer-reviewed randomized trial with 83 medical students found ~55% less time per case and higher answer-validity scores.
• In Intelligence-Based Medicine (2025), Arkangel AI's agent reached 90.26% accuracy on MedQA-style questions (DOI 10.1016/j.ibmed.2025.100274). A multiple-choice exam is not the same as real-world clinical performance.

Conflict of interest: the Arkangel AI studies cited here were run by Arkangel's own team. You can review them on our research page.

What are the risks of AI in healthcare?

These are the seven risks that matter most in clinical settings:

1. Errors and hallucinations. A language model can invent facts or references with complete confidence. In one study of ChatGPT-3.5, 47% of generated references were fabricated and 46% were real but inaccurate (Bhattacharyya et al., 2023).
2. Biased data. A widely used U.S. algorithm underestimated the needs of Black patients because it used healthcare spending as a proxy for illness (Obermeyer et al., 2019).
3. Automation bias. Clinicians tend to accept automated suggestions even when they are wrong (Goddard et al., 2012).
4. Privacy and security. Pasting identifiable patient information into a consumer chatbot can breach HIPAA or local privacy law.
5. No local validation. A model validated on one population can perform worse with different case mixes, languages or record systems.
6. Unclear accountability. If AI suggests and the clinician decides, teams must define who reviews, how it is documented and how errors are reported.
7. Unequal access. English-only tools, or tools restricted to some countries, widen gaps between health systems.

The same practices reduce all seven: verifiable sources, validation in the population where the tool will be used, access controls, audit trails and professional review of every output.

What does the WHO say about AI in health?

In 2021 the World Health Organization published Ethics and governance of artificial intelligence for health, built on six principles:

1. Protect human autonomy.
2. Promote human well-being, safety and the public interest.
3. Ensure transparency, explainability and intelligibility.
4. Foster responsibility and accountability.
5. Ensure inclusiveness and equity.
6. Promote AI that is responsive and sustainable.

In January 2024 it added guidance on large multi-modal models, the family that includes ChatGPT, with recommendations for governments, developers and providers: independent evaluation, transparency about training data and human oversight in clinical use.

Regulation: what is changing

• United States. The FDA regulates software that diagnoses or treats as a medical device. In January 2026 it updated its Clinical Decision Support Software guidance, which keeps four criteria separating software that supports a clinician from software regulated as a device. HIPAA governs how covered entities handle protected health information.
• European Union. The AI Act, Regulation (EU) 2024/1689, classifies many AI systems used in medical devices as high-risk.
• Latin America. Health agencies assess software by its intended use under each country's medical-device rules, and several governments are drafting national AI frameworks.

How to choose and use a clinical AI tool

• Visible sources: every answer should link to studies or guidelines you can open.
• Demonstrable security: ask for certifications with scope and validity dates, not just logos.
• Patient data: do not enter identifiable information into tools that are not contractually set up for it.
• Language and context: the tool should handle your language and cite guidelines relevant to where you practice.
• Published evidence: prefer tools evaluated in studies with declared design and conflicts of interest.

Arkangel AI answers clinical questions with traceable citations in English, Spanish, Portuguese and French. It is available on the web and on WhatsApp. It is ISO 27001:2022 certified, with HIPAA compliance and a BAA on Enterprise plans, 70+ security controls, penetration testing and continuous monitoring via Vanta (Trust Center). The free plan has limited searches; paid plans (Pro from $20/month) raise usage limits (pricing). Do not enter PHI on the free or Pro plans. To compare options, read best AI tools for doctors in 2026 or see AI for doctors.

Will AI replace doctors?

No. AI speeds up information search, document review and some image-reading tasks, but the physical exam, integrating a patient's context, communication and accountability for the decision remain human. The more likely change is in how doctors spend their time: less searching and typing, more verifying and deciding with the patient.

Frequently asked questions

What does the WHO say about AI in health?

The WHO asks that AI in health protect autonomy and privacy, be transparent, safe and overseen by humans, promote equity and be evaluated in real settings. Since 2024 it also has specific guidance for large language and multi-modal models, with recommendations for governments, developers and healthcare providers.

What are the main risks of AI in medicine?

The main risks are errors or fabricated references, biased training data, over-reliance on automation, privacy breaches, lack of local validation, unclear accountability and unequal access. They are reduced by verifiable sources, validation in the real population, security controls, audit trails and professional review of every output before it affects care.

Is it safe to use AI for medical diagnosis?

AI is safe as support when it is validated for that use, protects data and a clinician reviews its output; it should not diagnose on its own. The WHO and FDA both stress human oversight and transparency. Use tools that show their sources, and always document the final clinical decision yourself.

Is there free medical AI?

Yes, with limits. Some tools are free for verified clinicians or offer plans with a capped number of queries. Arkangel AI has a free plan with limited searches and paid plans that raise usage. Before choosing, compare cited sources, languages, privacy terms and how each vendor handles your data.

Sources

• WHO. Ethics and governance of artificial intelligence for health (2021) and guidance on large multi-modal models (2024).
• FDA. Artificial Intelligence-Enabled Medical Devices and Clinical Decision Support Software.
• European Union. Regulation (EU) 2024/1689.
• Topol EJ. High-performance medicine. Nat Med 2019.
• Bhattacharyya M et al. Fabricated and inaccurate references in ChatGPT-generated medical content. Cureus 2023.
• Obermeyer Z et al. Dissecting racial bias in an algorithm. Science 2019.
• Goddard K et al. Automation bias: a systematic review. JAMIA 2012.
• Arkangel AI. Mayo Clinic Proceedings: Digital Health, 2026 (authors are Arkangel staff) and Intelligence-Based Medicine, 2025.