AI Disclosure in Medical Research: What Singapore Hospital Teams Need to Know in 2026

A recent correspondence in JAMA highlights a critical gap: the scientific community has not yet established what "meaningful" AI disclosure looks like in medical research [2]. As Singapore hospitals accelerate clinical AI deployment and research output, this ambiguity creates compliance risk for institutional review boards (IRBs), research teams, and journal submission workflows. We've seen this confusion firsthand when advising hospital clusters on AI governance frameworks — researchers often don't know what to disclose, when, or how.

This post is for clinical researchers, hospital IRB administrators, research coordinators, and clinical informatics teams in Singapore navigating AI disclosure requirements in 2026.

Key takeaways

JAMA Network journals now track self-reported AI use, but correspondence published July 2026 questions whether disclosure standards are meaningful [2]
Singapore hospital researchers face disclosure ambiguity across three domains: AI-assisted writing, AI-driven analysis, and AI as the research subject
IRBs need operational definitions distinguishing grammar tools from generative content creation, and statistical software from black-box prediction models
Preprint servers show active methods development in clinical AI (ICU prediction architectures [7], synthetic medical imaging [9]), but peer-reviewed disclosure standards lag behind technical capability
Practical disclosure frameworks should specify tool names, versions, prompts (for generative AI), and human review steps — not just binary "AI used: yes/no" checkboxes

Why AI disclosure standards matter for Singapore hospital research

Singapore's research output is internationally visible. Papers from our hospital clusters appear in high-impact journals, inform clinical guidelines, and shape regional health policy. When AI tools contribute to that research — whether drafting methods sections, generating synthetic training data, or predicting patient outcomes — readers need to assess validity, reproducibility, and bias risk.

The July 2026 JAMA correspondence raises a foundational question: "has the scientific community established what meaningful AI disclosure entails?" [2]. The answer, based on current journal policies, is no. Most journals require some disclosure, but lack operational definitions. This creates three problems for Singapore hospital teams:

Compliance uncertainty: Researchers don't know if using Grammarly requires disclosure, or if ChatGPT for literature summarization crosses a line. IRBs receive inconsistent declarations.

Reproducibility gaps: A paper stating "AI was used for data analysis" tells readers nothing about model architecture, training data, hyperparameters, or validation methods. We've reviewed hospital AI research proposals where "machine learning" could mean logistic regression or a 50-layer transformer.

Governance lag: Hospital AI governance frameworks (see our clinical AI services) typically cover deployment of diagnostic or predictive models, but rarely address AI in the research process itself. This creates a blind spot for research ethics committees.

What types of AI use require disclosure in medical research?

We recommend Singapore hospital research teams adopt a three-domain disclosure framework:

1. AI-assisted writing and literature review

Examples: ChatGPT for drafting introductions, Claude for summarizing papers, Grammarly for grammar.

Disclosure threshold: Generative AI that produces original sentences or paragraphs requires disclosure. Grammar checkers and reference managers do not. The key distinction: did the tool create content or correct existing content?

What to disclose: Tool name, version, date accessed, specific sections where used (e.g., "ChatGPT-4 was used to draft an initial literature review outline on June 15, 2026; all content was subsequently verified and rewritten by authors").

2. AI-driven data analysis and prediction

Examples: Machine learning models for outcome prediction, LLMs for clinical note classification, computer vision for radiology feature extraction.

Disclosure threshold: Any AI model that processes patient data or generates results reported in the paper requires full methodological disclosure. This is not optional — it's a reproducibility requirement.

What to disclose: Model architecture, training data source and size, preprocessing steps, hyperparameters, validation approach, performance metrics, and code/model availability. For commercial tools (e.g., vendor-supplied risk scores), disclose the product name, version, and any known limitations.

The preprint on ICU time-series prediction [7] demonstrates good practice: it explicitly describes a "two-stream architecture" that decouples physiological signals from treatment protocols, addresses distribution drift, and provides implementation details. This level of transparency should be standard, not exceptional.

3. AI as the research subject

Examples: Studies evaluating LLM performance on clinical tasks, validating medical imaging AI, comparing ambient documentation tools.

Disclosure threshold: The AI system is the intervention being studied. Disclosure requirements are identical to any medical device or drug trial: full technical specifications, version control, and access conditions.

What to disclose: Everything in domain 2, plus: training data provenance, known biases, regulatory status (e.g., HSA classification), and conflicts of interest if the research team has commercial ties to the AI vendor.

A 2026 preprint on public perceptions of AI in healthcare [8] and another on synthetic medical image generation [9] both provide detailed methods sections that allow replication. Singapore hospital teams should use these as templates.

How Singapore hospital IRBs should operationalize AI disclosure

We've worked with institutional partners to develop AI disclosure checklists for research ethics review. Here's a practical framework:

Pre-submission checklist for researchers:

List every AI tool used in the research process (writing, analysis, visualization)
For each tool, answer: Did it generate new content/results, or modify existing work?
For generative AI: Provide tool name, version, date, prompts used, and human verification steps
For analytical AI: Provide full methods section per TRIPOD-AI or CONSORT-AI guidelines
Declare any commercial relationships with AI vendors

IRB review criteria:

Is the disclosure specific enough for a reader to assess validity? ("AI was used" is insufficient)
For patient data analysis, are model training data sources documented?
For generative AI in writing, is there evidence of human verification?
Does the study design account for AI-specific risks (e.g., hallucination, bias amplification)?

Post-approval monitoring:

Require version control: if researchers switch from GPT-4 to GPT-5 mid-study, that's a protocol amendment
For AI-driven analysis, require code and model archiving (institutional repository or public GitHub)
Flag papers for editorial review if AI use expands beyond the approved protocol

This is not theoretical. We've seen Singapore hospital research teams face journal rejection because AI disclosure was vague, and we've seen IRBs approve studies without realizing a commercial AI vendor had access to patient data during model training.

Why this matters in Singapore and Asia

Regulatory alignment: Singapore's PDPA and HSA frameworks govern AI deployment, but research use cases sit in a grey zone. Clear disclosure standards help hospitals demonstrate compliance when AI tools process patient data for research.

Regional leadership: Singapore hospital clusters publish high-impact research that influences clinical practice across Asia. If our disclosure standards are weak, downstream adoption decisions are uninformed.

Trust and transparency: Public trust in healthcare AI depends on transparency. The 2026 preprint on public perceptions [8] found that trust in AI-driven healthcare decisions correlates with perceived fairness and accountability. Vague disclosure undermines both.

Reproducibility crisis: Medical AI research already faces reproducibility challenges (see our post on clinical AI safety monitoring). Poor disclosure makes it worse. If a Singapore hospital publishes an ICU prediction model but doesn't disclose training data sources or hyperparameters, other hospitals can't validate or adapt it.

What to do next

For clinical researchers:

Adopt the three-domain disclosure framework (writing, analysis, subject) for every manuscript
When using generative AI, keep a log: tool name, version, date, prompts, and verification steps
For analytical AI, follow TRIPOD-AI or CONSORT-AI reporting guidelines (available at equator-network.org)
Disclose commercial relationships with AI vendors in conflict-of-interest statements

For hospital IRBs and research offices:

Update research ethics application forms to include structured AI disclosure questions
Provide researchers with a decision tree: "Does this AI use require disclosure?" (we can help design this — contact us)
Require code and model archiving for AI-driven analysis studies
Train IRB members on AI-specific risks (bias, drift, hallucination) so they can assess protocols effectively

For clinical informatics and governance teams:

Extend your AI governance framework (see our health data infrastructure post) to cover research use cases, not just deployment
Work with IRBs to define institutional AI disclosure standards — don't wait for journals to mandate it
Monitor preprint servers (arXiv, medRxiv) for emerging methods (like the ICU architecture [7] or synthetic imaging [9]) that your researchers might adopt

FAQ

Do I need to disclose using Grammarly or reference managers?

No. Tools that correct grammar, format citations, or organize references do not generate original content and do not require disclosure. The threshold is: did the tool create sentences, paragraphs, or results that appear in your paper? If yes, disclose. If it only corrected or formatted existing work, no disclosure needed.

What if I used ChatGPT to brainstorm ideas but didn't include any of its text?

If the AI-generated text does not appear in the manuscript and did not directly shape your analysis or conclusions, disclosure is optional but recommended for transparency. A simple acknowledgment ("ChatGPT was used for initial brainstorming; no AI-generated text appears in this manuscript") satisfies most journal policies.

Disclose what you can: product name, version, vendor, date accessed, and the specific task (e.g., "Vendor X Risk Score v2.1 was used to predict readmission risk; technical details are proprietary"). Then acknowledge the limitation: "Model architecture and training data are not publicly available, which limits independent validation." This signals to readers that reproducibility is constrained.

Should Singapore hospitals create institutional AI disclosure policies, or wait for journal mandates?

Create institutional policies now. Journal requirements are inconsistent and lag behind practice. An institutional policy protects your researchers (clear guidance reduces compliance risk), protects your hospital (demonstrates governance maturity), and protects the research community (improves transparency). We help hospital clusters design these policies — start a project if you need support.

Sources

[1] Trends in Newborn Hepatitis B Virus Vaccination—Reply. JAMA Network, July 21, 2026. https://jamanetwork.com/journals/jama/fullarticle/2850605

[2] Author AI Disclosure in JAMA Network Journal Submissions. JAMA Network, July 21, 2026. https://jamanetwork.com/journals/jama/fullarticle/2850541

[3] Mom, Doc, Patient, Home: Poetry and the Merging of Identities. JAMA Network, July 21, 2026. https://jamanetwork.com/journals/jama/fullarticle/2850534

[4] A Book for Medical Device Patents. JAMA Network, July 21, 2026. https://jamanetwork.com/journals/jama/fullarticle/2850453

[5] Finerenone in Patients With Chronic Kidney Disease Due to Glomerular Diseases Research Summary. JAMA Network, July 21, 2026. https://jamanetwork.com/journals/jama/fullarticle/2850123

[6] From Operations to Elderly Care Outcomes: A Thematic Review of Industrial Engineering and Decision-Support Approaches. arXiv preprint, July 21, 2026. https://arxiv.org/abs/2607.19075v1

[7] Biological Amnesia in ICU Time-Series Prediction: A Drift-Adaptive Two-Stream Architecture with Temporal Retrieval. arXiv preprint, July 21, 2026. https://arxiv.org/abs/2607.19020v1

[8] Public perceptions of AI-driven decision-making in healthcare: A structural equation modeling approach. arXiv preprint, July 21, 2026. https://arxiv.org/abs/2607.18884v1

[9] Local Label-Informed Feature Transfer for Generating Ground-Truth Medical Images: A Comparison of GAN- and Diffusion-Based Approaches. arXiv preprint, July 21, 2026. https://arxiv.org/abs/2607.18882v1

[10] The Evolving Role of Artificial Intelligence in Medical Genetics: Advancing Healthcare, Research, and Biosafety Management. Genes, December 1, 2025. https://pubmed.ncbi.nlm.nih.gov/41595426/

[11] Advancing a Research Agenda on the Ethical Challenges and Human Factors that will be shaping the evolution of AI/ML/DL enhanced NDE processes. e-Journal of Nondestructive Testing, July 1, 2026. https://doi.org/10.58286/33520

[12] Feasibility of Tailoring Artificial Intelligence-Assisted Ambient Scribes for Intensive Care Unit Rounds: Algorithm Development and Validation. JMIR Medical Informatics, July 7, 2026. https://pubmed.ncbi.nlm.nih.gov/42412948/

[13] Beyond the stigma: a community-based mixed methods hybrid photovoice–appreciative inquiry protocol to explore and enhance engagement of young men in mental health research. BMJ Open, July 1, 2026. https://bmjopen.bmj.com/content/bmjopen/16/7/e115182.full.pdf

[14] Towards patient-centered research in pregnancy-associated breast cancer: creating a science agenda through a priority-setting-partnership. Breast, May 1, 2026. https://pubmed.ncbi.nlm.nih.gov/42150215/

[15] Zero-Knowledge Process Verification: A Comprehensive Framework for a Distributed Healthcare System. Blockchain in Healthcare Today, 2026. https://pubmed.ncbi.nlm.nih.gov/42205846/

[16] Artificial intelligence-generated synthetic data for cancer research and clinical trials. Nature Reviews Cancer, May 2026. https://pubmed.ncbi.nlm.nih.gov/41720945/

[17] Integrating diversity, equity, and inclusion in generative AI applications for healthcare education: a scoping review. International Journal of Medical Informatics, July 1, 2026. https://doi.org/10.1016/j.ijmedinf.2026.106582

[18] Symptom burden in multiple long-term conditions: An AI-supported, mixed-methods concept elicitation study. JRSM Open, July 1, 2026. https://doi.org/10.1177/20542704261459717

AI Disclosure in Medical Research: What Singapore Hospital Teams Need to Know in 2026