The Day When "Non-Human Subject" Research Increases - The Temptation to Skip Review: Research Accelerated by Synthetic Medical Data, Shaken Trust

Universities Skipping Ethical Review with "Synthetic Medical Data"?—AI Research Caught Between Speed and Trust

Introduction

"If it's 'synthetic medical data' generated by AI, it doesn't involve human subjects, so it's exempt from ethical review (IRB)."—This rationale has led several medical research institutions in Canada, the US, and Italy to bypass the usual ethical reviews, as reported by Nature, sparking heated debates on social media both domestically and internationally. The backdrop to this issue is the challenge faced by medical AI to balance patient privacy protection with research speed.Nature

WebProNews also covered this trend, highlighting the benefits of accelerated research and the concerns over the erosion of ethical standards. This article will summarize the key points from the original report, cross-referencing the latest research findings, governance guidelines, and social media reactions to outline the critical issues that businesses, universities, and media stakeholders should consider.WebProNews

What's Happening: Synthetic Data and the "Non-Human Subject Research" Argument

According to the Nature article, institutions like the University of Washington School of Medicine (US), Ottawa Children's Hospital/Ottawa Hospital (Canada), and IRCCS Humanitas Research Hospital (Italy) are expanding operations that use AI-generated synthetic data as "not human subject research", thus not requiring IRB review. The rationale is that it does not contain identifiable information from real patients. In the US, this interpretation is said to align with the 1991 Common Rule. However, some facilities require IRB approval for accessing real data to create synthetic data, indicating inconsistent boundaries.The process of accessing real data to create synthetic data requires IRB approval, and the demarcation is not uniform.Nature

What is Synthetic Data?

It is a method where a generative model trained on real patient data creates fictitious records that retain statistical properties. Proponents argue it facilitates sharing and advances inter-facility collaboration. However, depending on how it's created, concerns about bias and information leakage may persist.Nature

Is Privacy Really Safe? Re-identification and Membership Inference

"Anonymization or synthesis does not necessarily equal safety." Stanford HAI has warned that anonymization relying on HIPAA's "18-item removal" cannot completely eliminate re-identification risk, as individuals could be identified by matching with external data.Stanford HAI

Moreover, research in 2024 demonstrated that even models trained on anonymized clinical notes or synthetic notes could be subject to "membership inference attacks" to deduce whether they were part of the training data. This challenges the rationale for skipping reviews on the grounds of "no real patient involvement."Nature

Ethical and Legal Grey Areas: Review is Not a "Zero or Full" Choice

Frameworks supporting AI research reviews are being developed even in the world of IRBs. The MRCT Center, based at Harvard, plans to release an IRB Review Framework for AI Research in 2025. It provides practical checklists from initial questions about whether it's research, human participation, and AI usage (intervention or operational support) to fairness, transparency, consent, and data governance. It encourages consistent oversight, including **"light-touch reviews"** based on risk, rather than a binary choice of full review or none.Brigham and Women's Hospital Clinical Trials Center

Research Speed vs. Social Trust: Benefits and Side Effects

The benefits emphasized by proponents are clear. Synthetic data makes cross-facility data sharing easier, increasing the speed and scale of research. It enables simulations of rare diseases, drug reactions, and long-term follow-ups that are difficult to verify with real data alone—points highlighted by stakeholders in Nature.Nature

However, if the **"review skip" prioritizing speed becomes common, it accumulates an "invisible cost" of eroding social trust**. We must not forget the growing public backlash against "stealth experiments" using synthesis and AI in recent years.The Washington Post

How Social Media Reacted: The Atmosphere of Approval and Disapproval

• Nature's original post served as a starting point, widely shared by community accounts of medical libraries and emergency physicians. Interest from the clinical field was high, with strong reactions to headlines suggesting "bypassing usual reviews."X (formerly Twitter)X (formerly Twitter)

• Some researchers and educators posted critical comments, describing it as **"deeply dubious."** Concerns were raised that exceptions to ethical reviews might ultimately encourage research misconduct or algorithm "runaway."X (formerly Twitter)

• Retraction Watch also shared the article. The dissemination by this media, which has been tracking inappropriate AI research practices and retraction cases, symbolizes the growing "watchful eye."X (formerly Twitter)

Summarizing the social media atmosphere, the clinical and educational communities are mixed with "cautious optimism" and "distrust of procedural neglect." The medical and ethical clusters appear more cautious than the general tech community, indicating a temperature difference.

For Practitioners: "7 Self-Checks" for Synthetic Medical Data

Even with "review skips," there is much to be done. This is the reality pointed out by the Nature article, MRCT guidelines, and privacy research. Before starting a project, check at least the following seven points.

1. Audit of the Generation Process: Document protocols for evaluating synthetic degree (real-to-synthetic relationship, statistical similarity) and "reverse lookup" possibility.Nature

2. Privacy Threat Model: Evaluate membership inference, re-identification, and linkability (using differential privacy if necessary).Nature

3. Data Card/Model Card: Publicize data origin, filters, biases, and caution points for unexpected use (internally if necessary).Brigham and Women's Hospital Clinical Trials Center

4. Fairness Testing: Check for gaps in representativeness (handling of minorities and rare diseases).Brigham and Women's Hospital Clinical Trials Center

5. Human-in-the-Loop: If involved in decision-making, specify human intervention points and stop valves (stop conditions).Brigham and Women's Hospital Clinical Trials Center

6. Transparency and Disclosure: Clearly indicate the use of synthetic data in papers and reports, and note limitations and assumed risks.Brigham and Women's Hospital Clinical Trials Center

7. Establishing "Lightweight Reviews": Even if positioned as outside the scope of IRB, establish an AI research-specific review window, and systematize annual audits and post-reviews.Brigham and Women's Hospital Clinical Trials Center
   
   

Common Misunderstandings (FAQ)

• Q: Is there zero personal information because it's synthetic?
  A: If real data is used as material for generation and it's statistically "too similar," there may be room for membership inference or re-identification. Risk assessment, including processing steps, is essential.Nature

• Q: Is IRB unnecessary and not a problem?
  A: Even if technically deemed unnecessary, from the perspective of social trust and research reproducibility, going through light-touch third-party reviews often reduces long-term costs.Brigham and Women's Hospital Clinical Trials Center

• Q: Is synthetic data a "bias remover"?
  A: It may expand or amplify the distortions of the original data. Testing for representativeness and confirming out-of-distribution behavior are essential.Brigham and Women's Hospital Clinical Trials Center
  ##