From the Democratization of Medical Knowledge to the Democratization of Medical Data

As large language models (LLMs) rapidly enter the healthcare sector, artificial intelligence is evolving beyond providing general health information and is beginning to interact directly with patients’ most sensitive personal health data.

A recent Viewpoint article in JAMA, “When Patients Share Everything With an AI Chatbot: Risks and Opportunities of Large Language Models,” examines the opportunities and challenges posed by a new generation of health-focused AI platforms such as ChatGPT Health. The authors argue that as patients gain the ability to synchronize their complete electronic health records (EHRs) with AI systems, healthcare is moving from the democratization of medical knowledge toward the democratization of medical data.

While this transformation has the potential to advance personalized medicine, it also raises unprecedented concerns regarding privacy, bias, and regulatory oversight.

What Is the Potential Value of AI Access to Complete Medical Records?

In theory, if AI systems can integrate patients’ medical histories, health monitoring data, wearable device information, and lifestyle records, they may provide several important benefits:

*More personalized health recommendations

*Earlier identification of rare diseases

*Enhanced epidemic and public health surveillance

*Greater utilization of Real-World Data (RWD)

*Accelerated clinical research and drug development

For researchers, large-scale and real-time integration of health data may become a critical foundation for the future of precision medicine.

Medical Records Are Not Entirely Objective

However, the authors highlight an important reality:

Electronic health records are not neutral repositories of facts.

In addition to objective laboratory and diagnostic findings, medical records often contain subjective assessments and interpretations made by healthcare professionals.

Examples include:

1. Behavioral descriptions of patients

2. Preliminary diagnostic impressions

3. Unconfirmed clinical assumptions

4. Potentially biased language or documentation

If AI systems learn directly from these records, they may reproduce—or even amplify—existing biases.

For example, a patient suffering from chronic pain may have previously been labeled as “drug-seeking.” Even if subsequent evaluations confirm a legitimate physiological cause for the pain, an AI system may still be influenced by earlier documentation and provide less appropriate recommendations.

In other words, AI systems may learn not only medical knowledge but also the biases embedded within healthcare systems.

AI May Reinforce Existing Health Disparities

The authors further note that healthcare systems already face significant health disparities.

Patients from different racial, ethnic, socioeconomic, and demographic backgrounds may experience unequal diagnosis and treatment.

If such disparities are reflected in EHR data and AI systems treat these records as objective truth, future applications may generate:

*Biased diagnoses

*Biased recommendations

*Biased risk assessments

As a result, existing healthcare inequities could become further entrenched.

Therefore, the risks associated with AI may stem not only from the model itself but also from the data used to train and inform it.

Can HIPAA Protect Patient Data Uploaded to AI Platforms?

Another key issue discussed in the article is data privacy.

Many patients assume that their medical information remains protected under the Health Insurance Portability and Accountability Act (HIPAA).

However, the authors point out that once patients voluntarily upload their medical records to a commercial AI platform, those data may no longer be fully protected under HIPAA.

The reason is that most AI platforms are not considered HIPAA-covered entities.

Consequently:

1. HIPAA restrictions on data use may not apply.

2. HIPAA security requirements may not apply.

3. HIPAA breach notification obligations may not apply.

Although AI companies may promise strong privacy protections, corporate privacy policies are fundamentally different from legally enforceable regulatory safeguards.

Lack of Transparency May Be the Greater Concern

According to the authors, the most significant challenge may not be data breaches, but rather the lack of transparency.

Currently, independent researchers have limited ability to evaluate:

*Whether AI systems exhibit bias against specific populations

*Whether safety incidents have occurred

*Whether inappropriate medical recommendations are being generated

*Whether AI is influencing patients’ healthcare-seeking behavior

Because these data remain under the control of platform developers, external validation is often impossible.

As a result, even when AI companies report strong performance, there may be insufficient independent evidence to verify such claims.

PI-Union Medical Science Commentary

Health-focused AI platforms such as ChatGPT Health represent an important milestone in the evolution of AI-powered healthcare. For the first time, patients may be able to provide AI systems with comprehensive health records for personalized analysis.

However, when AI begins reading entire medical records, it receives not only information about diseases and treatments but also decades of accumulated clinical assumptions, documentation biases, and systemic healthcare challenges.

For this reason, the future development of healthcare AI should not focus solely on technological innovation. Equal attention must be given to data governance, clinical evidence generation, regulatory oversight, and ongoing performance monitoring.

Only through robust safeguards can AI become a tool for improving healthcare outcomes rather than amplifying existing inequities and risks.

Reference: When Patients Share Everything With an AI Chatbot–Risks and Opportunities of Large Language Models

Reviewer: PI-Union Medical Science Ltd.

* E-mail: piunion@pi-union.com

* Official Website: https://pi-union.com/

* Facebook: www.facebook.com/piunion2020

* Youtube: www.youtube.com/@pi-union

* Instagram: www.instagram.com/piunion2020

* LINE: @654eukag

從醫療知識民主化到醫療資料民主化

隨著大型語言模型（Large Language Models, LLMs）快速進入醫療領域，人工智慧不再只是提供一般健康資訊，而開始直接接觸患者最敏感的個人健康資料。

近期刊登於《JAMA》的 Viewpoint 文章〈When Patients Share Everything With an AI Chatbot: Risks and Opportunities of Large Language Models〉，探討了新一代健康型 AI 平台（如 ChatGPT Health）所帶來的機會與挑戰。作者指出，當患者能夠將完整電子病歷（Electronic Health Records, EHRs）直接同步至 AI 系統時，我們正從「醫療知識的民主化」邁向「醫療資料的民主化」。

這項變革可能促進個人化醫療發展，但同時也引發前所未有的隱私、偏見與法規問題。

AI讀取完整病歷：潛在價值何在？

理論上，如果 AI 能夠整合患者的病歷資料、健康監測數據、穿戴裝置資訊以及生活型態紀錄，將可能帶來多項效益：

* 提供更個人化的健康建議

* 協助罕見疾病的早期辨識

* 強化流行病監測能力

* 促進真實世界資料（Real-World Data, RWD）的運用

* 加速臨床研究與藥物開發

對於研究人員而言，大規模且即時的健康資料整合，更有機會成為未來精準醫療（Precision Medicine）的重要基礎。

醫療紀錄並非完全客觀

然而，作者提醒我們一個重要事實：

「電子病歷不是中立的事實倉庫。」

病歷中的內容除了客觀檢驗數據外，也包含醫護人員的主觀判斷與描述。

例如：

1. 對患者的行為評價

2. 初步診斷推測

3. 尚未證實的臨床印象

4. 可能帶有偏見的文字紀錄

如果 AI 系統直接學習這些內容，可能會將既有偏見複製甚至放大。

例如某位慢性疼痛患者曾被標註為「疑似尋求藥物（drug-seeking）」，即使後續證實其疼痛有明確生理原因，AI 仍可能受到早期紀錄影響，而提供較不適當的建議。

換言之，AI 不只是學習醫學知識，也可能學習醫療體系中的偏見。

AI可能複製醫療不平等

作者進一步指出，醫療體系本身就存在健康不平等（Health Disparities）問題。

不同種族、性別、社經背景的患者，在診斷與治療過程中可能面臨不同待遇。

若這些偏差已存在於電子病歷中，而 AI 又將其視為「真實資料」進行學習，未來可能形成：

* 偏差診斷（Biased Diagnosis）

* 偏差建議（Biased Recommendations）

* 偏差風險評估（Biased Risk Assessment）

最終使既有醫療不平等被進一步強化。

因此，AI 的風險不一定來自模型本身，而可能來自模型所學習的資料。

HIPAA保護不了患者主動上傳的資料？

文章另一項值得關注的議題是資料隱私。

許多患者認為醫療資料受到 HIPAA（Health Insurance Portability and Accountability Act）保護，因此相當安全。

然而作者指出：

當患者自行將病歷上傳至商業化 AI 平台後，這些資料未必仍受到 HIPAA 的完整保障。

原因在於：

AI 平台通常不是 HIPAA 所定義的醫療照護提供者（Covered Entity）。

因此：

1. HIPAA 的資料使用限制可能不適用

2. HIPAA 的資訊安全規範可能不適用

3. HIPAA 的資料外洩通報義務可能不適用

雖然企業可能承諾保護使用者隱私，但企業政策與法律保障仍有本質上的差異。

缺乏透明度是更大的問題

作者認為，目前最大的隱憂或許不是資料外洩，而是缺乏透明度。

目前外界難以得知：

* AI 是否對特定族群產生偏見

* 是否曾發生安全事件

* 是否提供不當醫療建議

* 是否影響患者就醫行為

由於相關資料掌握在平台開發者手中，獨立研究人員往往無法進行驗證。

因此，即使 AI 系統宣稱具有良好效能，也缺乏足夠的第三方證據支持。

百聯醫學評論

ChatGPT Health 類型的產品代表醫療 AI 發展的重要里程碑，也讓患者首次有機會將自己的完整健康資料交由 AI 分析。

然而，當 AI 開始閱讀完整病歷時，它接收到的不只是疾病資訊，也包括醫療體系長期累積的偏見、誤判與制度性問題。

因此，未來醫療 AI 的發展不應只關注技術創新，更需要建立完善的資料管理、臨床證據與法規監管機制，才能真正讓 AI 成為改善健康照護的工具，而非放大既有問題的新風險來源。

Reference: When Patients Share Everything With an AI Chatbot–Risks and Opportunities of Large Language Models

Reviewer: PI-Union Medical Science Ltd.

* E-mail: piunion@pi-union.com

* Official Website: https://pi-union.com/

* Facebook: www.facebook.com/piunion2020

* Youtube: www.youtube.com/@pi-union

* Instagram: www.instagram.com/piunion2020

* LINE: @654eukag