時事英豪｜看新聞學英文《AI臨床表現勝醫生　缺乏問責難替代》

【編者按】為回應家長和學生對提升英文能力的需求，點教育推出欄目「時事英豪」。本欄目精選社會熱點，以中英對照形式解構新聞時事，並提煉實用英文詞彙，文末更設有思考題，助學生在了解時事的同時鍛煉語言能力。

儘管科技飛速發展，人們還是不願將性命攸關的醫療托付於人工智能。

內容如下：

美國非牟利衞生政策研究機構凱澤家族基金會（KFF），今年較早時發布的民調結果顯示，約三分之一美國成年人正使用人工智能（AI）獲取健康信息與建議。

The Kaiser Family Foundation (KFF), a US nonprofit health policy research organization, released polling results earlier this year showing that about one-third of American adults are using artificial intelligence (AI) to obtain health information and advice.

美國AI公司OpenAI於今年7月上線ChatGPT Health功能，允許用戶整合醫療紀錄與健康數據，從與聊天機器人的對話中獲取個人化的健康資訊。

In July of this year, the US AI company OpenAI launched the ChatGPT Health feature, allowing users to integrate medical records and health data and obtain personalized health information through conversations with the chatbot.

但有分析認為，即使AI看病表現或優於人類，但人們還沒準備好把性命交給AI。

However, some analysts believe that even if AI performs better than humans in medical diagnosis, people are not yet ready to entrust their lives to AI.

業內研究表明，AI在部分臨床任務中的表現已取得突破性成就，或已超越人類醫生。

Industry research shows that AI has already achieved breakthrough results in some clinical tasks and may have surpassed human doctors.

兩篇發表於科學期刊《自然》的研究顯示，德國學術團隊開發的MIRA，與Google研發的AMIE這兩款AI智能體，在虛擬測試中已能媲美、甚至超越人類醫生的臨床決策能力。

Two studies published in the scientific journal Nature show that two AI agents (MIRA, developed by a German academic team, and AMIE, developed by Google) were able to match or even surpass human doctors' clinical decision-making capabilities in virtual tests.

在500多宗急症病例評估中，AI診斷準確率達87.8%，優於6名跨專科醫生組成的人類專家組的78.1%。

In the evaluation of more than 500 emergency cases, AI achieved a diagnostic accuracy rate of 87.8%, outperforming the 78.1% achieved by a human expert panel composed of six multidisciplinary physicians.

由哈佛醫學院等機構組成的研究團隊，在科學期刊《科學》上發表的論文亦指出，OpenAI的模型，在6項臨床推理任務中表現遠勝人類，但在「不可出錯診斷」、即一旦出錯就會搞出人命的測試中，AI模型還未與人類拉開差距。

A research team composed of institutions including Harvard Medical School also pointed out in a paper published in the scientific journal Science that OpenAI's model performed far better than humans in six clinical reasoning tasks, but in "diagnoses that cannot afford errors"—that is, tests in which a mistake could cost a person's life—AI models have not yet opened a gap with humans.

不過研究團隊坦承，該研究的對照組是內科醫生，而非急症室醫生，結論有局限性。

However, the research team acknowledged that the control group in the study was internists rather than emergency room doctors, so the conclusions have limitations.

同期的類似研究則獲得截然不同的結果。美國醫學會期刊《JAMA Network Open》刊文稱，業界現有的大型語言模型（LLM），包括GPT-5、Claude 4.5 Opus、Gemini 3、Grok 4等，在臨床推理早期的鑒別診斷階段，錯誤率普遍逾80%。

Similar studies published around the same time reached sharply different results. An article in the American Medical Association journal JAMA Network Open stated that existing large language models (LLMs) in the industry, including GPT-5, Claude 4.5 Opus, Gemini 3, Grok 4, and others, generally had error rates exceeding 80% in the early differential diagnosis stage of clinical reasoning.

分析認為，AI在信息不足、需逐步推理的早期階段，往往容易過早作出結論。相比之下，臨床醫生通常會在初期保留多種可能性，並隨着檢查結果修正判斷，最終才形成定論。

Analysts believe that in the early stage, when information is insufficient and step-by-step reasoning is required, AI is often prone to reaching conclusions prematurely. By contrast, clinicians usually retain multiple possibilities in the initial stage and revise their judgments as test results come in, ultimately forming a conclusion.

重點詞彙：