OpenAI 向免费用户提供 GPT-5.5 Instant 的劣质健康建议，付费用户可用 GPT-5.6 Sol

OpenAI 正在向美国 18 岁及以上用户推送“ChatGPT 健康”功能，此前自一月起已开始发布和测试该功能。订阅服务甚至可能真正拯救你的生命。

用户可以连接 Apple Health、医疗记录和健康应用，以查看化验结果、为医生预约做准备，并分析睡眠或活动数据。OpenAI 表示不会将连接的健康数据用于模型训练或广告。

使用免费版 ChatGPT 的用户获得的健康建议质量较低：https://openai.com/index/health-in-chatgpt/。OpenAI 使用 GPT-5.5 Instant 支持该功能：https://the-decoder.com/chatgpts-new-health-upgrade-beats-doctor-written-answers-openai-says/，其在健康基准测试中的得分低于为付费用户保留的新旗舰模型 GPT-5.6 Sol：https://openai.com/index/gpt-5-6/。OpenAI 很可能会从伦理角度为这种双层系统进行辩护，指出两种模型在 HealthBench Professional 测试中都超过了医生的答案：https://the-decoder.com/openai-says-its-latest-models-outperform-doctors-in-medical-benchmark/。广告

即便基准测试结果看似决定性，其也来自用于测量知识的人工测试环境，而医生的成绩可能因多种原因较低。他们可能面临时间压力、疲劳，或在没有患者记录或同事输入等工具的情况下进行测试。广告 DEC_D_Incontent-1

基准测试也无法覆盖实际医疗检查中发生的许多情况，在真实场景中，医生可以面对面检查患者，捕捉非语言线索，并利用多年的经验评估整体健康状况。OpenAI 在公告中反复声明，ChatGPT 仍可能出错，不能替代医疗建议。公司表示有超过 260 名医生参与了健康功能的开发。

OpenAI 的早期测试还发现，超过 70% 的参与者在专门的健康板块之外提出健康问题，因为切换到该板块太麻烦。公司此后已在任何对话中提供健康功能，同时保留独立的健康板块以管理数据和访问以往的健康对话。广告

这些风险并非假设性的。在最新的放射学基准测试 RadLE 2.0 中，测试的 16 个 AI 模型没有一个的表现能与人类放射科医生相比：https://the-decoder.com/ai-chatbots-reading-x-rays-can-be-dangerously-confident-even-when-theyre-wrong/。主要问题是，聊天机器人在给出错误结果时会高度自信，而不是承认它们已经达到了极限，而人类放射科医生在承认不确定性方面要好得多。这种过度自信、说服力和谄媚的混合——聊天机器人验证用户而不是挑战错误假设——也导致了严重的心理健康危害：https://the-decoder.com/openais-sam-altman-warned-of-ais-superhuman-persuasion-in-2023-and-2025-proved-him-right/。Ad DEC_D_Incontent-2

与此同时，一些报告显示，AI 在健康数据中发现了医学专业人员可能遗漏的模式，且有时结果非常显著：https://the-decoder.com/chatgpt-helped-identify-a-genetic-mthfr-mutation-after-a-decade-of-missed-diagnoses/。用于电子健康记录的系统 MIRA 和 AMIE 在模拟诊疗中表现大致与初级保健医生相当：https://the-decoder.com/ai-systems-rival-doctors-in-new-nature-studies-but-one-result-suggests-the-tech-wont-age-well/。一位研究人员将这样的 AI 代理比作飞机的自动驾驶系统：“这些系统可以通过接管例行任务来支持和减轻医疗专业人员的负担，但最终责任始终由医生承担。” Ad

保持对 AI 的关注。内容清晰、实用，没有冗余。

关注 The Decoder 获取 AI 新闻、背景故事和专家分析。

The Decoder：https://the-decoder.com/

OpenAI is rolling out "Health in ChatGPT" to U.S. users aged 18 and older after unveiling and testing the feature since January. Paying for a subscription could literally save your life.

Users can connect Apple Health, medical records, and wellness apps to review lab results, prepare for doctor's appointments, and analyze sleep or activity data. OpenAI says it won't use connected health data for model training or advertising.

Users on the free version of ChatGPT receive lower-quality health advice：https://openai.com/index/health-in-chatgpt/. OpenAI powers the feature with GPT-5.5 Instant：https://the-decoder.com/chatgpts-new-health-upgrade-beats-doctor-written-answers-openai-says/, which scores lower on health benchmarks than the new flagship model, GPT-5.6 Sol：https://openai.com/index/gpt-5-6/, reserved for paying users. OpenAI will likely defend this two-tier system on ethical grounds by pointing out that both models beat doctors' answers on the HealthBench Professional：https://the-decoder.com/openai-says-its-latest-models-outperform-doctors-in-medical-benchmark/ test. Ad

Even when benchmark results appear decisive, they come from artificial test environments designed to measure knowledge, and doctors may score lower for several reasons. They may be under time pressure, dealing with fatigue, or taking the test without tools such as patient records or input from colleagues. Ad DEC_D_Incontent-1

Benchmarks also can't capture much of what happens during an actual medical exam, where doctors can examine patients in person, pick up on nonverbal cues, and draw on years of experience to assess their overall condition. OpenAI itself repeatedly states in the announcement that ChatGPT can still make mistakes and can't replace medical advice. The company says more than 260 physicians helped develop the Health features.

OpenAI's early tests also found that more than 70 percent of participants asked health questions outside the dedicated Health section because switching to it was too cumbersome. The company has since made Health available in any conversation while keeping the separate Health section for managing data and accessing past health chats. Ad

Those risks aren't hypothetical. In the latest radiology benchmark, RadLE 2.0, none of the 16 AI models tested performed as well as human radiologists：https://the-decoder.com/ai-chatbots-reading-x-rays-can-be-dangerously-confident-even-when-theyre-wrong/. The main problem was that chatbots gave incorrect findings with high confidence instead of admitting when they had reached their limits, while human radiologists were far better at acknowledging uncertainty. That mix of overconfidence, persuasion, and sycophancy, where chatbots validate users rather than challenge false assumptions, has also contributed to serious mental health harms：https://the-decoder.com/openais-sam-altman-warned-of-ais-superhuman-persuasion-in-2023-and-2025-proved-him-right/. Ad DEC_D_Incontent-2

At the same time, some reports show AI spotting patterns in health data that medical professionals miss, sometimes with striking results：https://the-decoder.com/chatgpt-helped-identify-a-genetic-mthfr-mutation-after-a-decade-of-missed-diagnoses/. MIRA, a system for electronic health records, and AMIE both performed about as well as primary care doctors in simulated consultations：https://the-decoder.com/ai-systems-rival-doctors-in-new-nature-studies-but-one-result-suggests-the-tech-wont-age-well/. One researcher compared AI agents like these to an airplane's autopilot: "These systems can support and relieve medical professionals by taking over routine tasks, but ultimate responsibility will always remain with the physicians." Ad

Stay in the loop on AI. Clear, useful, no fluff.

Follow The Decoder for AI news, background stories and expert analyses.

The Decoder：https://the-decoder.com/