ChatGPT for Doctors | GPT-5.4 Surpasses Physicians in 7,000-Case Evaluation
機械翻訳 / Machine-translated
A
AI Friends
2026年4月26日・31分で読める
機械翻訳 / Machine-translated
@aifriends
AI Friends（https://aifriends.jp）のクロスポスト公式アカウント。AIツールの紹介・使い方・できることを、中学生でもわかるやさしい日本語で届けます。
"Doctor, the referral letter still isn't done" — the reality of physicians exhausted by paperwork between appointments is finally about to change.
On April 22, 2026, OpenAI announced "ChatGPT for Clinicians," a free AI tool for U.S. healthcare professionals.
Powered by GPT-5.4, it has demonstrated performance that surpasses human physicians.
Here is a comprehensive breakdown of what is changing, and when Japanese physicians can expect access.
Let's start by getting an overview of this release.
OpenAI officially launched "ChatGPT for Clinicians" on April 22, 2026 (U.S. time).
It is targeted at verified U.S. healthcare professionals across four categories: physicians, nurses, pharmacists, and PAs (Physician Assistants — licensed professionals who support physicians).
The price? Completely free.
"The latest cutting-edge AI for professionals, entirely at no cost for those with medical credentials" — this is a remarkably generous offering. It symbolizes OpenAI's move to place healthcare at the top of its strategic priorities.
The features are organized around three main areas of clinical work: (1) Care Consultation (a sounding board for difficult clinical decisions), (2) Documentation (drafting referral letters, patient-facing explanations, medical records, and other documents), and (3) Medical Research (summarizing recent literature and organizing evidence).
Think of it as "hiring a veteran attending physician with 30 years of experience, a secretary, and a medical school librarian — all for free, 24 hours a day." All of it is integrated through GPT-5.4's reasoning capabilities and medical literature search, accessible from a single screen.
The backdrop is a dramatic surge in AI adoption among U.S. physicians over the past year.
Usage jumped from 48% in 2025 to 72% in 2026 — a 24-point increase in just one year.
More than two out of every three physicians are now using some form of AI on a daily basis.
OpenAI's strategy has evolved from "handing over office supplies" to "providing a dedicated workbench." Making it free is almost certainly a move to quickly dominate the U.S. physician market.
The biggest headline is performance that surpasses human physicians in evaluation. Let's look at the remarkable figures.
OpenAI simultaneously announced a new public benchmark called "HealthBench Professional." It uses real clinical conversation cases and scoring rubrics written by physicians, evaluated through a multi-stage review process involving multiple doctors.
Rather than "AI solving a test it created itself," this is designed more like "having AI take a board exam written by practicing physicians" — a rigorous and objective design. It is widely expected to become the industry standard for objectively comparing medical AI performance.
The results were striking.
GPT-5.4 in the ChatGPT for Clinicians workspace recorded a score of 59.0, surpassing the base GPT-5.4, other OpenAI models, external medical AIs, and human physicians.
Notably, it also outperformed physician teams given unlimited time and access to web search.
The impact is comparable to "a high schooler who won a national competition in the qualifying rounds going on to beat professional athletes in the final." That said, 59 points is not a perfect score, so the premise remains unchanged: AI is not infallible and physician oversight is essential.
The safety figures are equally impressive.
Physician advisors reviewed approximately 7,000 test conversations, and 99.6% were rated "safe and accurate."
This reflects an extreme reduction in hallucination — the phenomenon where AI confidently states incorrect information.
This is roughly "only 4 errors out of 1,000 responses." Of course, even a 0.4% risk cannot be tolerated in clinical settings, which is why the system is designed so that final decisions always rest with the physician.
Here is an organized look at what healthcare professionals can actually do with this tool.
This feature provides evidence-backed answers drawn from peer-reviewed medical literature, with citations automatically attached to each response — allowing physicians to immediately jump to primary sources for verification.
Think of it as receiving "a mini-report with references, not just an answer." It significantly reduces the time spent searching UpToDate or PubMed, making evidence-based medicine (EBM) much more practical in everyday use.
This feature lets users save templates for recurring tasks — such as referral letters and patient explanations — and reuse them repeatedly.
Once you register a "referral letter template formatted for your institution," all you need to do next time is enter the patient details and it generates instantly.
It's as easy as "saving a cooking recipe on your phone and just entering the ingredients each time." Customization to match facility-specific protocols and regional care coordination formats is also possible.
One feature not to overlook is automatic tracking of CME (Continuing Medical Education) credits. U.S. physicians are required to earn CME credits on a regular basis, and under this system, simply conducting evidence-based research within ChatGPT counts toward those credits.
"You study, and your certification renewal points accumulate at the same time" — two birds with one stone.
This addresses one of the most common frustrations for U.S. physicians — being too busy to attend training sessions — and has been well received by the medical community.
For privacy compliance, users can enter into a Business Associate Agreement (BAA) compliant with HIPAA (the U.S. healthcare data protection law).
This is a legally binding agreement stating "data you use will be properly protected as patient information." Conversation content is not used for model training, and multi-factor authentication and data encryption are included as standard.
This proactively addresses physician concerns about data leakage.
The medical AI market is already highly competitive. Here is a breakdown of the major players.
The positioning is clear: "Nuance for charting, OpenEvidence for literature, Hippocratic for patient interaction, ChatGPT for Clinicians for all of the above — free." This is a direct collision for OpenEvidence in particular, and the battle for physicians' attention has begun in earnest.
The biggest question for Japanese physicians is: "When can we use this in Japan?"
To get straight to the point: as of April 2026, Japanese physicians are not eligible for the free ChatGPT for Clinicians tier. OpenAI launched it exclusively for verified U.S. healthcare professionals, and announced plans to expand to other countries over the coming months through an organization called the Better Evidence Network. The realistic timeline for Japan is the second half of 2026 through the first half of 2027.
This follows OpenAI's typical rollout pattern: "Run it in the U.S., then take it global once things are working smoothly."
That said, Japanese physicians are already making creative use of the general ChatGPT and ChatGPT Enterprise. At Iryohojin Shaidan Morgenrot, doctors paste patient records from the EHR and automatically generate referral letters (診療情報提供書), cutting the time required to roughly one-tenth.
"A ten-minute document done in one minute" — a dramatic efficiency gain. Other documented uses include drafting patient-facing explanations, summarizing medical literature, and generating discharge summaries.
Japan also faces its own specific challenges.
Hallucination — where AI confidently produces plausible-sounding but incorrect information — can be fatal in clinical settings. The rule "never delegate anything resembling a medical diagnosis to AI" must be strictly enforced.
Standard practice has also emerged around anonymizing all inputs: patient names, dates of birth, My Number IDs, EHR IDs, and addresses must never be entered.
Ministry of Health, Labour and Welfare and Japan Medical Association official guidelines are still being developed, meaning institutions adopting these tools currently need to create their own internal rules.
Dr. Yamada sees more than 20 outpatients a day and spends evenings writing five or six referral letters. By pasting a patient chart summary and asking "Please write a referral letter for this patient to the cardiology department at ○○ Hospital," ChatGPT for Clinicians (or the general ChatGPT version in Japan) generates a draft in three seconds, covering past medical history, chief complaint, clinical course, and reason for referral.
"Fifteen minutes per letter drops to one minute." Leaving the office at 9 PM becomes leaving at 7 PM — making it home in time for dinner with the family.
During a night shift, a family member of a new patient calls asking: "What are the side effects of that medication?"
Sasaki asks ChatGPT: "Taking into account this patient's age, medical history, and current medications, please explain the main side effects of ○○ in simple language for a family member." A medically accurate explanation written at an accessible level is generated instantly.
"Thirty minutes spent anxiously drafting an explanation is now done in three minutes." With a final check always performed by the nurse before use, safe operation is maintained.
Preparing for a case conference, Dr. Suzuki asks: "Summarize the latest treatment evidence for ○○ syndrome from 2025 onward across five key papers, including sample sizes and limitations for each study." A primary-source summary with citation links is ready in 30 seconds.
"Reading five papers on PubMed until midnight is consolidated into actually reading the papers themselves." Freed from the literature-search work that used to steal thinking time, residents can now devote that time to clinical reasoning and communicating with patients.
A. The U.S.-only free version of "ChatGPT for Clinicians" is not currently available to Japanese physicians.
However, general ChatGPT Plus (USD 20/month) and ChatGPT Enterprise are available in Japan, and the general-purpose GPT-5.4 they run on is at a comparable performance level.
With well-designed medical prompts, Japanese physicians can achieve equivalent workflow efficiencies starting today.
For institutional adoption, signing a data protection agreement through ChatGPT Enterprise is the recommended approach.
A. Not in the short term. The "AI surpassing physicians" result on HealthBench Professional is a test of written response quality only. Actual clinical care involves physical examination, patient trust relationships, and managing unexpected complications — areas where AI still struggles.
It is similar to "a student who scores perfectly on written exams not automatically becoming a world-class surgeon." The realistic future is one where the "paperwork and search" portion of a physician's work becomes automated, freeing them to focus on interpersonal care.
A. The U.S. version is legally secure when a HIPAA-compliant BAA is in place. For Japanese users of general ChatGPT, the absolute rule is: never enter directly identifiable information such as names, dates of birth, My Number IDs, EHR patient IDs, or addresses.
The safe approach is to enter only anonymized information, such as "55-year-old male, history of hypertension and diabetes." When in doubt, consult your institution's medical information officer.
A. Final responsibility always rests with the physician. AI is positioned as a "colleague offering reference opinions" — the legal duty to assess the validity of a diagnosis belongs to the physician.
"The AI said so" does not constitute a defense in either Japan or the United States. Even at 99.6% safety, never forget that the system is designed on the premise that physicians must cover the remaining 0.4% error risk.
A. "ChatGPT for Clinicians" is for healthcare professionals only. For general consumers, OpenAI is preparing a separate service called "ChatGPT Health." Regular ChatGPT can also be used for preliminary health research, but it should be limited to general information gathering — any concerning symptoms should always be discussed directly with a physician.
The smart approach is to "use it as a prompt to seek care, not as a definitive answer."
The era of AI freeing physicians from their "paperwork hell" has finally begun.
The winners going forward will not be "physicians whose jobs are taken by AI," but "physicians who master AI to reclaim more time with their patients." In Japan, too, with creative use of general ChatGPT, tenfold efficiency gains in referral letter drafting, literature search, and patient explanations are within reach — starting today.
"One year of being too afraid of AI to try it" is "one year of falling further behind physicians who are already using it."
Be bold — but be careful.
For physicians, ChatGPT is becoming a reliable new colleague.
This article is a cross-post from AI Friends.
