Doctors at Beth Israel Deaconess Medical Center in Boston tried something new last year. Nearly 100 patients scheduled urgent primary care visits. Then they chatted at home with a Google AI system called AMIE days before seeing their physician.

The bot asked questions. It explored symptoms. It listed possible causes and treatments. It produced a summary for the doctor. Results surprised many. The technology proved helpful in three-quarters of cases. It shaped how physicians approached care in more than half. Physicians described the performance as on par with a strong medical resident.

But this success arrives against a backdrop of sharp warnings. Other AI health tools have stumbled badly on emergencies. OpenAI’s ChatGPT Health, for instance, undertriaged over half of serious cases in independent tests. The gap between promise and peril has never looked wider.

The Beth Israel study, published Thursday in The Lancet, marks one of the first times a conversational diagnostic AI faced real patients in a live clinical workflow. Researchers from Google and the hospital worked together. They enrolled 114 adults between April and November 2025. Ninety-eight completed both the AI chat and the appointment. Supervising doctors monitored every exchange in real time.

Not one conversation triggered a safety stop. That fact stands out. The system never crossed predefined red lines that would require immediate human takeover. Supervisors spotted just one hallucination across hundreds of turns. They added clarifying details in five others. Adam Rodman, director of AI programs at BIDMC’s Carl J. Shapiro Institute, called the work an effort to establish baseline characteristics of real-world AI conversations.

Peter Brodeur, a resident physician at the time and co-first author, reviewed transcripts with colleagues. “They compared it to a high-performing medical student or resident,” he said. The AI’s differential diagnosis included the correct final condition in its top seven possibilities 90 percent of the time. The top three matched 75 percent. The single top guess landed right 56 percent. When later confirmed by diagnostic tests, accuracy held up.

Physicians received transcripts and summaries before seeing patients. In post-visit surveys covering 60 cases, doctors said the materials helped them prepare in 75 percent of encounters. The information may have changed their management plan in more than half. Patient trust in AI rose after the experience. Acceptance of the technology improved.

The findings contrast sharply with recent examinations of consumer-facing tools. A February study in Nature Medicine tested OpenAI’s ChatGPT Health on 60 clinician-written vignettes. The system undertriaged 52 percent of gold-standard emergencies. Cases of diabetic ketoacidosis or impending respiratory failure often received advice to seek care in 24 to 48 hours instead of heading to the emergency department. Textbook events like stroke triggered correct responses. Nuanced presentations did not.

That trial, led by Mount Sinai researchers, raised immediate safety questions. It appeared weeks after OpenAI launched the dedicated health chatbot in January 2026. The company reported 40 million daily health queries to its models. Many occurred outside clinic hours. Users now upload records from Epic systems, Apple Health and wellness apps. Yet the terms still declare the service is not for diagnosis or treatment.

Similar caution appears in regulatory reviews. A meta-analysis in BMC Emergency Medicine examined 11 studies with over 3,000 cases. Large language models showed moderate ability to spot highest-acuity patients but limited sensitivity. Pooled sensitivity for critical cases reached only 61 percent. Specificity stood at 97 percent. Authors concluded current models are not ready for standalone emergency triage.

Health systems have moved anyway. Britain’s NHS began rolling out an AI triage feature in its app this year. Early pilots cut telephone queues by 29 percent at one practice. The tool asks structured questions then directs patients to GP appointments, pharmacies, emergency care or self-management. NHS England aims to reach all app users by April 2028. Rapid Health, the company behind the Smart Triage system, reported the upgrade now serves practices covering three million patients. Carmelo Insalaco, its chief executive, said the project proves digital gains need not wait for later targets.

In the United States, commercial offerings multiply. Amazon launched a health AI assistant for One Medical members in late September. It pulls from medical records to answer questions, suggest care settings and book visits. Guardrails route complex cases to clinicians. OpenAI expanded ChatGPT Health to all U.S. users over 18. Anthropic and others followed with their own medical features. A McKinsey analysis from September estimated AI could handle 16 to 22 percent of U.S. outpatient care today. One in three adults already consults chatbots for symptoms or test explanations monthly.

Yet enthusiasm collides with documented failures. A Florida pastor sued OpenAI in July after ChatGPT allegedly downplayed symptoms of a pulmonary embolism. The suit claims the model discouraged medical consultation. It also questioned the reliability of crisis safeguards. Separate research found rules-based triage chatbots still outperform large language models on accuracy even if users prefer the conversational style of AI.

Google’s AMIE differs in training. The company built it specifically for clinical reasoning rather than general search. A more powerful Gemini variant powers the system. In the Beth Israel trial, AMIE generated summaries that doctors found valuable for preparation. The AI explored patient histories conversationally. It offered differential lists for discussion during the visit. And it avoided the overtriage or dangerous undertriage seen elsewhere in some scenarios.

Still, authors stress limits. The study was small, single-center and focused on feasibility. All patients had already booked appointments. No one relied on the AI for final decisions. Larger trials must test scalability, diverse populations and long-term outcomes. Safety supervisors remained in the loop. Real deployment without such oversight would look different.

Physicians themselves face changing roles. Some report AI makes them better at spotting overlooked patterns. Others worry trainees lean too heavily on tools and lose diagnostic muscle. A New York Times opinion piece in September captured the tension. One doctor described using AI to select antibiotics, interpret labs and consider rare conditions. The same article questioned whether early exposure dulls the reasoning skills students need to develop.

Insurance battles add another layer. Hospitals use AI to identify more secondary diagnoses and submit richer claims. Insurers deploy their own systems to challenge those codes. A Blue Cross Blue Shield report tied hospital AI coding changes to nearly $1 billion in extra costs over two years. Patients land in the middle with surprise bills or denied care. The risk of bots fighting bots grows.

Google published its first-ever Lancet paper on the AMIE results. The company’s research blog highlighted improved patient-physician relationships as a potential outcome. Summaries let doctors enter visits better informed. Conversations focus faster on what matters. Clinicians gain time for complex cases. But only if the technology earns trust across the board.

Recent deployments offer mixed signals. Manchester GP practices tested AI receptionists that handle calls, produce summaries and flag safety issues. Ninety-five percent of calls resolved same-day. Receptionist workload dropped by half. Patient reactions split. Some appreciated speed. Others preferred human contact. Similar agentic AI pilots at hospitals analyze feedback or monitor recovery in real time.

The BIDMC trial adds concrete data to this picture. Zero safety stops. High diagnostic overlap. Physician approval in most cases. Patient attitudes shifting positive. These numbers matter for an industry strained by demand. Urgent care clinics overflow. Primary doctors burn out. Wait times stretch. Tools that safely offload initial information gathering could ease pressure without sacrificing quality.

But the path forward demands caution. The same week the Lancet paper appeared, other headlines reminded readers of AI’s blind spots in high-stakes medicine. One system flags hospital patients at risk of malnutrition or delirium from record notes. Another helps interpret ECGs for hidden heart damage. Each arrives with guardrails and calls for validation. None claims to replace judgment.

So the AMIE results feel like progress. Not transformation. Not yet. They show conversational AI can sit inside real workflows, earn clinician respect and avoid obvious harm in a carefully supervised setting. That combination has been rare. It deserves attention from health system leaders, regulators and technology builders alike.

Further studies will test AMIE and competitors in broader populations. They must measure hard outcomes. Reduced emergency visits. Faster diagnoses. Lower costs. Improved satisfaction. Until those answers arrive, adoption will remain measured. The gap between lab success and bedside reliability is exactly what this trial begins to close.