600,000 People Ask ChatGPT About Their Health Every Week — And 70% Do It After Hours

OpenAI’s data reveals a healthcare system failure hiding in plain sight.

The numbers are striking: 600,000 health-related queries flow through OpenAI’s systems every single week. But the pattern behind those queries tells an even more important story. Seventy percent happen after traditional healthcare hours — evenings, weekends, the middle of the night.

This isn’t people asking about celebrity diets or workout tips. OpenAI specifically identified these queries as coming from “hospital deserts” — areas where access to medical care is limited or nonexistent. People are asking whether they should drive hours for emergency care. They’re trying to decide if their symptoms warrant a trip they can’t afford, or if they can wait until morning.

The AI isn’t replacing doctors. It’s filling a vacuum the medical system created.

The Scale of the Phenomenon

OpenAI’s health query data is just one window into a much larger trend. Microsoft reports that its Copilot platform receives 50 million health-related questions every day. Health is the most popular topic on the Copilot mobile app — more than weather, news, or entertainment.

Karan Singhal, who leads OpenAI’s Health AI team, says the company noticed a rapid rise in health queries even before building dedicated health products. The demand was already there. Users were already treating chatbots as medical advisors. OpenAI just made it official.

This surge isn’t driven by technophiles or early adopters. It’s driven by people who can’t get answers any other way.

Hospital Deserts and Healthcare Inequality

The term “hospital desert” describes areas where residents lack reasonable access to emergency medical care. In the United States, an estimated 30 million people live in such areas. Rural communities are particularly affected — hospitals have been closing at an accelerating rate, leaving gaps that can span hundreds of miles.

The consequences are measurable. People in hospital deserts face higher mortality rates from treatable conditions. Emergency response times stretch into hours rather than minutes. And the psychological toll of knowing help is far away drives people to seek alternatives.

AI chatbots have become that alternative. They’re available 24 hours a day. They don’t require insurance. They don’t judge. And they respond instantly — even at 2am when a parent is deciding whether their child’s fever warrants a three-hour drive to the nearest emergency room.

The After-Hours Pattern

The 70% after-hours figure is the most telling detail. It reveals that people aren’t using AI as a convenience — they’re using it as a last resort when traditional options are closed.

Healthcare doesn’t operate on a 24/7 schedule for most people. Urgent care centers close. Doctor’s offices don’t take calls at midnight. Telemedicine services often have limited hours or require appointments. The emergency room is always open, but it’s expensive, intimidating, and potentially far away.

AI fills the gap. It provides something that feels like medical guidance when nothing else is available. The question is whether it provides good guidance.

What the Research Shows

The evidence on AI health advice is mixed — and that’s being generous.

Google’s study of its Articulate Medical Intelligence Explorer (AMIE) chatbot showed diagnoses matching physician-level accuracy in controlled settings. None of the test conversations raised major safety concerns. This suggests that, in principle, AI can provide useful health guidance.

But other studies paint a more concerning picture. A Mount Sinai study found that ChatGPT Health sometimes over-recommends care for mild conditions — potentially driving unnecessary medical visits. Worse, it occasionally fails to flag genuine emergencies. The tool that tells you to relax about chest pain might be right 99 times out of 100, but the 100th case is the one that matters.

All six academic experts consulted by MIT Technology Review agreed on two points: AI health tools have real potential for underserved populations, and they’re being released without adequate independent testing.

The Liability Problem

Here’s where the story gets complicated. Every major health AI platform carries disclaimers warning users not to seek diagnoses from chatbots. The terms of service are clear: this is not medical advice.

But as Adam Rodman, a physician and researcher at Beth Israel Deaconess Medical Center, notes bluntly: “We all know that people are going to use it for diagnosis and management.”

The disclaimers protect the companies. They don’t protect the users. And they certainly don’t address the underlying problem: people need medical guidance, and they’re getting it from the only source available.

The liability question becomes acute when things go wrong. If someone delays seeking care because a chatbot minimized their symptoms, who bears responsibility? The user who “should have known better”? The company that provided the tool? The healthcare system that wasn’t there when needed?

The Testing Gap

Tech companies test their own products extensively. OpenAI developed HealthBench, a benchmark scoring LLM responses to realistic health conversations. GPT-5 scored significantly better than previous models — though well short of perfect.

But internal testing has blind spots. Oxford researcher Andrew Bean found that even when an LLM correctly identifies a condition, a non-expert user asking the same question with LLM assistance arrives at the right answer only about one-third of the time. Users omit key details from their prompts. They misread responses. They lack the medical knowledge to interpret what the AI is telling them.

Bean argues that no matter how rigorous a company’s internal research is, it cannot replace independent evaluation. “The evidence base really needs to be there,” he says.

Stanford’s MedHELM framework offers one of the more comprehensive third-party evaluation systems. OpenAI’s GPT-5 holds the top MedHELM score. But even MedHELM evaluates only single responses — not the multi-turn conversations that real users typically have with health chatbots.

The Google Counterexample

Google’s approach offers an instructive contrast. Despite its AMIE chatbot producing strong results in testing, the company has decided not to release it publicly yet. It cites unresolved issues in equity, fairness, and safety testing.

This stands in contrast to competitors who launched first and tested later. Google’s caution suggests that responsible deployment is possible — it just requires accepting that someone else might capture the market while you verify your product works safely.

The Real Question

The debate about AI health tools often focuses on whether they’re good enough to replace doctors. This misses the point.

For millions of people in hospital deserts, there is no doctor to replace. The choice isn’t between AI and a physician — it’s between AI and nothing. Between guidance that might be flawed and no guidance at all.

No expert interviewed demanded perfection from health AI before release. Doctors make mistakes too. For someone with rare access to a physician, a chatbot that is sometimes wrong may still be a meaningful upgrade over having no guidance at all — as long as its errors stay manageable and rare.

The current evidence doesn’t confirm whether available tools improve outcomes or create new risks. More rigorous, independent testing is the only way to answer that question with confidence.

But the testing needs to happen in the context of reality: a healthcare system that leaves 30 million people without reasonable access to emergency care, and where 600,000 people every week turn to AI because they have nowhere else to go.

The Stakes Just Changed

OpenAI’s data reveals something important about the AI healthcare moment. This isn’t a technology story about better algorithms or larger models. It’s an access story about a system that isn’t meeting people’s needs.

The 600,000 weekly queries aren’t a validation of AI’s medical capabilities. They’re an indictment of healthcare’s geographic and economic accessibility. The 70% after-hours pattern isn’t a convenience metric — it’s a signal of desperation.

AI companies are stepping into a gap the medical system created. Whether they can do so responsibly — with proper testing, appropriate humility about limitations, and genuine accountability when things go wrong — will determine whether this moment becomes a genuine expansion of healthcare access or a cautionary tale about technology rushing in where institutions have failed.

The people asking ChatGPT about their symptoms at 2am don’t care about benchmark scores or evaluation frameworks. They care about whether they need to drive three hours to an emergency room, or whether they can wait until morning.

They’re getting answers from the only source that answers. Whether those answers are good enough is a question we haven’t adequately answered — even as millions of people are already acting on them.

Related Reading

• Foxconn’s $66.6 Billion Quarter Proves AI Infrastructure Is the Real Money — The infrastructure behind AI’s healthcare expansion
• Are We Being Trained by Our Own AI Constructs? — How AI tools reshape human decision-making
• Anthropic Launches AnthroPAC — When AI safety companies enter the political arena

Sources

1. OpenAI Launches ChatGPT Health — Medical Economics 2. AI Health Tools: Promise, Risks, and Gaps — DistilINFO 3. Why People With Chronic Illness Are Turning to AI Chatbots — The New York Times 4. OpenAI for Healthcare Aims to Streamline Clinical Workflows — AJMC 5. Medical AI’s 1% Problem: Why Billions Still Await Revolution — Forbes