Would you see a doctor who is wrong 30% of the time?

This question is more than fair: if you were sure that a certain doctor gets up to 60% of diagnoses and prescriptions wrong, would you see them? What if it were 30%? Would you go? I think the overwhelming majority of people, once they had this information, would look for another professional—perhaps one about whom this statistical certainty, and all the psychological weight it carries, could not be had.

Human beings love not knowing things. Ignorance is bliss, as the saying goes: not being sure of things, such as the date of our own death, gives us peace. One of the reasons diseases like cancer are so cruel is the suffering of patients who learn that often nothing more can be done and that the only certainty is the imminent end. The suffering of knowing that in three months you will no longer be among the living is greater than the certainty of death itself. We suffer more from knowing than from the reason to suffer; after all, unlike other animals, we have a sharp sense of the future, which is the source of our anxieties.

And it is precisely because of this mix of ignorance and anxiety that millions of people have started consulting artificial intelligences as if they were trusted physicians, despite the high hallucination rates these models produce. When a strange pain strikes at two in the morning, the chatbot is right there: no waiting room, no expensive appointment, no judgment, and an instant answer written in a calm, empathetic, persuasive tone. Yet behind that polished prose there is no clinical reasoning—only a statistical text generator that would rather invent a plausible-sounding answer than admit it has no idea what it is talking about.

What hallucination benchmarks actually show

In doubt? Look at the continuously updated compilation by Suprmind tracking the major independent benchmarks in the field. In the earlier version of this post, I already pointed to a figure that should send a chill down any patient's spine: on Artificial Analysis's AA-Omniscience leaderboard, Claude Fable 5.1, despite holding the highest raw accuracy in the table (68.4% on v1.1 and 67.2% on v1), posted a 72.6% hallucination rate on v1 (and 70.0% on v1.1). In plain English: when the "smartest" model on the leaderboard hits a gap in its knowledge, it guesses wrong more than seven times out of ten instead of simply saying "I don't know."

That gap between "knowing a lot" and "knowing your own limits" is the blind spot that misleads everyday users. When AI labs advertise "hallucination rates of just 2% or 3%," they are almost always pointing to source-grounded summarization benchmarks such as the Vectara Hallucination Leaderboard, where the model is handed the full document up front and only needs to avoid distorting it. In the real world—when a patient types vague symptoms without a complete medical chart—what matters is behavior on open-ended queries. And on that front, the November 2026 numbers tell a very different story across major model families:

• Claude family (Anthropic): for a long time, the industry story was that Claude would rather abstain than fabricate. The 2026 releases broke that reputation in pursuit of leaderboard accuracy: beyond the 70.0%–72.6% of Fable 5.1 and 63.6% of Fable 5, Claude Opus 5.5 posts a 50.3% hallucination rate on v1.1 (58.6% on v1)—well above Claude Opus 4.8 (39.3% on v1), which still leaned toward refusing when uncertain.
• GPT family (OpenAI): in the mid-2026 generation, GPT-5.5 hit 89.0% hallucination in its final v1 record and GPT-5.6 Sol reached a staggering 92.2% (GPT-5.6 Luna hit 92.3% on v1.1): when they did not know the answer, they fabricated one more than nine times out of ten. The newer generation cut back the worst excesses, yet remains far from safe: GPT-6 Astra still hallucinates 49.8% on v1.1 (51.3% on v1), GPT-6.1 Sol sits at 51.9%, and even OpenAI's best-calibrated model on the board, GPT-6.1 Terra, guesses wrong 34.5% of the time when facing uncertainty.
• Gemini family (Google) and open/competing models: Gemini 3 Pro and Gemini 3 Flash reached 91.5% and 93.0% hallucination on v1, and the public Gemini 3.8 Flash still posts 55.1% on v1.1. Among open-weight and competing models, DeepSeek V4 Pro hit 94.1% on v1.1 (94.8% on v1) and DeepSeek V4.1 Flash hit 96.5% on v1. Newer releases like Gemini 4 Argon proved that a model can be trained to stay quiet—dropping to 4.3% hallucination on v1.1—but at a telling cost: it answers only 48.3% of questions correctly and declines most of the rest. On xAI's side, Grok 4.1 Fast scored 73.4% on v1, while Grok 4.7 dropped to 20.0% on v1.1 (at 42.7% accuracy).

As Suprmind's own report notes, 52 of the 124 models on the AA-Omniscience v1.1 board have a negative index—meaning they give more wrong answers than right ones. Worse still, formal impossibility proofs published in 2024 and 2025 (by Xu, Jain, and Kankanhalli, and by Karpowicz), whose conclusions OpenAI itself publicly acknowledged, show that hallucinations are not a temporary bug awaiting the next point release. They are a mathematically inevitable property of architectures that generate language by predicting probable token sequences. You can reduce them with web search, cross-checking, and cautious training; eliminating them is mathematically impossible.

And there is a cruel multiplier known as the confidence paradox: the larger and more articulate the model, the more convincing its lies sound. In an analysis of nearly 18,000 real user questions tracked by Suprmind, the models' assertiveness tone barely budged between trivial prompts and high-stakes decisions (shifting less than 0.09 points on a 1-to-10 scale). A chatbot speaks with the exact same clinical poise whether it is reciting a textbook guideline or inventing a nonexistent dosage.

"Doctor Chatbot": when a hallucination leaves the screen and lands in the ICU

This is not abstract theory. According to OpenAI itself, more than 230 million people ask health and wellness questions on its platform every week—and reporting based on company figures puts usage at over 40 million people every day turning to ChatGPT for health advice. The fallout from consulting "doctors" that fabricate 30% to 90% of what they don't know is already showing up in peer-reviewed medical literature.

The most striking case was published in August 2025 in Annals of Internal Medicine: Clinical Cases under the title A Case of Bromism Influenced by Use of Artificial Intelligence (and covered by The Guardian and Live Science). A 60-year-old man with no prior psychiatric history read about the downsides of table salt (sodium chloride) and asked ChatGPT how to eliminate chloride from his diet. The chatbot suggested sodium bromide as a replacement for chloride—a toxic chemical used in industrial and cleaning applications and long abandoned in medicine for causing severe neurological harm.

Trusting the reply, the patient bought sodium bromide online and ingested it for three months. He arrived at the emergency department with bromism, a 19th-century toxidrome: acute paranoia (believing his neighbor was trying to poison him), auditory and visual hallucinations, severe electrolyte abnormalities, extreme thirst, facial acne, and insomnia. He tried to flee the hospital within 24 hours and required a multi-week involuntary psychiatric admission. When his physicians later asked ChatGPT the same question themselves, it again suggested bromide without issuing a specific health warning or asking why the user wanted to replace dietary chloride—the very first question any human clinician would ask. Even worse, because the "consultation" happened inside a private chat window, doctors initially had no way to trace what had poisoned him.

Anyone assuming newer models or dedicated health features would fix this got a rude awakening in February 2026. Researchers at the Icahn School of Medicine at Mount Sinai published in Nature Medicine the first independent safety evaluation of ChatGPT Health—the dedicated feature OpenAI launched in January 2026 to link medical records and wellness apps. The findings, detailed by The Guardian and reviewed by the National Institutes of Health (PMC/NIH), are alarming:

• Emergencies missed more than half the time: across nearly 1,000 responses to 60 realistic clinical scenarios vetted by three independent physicians, ChatGPT Health under-triaged 51.6% of true emergencies, telling patients experiencing diabetic ketoacidosis, severe asthma flares, or respiratory failure to stay home or book a routine visit. In one simulation, eight times out of ten (84%), it sent a suffocating woman to a future appointment she would not live to attend—while, at the other extreme, sending 64.8% of completely stable individuals to the emergency room.
• Deadly sycophancy (anchoring bias): merely mentioning that a "friend" or relative thought the symptoms were nothing serious made the model nearly 12 times more likely to downplay a medical emergency.
• A suicide guardrail that vanishes next to normal labs: when researchers tested a fictional 27-year-old patient thinking about taking a large number of pills, the crisis-help banner appeared every time. Yet when they pasted normal lab results into the exact same prompt—same patient, same words, same severity—the crisis banner disappeared in 16 out of 16 attempts (0% trigger rate).

Nor is this unique to OpenAI. A multi-model audit published in npj Digital Medicine tested ChatGPT, Gemini, Claude, and Llama on patient-posed medical questions and documented dangerous omissions in obstetric emergencies (such as miscarriage warning signs), unsafe advice telling users to give plain water to infants, and false reassurance treating chest or epigastric pain as "benign heartburn" without checking cardiac history. The NIH/PMC review also documents inaccurate and potentially harmful health advice from Google AI Overviews on pancreatic cancer and liver function tests.

There is also a rarely discussed gulf between passing a medical exam and treating a real human patient. In a randomized study from the University of Oxford published in Nature Medicine (Oxford press release), language models scored 95% accuracy when evaluated in isolation on clinical vignettes; when ordinary members of the public used the exact same chatbots to navigate those symptoms, correct identification plummeted to 35%—no better than a standard web search. The reason is straightforward: laypeople do not know which clinical details matter, and chatbots rarely conduct disciplined history-taking before jumping to confident conclusions. Worse still, when a study in Communications Medicine planted a single fabricated detail inside 300 clinical vignettes, six leading models built their answers on the fake premise in 50% to 82%–83% of cases (even with explicit safety prompting, the average error rate remained a dangerous 44%). Small wonder the independent patient-safety nonprofit ECRI ranked the misuse of AI chatbots as the #1 health technology hazard for 2026.

On this very blog, I wrote in On mechanical parrots about how these models stay so trapped in their training statistics that they ignore explicit instructions whenever + and - diff signs are flipped, and in Whoever writes does not approve, how even for routine programming tasks you have to surround AI with automated tests and skeptical reviewers because it cheerfully reports "all done" on broken code. In software, a failing test breaks the build; in medicine, the patient swallows the hallucination as if it were a prescription.

When hallucinations leave the clinic and reach the control tower: yesterday's FAA outage

If the fallout stopped at home health chats, it would already be a crisis. Instead, the same commercial rush—pitching flawed probabilistic models as infallible replacements or "copilots"—is spreading into domains where the tolerance for error should be zero.

In law, Damien Charlotin's database now tracks 2,266 court rulings worldwide (1,535 in U.S. courts) where a party relied on AI-hallucinated material—including attorneys for the U.S. Department of Health and Human Services (HHS) who, in October 2026, filed exceptions containing 22 fabricated, erroneous, or irrelevant footnotes out of 29. Nor is it just careless free-tier ChatGPT users: a Stanford study tested commercial legal AI tools explicitly marketed as "hallucination-free" and found error rates from 17% (Lexis+ AI and Ask Practical Law AI) to 33% (Westlaw AI-Assisted Research). Even top academic peer review is cracking: an audit of NeurIPS 2025 turned up 100 phantom AI-generated citations across 53 accepted papers.

Yet nothing illustrates the recklessness of bolting AI onto critical infrastructure quite like what happened yesterday, Friday, November 13, 2026, across U.S. airspace.

Earlier this year, the FAA (Federal Aviation Administration) signed a 12-year, $875 million (nearly $1 billion) contract with software firm Air Space Intelligence (ASI) to deploy an AI platform called SMART (Strategic Management of Airspace, Routes and Trajectories), publicly billed as a predictive "brain" for air traffic control that analyzes schedules, weather, and runway capacity to reroute flights and ease bottlenecks (Futurism, Forbes, FAA SMART One-Pager). This week, after initial deployment in Washington, D.C., the FAA expanded the system nationwide across its 22 regional Air Route Traffic Control Centers (ARTCCs).

Yesterday, just days after that nationwide AI rollout, a severe telecommunications and data-inconsistency failure hit all 22 regional centers mid-afternoon. The FAA was forced to issue an extraordinary nationwide ground stop, halting departing flights across the United States for more than four hours—only the second nationwide grounding since the September 11, 2001 attacks, and the longest since then. According to FlightRadar24 data cited by CBC News, the immediate toll was staggering: more than 6,600 flights canceled, 8,600 delayed, and hundreds of aircraft diverted mid-air (PBS NewsHour, Futurism, Inc.).

Intellectual honesty requires stating the exact sequence of claims. Last night, asked at the White House about the striking timing, Homeland Security Secretary Markwayne Mullin initially said investigators were looking into whether there was any link to the AI rollout. Hours later, both Air Space Intelligence and federal officials stated that SMART did not cause the outage: the vendor categorically denied any involvement, and Transportation Secretary Sean Duffy blamed antiquated routers and switches that became overloaded with data packets (Business Insider, CBC News).

Even accepting the official explanation that aging hardware—not the SMART algorithm—triggered the shutdown, yesterday's breakdown exposes a dangerous inversion of priorities and a systemic risk nobody should brush off. Look at the picture officials themselves just painted: U.S. air traffic control suffers from a chronic shortage of thousands of human controllers and runs on telecom gear so outdated that the Transportation Secretary immediately demanded $50 billion from Congress to replace switches and copper wiring. Yet the marquee initiative was pouring nearly $1 billion into an AI layer on top of fragile legacy infrastructure—over the skepticism of aviation veterans who had already urged the FAA to narrow the rollout's scope.

Worse than network overload is the cognitive and operational hazard ahead. As Futurism noted, roughly 5,000 aircraft can be airborne over the U.S. at any given moment. When overworked controllers start staring at screens fed by opaque probabilistic models marketed as predictive optimization, automation bias sets in—the human tendency to trust the machine's dashboard rather than challenge it in time. In a home symptom chat, a 30% error rate or a 70% hallucination rate quietly harms or kills one patient at a time; in an air traffic control center or power grid, statistical guesswork leaves zero margin for a second opinion.

Why self-regulation has failed—and what serious regulation must demand

In light of all this, the promise that "the market and AI labs will self-regulate" has already collapsed against the evidence. That is why there is an urgent need for clear regulation covering the research, development, and deployment of these models—to protect users from predictable harm, especially when misuse or miscalibration can injure them or put entire public systems at risk.

Serious regulation does not mean banning research or stifling innovation; it means closing four glaring loopholes that currently operate as a legal Wild West:

1. End the regulatory loophole for general-purpose models: dedicated clinical software or AI-enabled medical hardware must clear strict regulatory approval (such as FDA or Anvisa certification) before touching a patient. Meanwhile, general-purpose chatbots—and hybrid offerings like ChatGPT Health—dispense medical triage to tens of millions of people daily behind a fine-print disclaimer ("not a substitute for professional medical advice") while public marketing sells a 24/7 health companion. Any system that functions in practice as clinical triage or therapeutic guidance should meet medical-device safety standards.
2. Mandatory independent auditing and real transparency before launch: benchmarks from Artificial Analysis and Suprmind show that vendor-run tests routinely contradict independent evaluations—as seen across gaps in models like Grok 4.6, Claude Haiku 5.5, and Tencent Hy4—or quietly hide regressions when labs trade calibration for leaderboard accuracy (as in the shift from Claude Opus 4.8 to Opus 5 and Fable 5.1). Models deployed in healthcare, law, finance, or public infrastructure cannot be graded exclusively by the marketing teams selling them.
3. No silent replacement of human specialists and no AI band-aids over decaying infrastructure: probabilistic software cannot replace accountable human judgment in medical emergencies, courtrooms, or air traffic control—nor serve as a shiny budget excuse to leave controller rosters understaffed and core communications hardware crumbling, as yesterday's FAA breakdown laid bare. High-stakes systems require deterministic audit trails, full decision logging, and qualified humans empowered to override algorithmic prompts.
4. A legal duty of calibration and proportional civil liability: when a model is trained to guess confidently 70% or 90% of the time it does not know an answer, that is not an act of nature—it is a training choice. The leap from Gemini 3 Pro (91.5% hallucination) to Gemini 4 Argon (4.3%), or the more cautious refusal profile of models like Claude Opus 4.8, proves that labs can teach models to recognize uncertainty and abstain, even if doing so lowers raw answer counts on vanity leaderboards. The law must reward safe abstention and hold companies civilly liable when they sell tools known to encourage self-harm, toxic ingestion, or fatal misdiagnosis.

Let us return to the title question. You would never walk into a doctor's office if a sign on the door warned that they make up 30% to 70% of what they don't know with a reassuring smile—and you certainly wouldn't board a plane knowing the control tower relied on statistical guesswork of that order. Demanding clear rules, independent audits, and hard boundaries on where this technology may operate is not anti-progress: it is basic common sense so that the convenience of an instant answer stops costing human lives.

Sources and references

• Hallucination comparison data and cross-benchmark report (AA-Omniscience, Vectara, FACTS, HalluHard): AI Hallucination Rates, Statistics & Benchmarks in 2026 — Suprmind
• AA-Omniscience leaderboard: AA-Omniscience Evaluation — Artificial Analysis
• Grounded summarization faithfulness leaderboard: Vectara Hallucination Leaderboard (GitHub)
• Mathematical impossibility of zero hallucinations: OpenAI admits AI hallucinations are mathematically inevitable — Computerworld
• ChatGPT-induced bromism clinical case: A Case of Bromism Influenced by Use of Artificial Intelligence — Annals of Internal Medicine: Clinical Cases | The Guardian | Live Science
• ChatGPT Health safety evaluation on emergencies and suicide risk: ChatGPT Health performance in a structured test of triage recommendations — Nature Medicine | The Guardian | Introducing ChatGPT Health — OpenAI
• Real-world patient use vs. isolated benchmark scores (Oxford): Reliability of LLMs as medical assistants for the general public: a randomized preregistered study — Nature Medicine | University of Oxford
• Multi-model unsafe medical answers (ChatGPT, Gemini, Claude, Llama) and vulnerability to planted false details: Large language models provide unsafe answers to patient-posed medical questions — npj Digital Medicine | Multi-model assurance analysis showing large language models are highly vulnerable to adversarial hallucination attacks during clinical decision support — Communications Medicine
• Broad review on chatbot health risks and Google AI Overviews: The risks of AI-generated health advice — PubMed Central / NIH
• 2026 hospital and patient safety ranking: Misuse of AI Chatbots Tops Annual List of Health Technology Hazards — ECRI
• Hallucinations in law and commercial legal AI tools: AI Hallucination Cases Database — Damien Charlotin | AI on Trial: Legal Models Hallucinate in 1 out of 6 (or More) Benchmarking Queries — Stanford HAI
• FAA SMART AI system ($875M contract): SMART One-Pager (PDF) — FAA | The FAA Is Spending Nearly a Billion Dollars to Make AI the "Brain" of Air Traffic Control Towers — Futurism | FAA Bets $875 Million On AI To Reduce Flight Delays — Forbes
• U.S. nationwide FAA ground stop incident (Nov 13, 2026): FAA grounds all flights at U.S. airports due to air traffic control communications problems — PBS NewsHour | U.S. Federal Aviation Administration lifts nationwide ground stop after 4 hours — CBC News | Entire US Airspace Shut Down Days After the FAA Rolled Out an AI Air Traffic Tool — Futurism | FAA SMART AI System Vendor, Microsoft Say Ground Stop Wasn't Their Fault — Business Insider | This Week, the FAA Rolled Out AI Air Traffic Management. Then It Grounded Flights Nationwide — Inc.