The History of AI in Healthcare Didn’t Start With ChatGPT

Every headline about generative AI in medicine makes it sound like this is the first time software has tried to think like a clinician. It isn’t. The history of AI in healthcare stretches back more than 50 years, through at least two “AI winters,” a $62 million cautionary tale, and a slow regulatory grind that only recently produced more than a thousand FDA-cleared AI devices. Understanding that arc matters because the same failure patterns (workflow mismatch, unvalidated performance claims, and hype outrunning evidence) keep reappearing with each new wave of technology.

Quick Reference: Key Eras in the History of AI in Healthcare

Era	Approx. Years	Defining System or Event	Why It Mattered
Expert systems	1970s-1980s	MYCIN, INTERNIST-1	Proved rule-based reasoning could rival specialists, but was never clinically deployed
AI winter	Late 1980s-1990s	Funding and interest collapse	Expert systems proved too rigid and costly to maintain
Quiet machine learning years	1995-2011	First FDA-cleared AI device (1995)	AI entered devices slowly, mostly in imaging and lab screening
Deep learning breakthrough	2012-2016	AlexNet, imaging AI	Neural networks made image-based diagnosis practical
Big hype, big failure	2011-2017	IBM Watson for Oncology	$62M MD Anderson project shelved without treating a patient
Autonomous AI arrives	2018	FDA authorizes IDx-DR	First AI diagnostic allowed to operate without physician review
Predictive analytics caution	2019-2021	Epic Sepsis Model validation studies	Real-world performance far below vendor marketing claims
Generative AI era	2022-present	ChatGPT, Med-PaLM 2, GPT-4	LLMs pass medical exams and enter documentation, but none are FDA-cleared as devices yet

The Expert System Era: How MYCIN Launched AI in Medicine

The most cited starting point in the history of AI in healthcare is MYCIN, built at Stanford in the early 1970s. MYCIN was designed to diagnose infectious blood diseases and recommend antibiotics, adjusting dosage for a patient’s body weight, and it operated using a knowledge base of roughly 500 hand-coded rules. AI pioneer Allen Newell called it “the granddaddy” of all expert systems, “the one that launched the field”.

MYCIN’s diagnostic performance was genuinely impressive for its era. Research at Stanford Medical School found MYCIN had a correct diagnosis rate of about 65%, better than non-specialist physicians, though still behind infectious disease experts averaging around 80%. Yet MYCIN was never used on real patients. The reasons echo through every subsequent AI failure in medicine: workflow integration was clumsy on 1970s mainframes, and nobody had resolved who would be liable if the software’s recommendation harmed a patient. Around the same time, Rutgers and Pittsburgh researchers built INTERNIST-1, a ranking-based diagnostic tool for general internal medicine, and the NIH funded early networking infrastructure like the SUMEX-AIM system to connect biomedical AI researchers. None of these systems reached routine clinical use, and by the late 1980s funding for symbolic, rule-based AI collapsed into what researchers call an AI winter.

The Quiet Machine Learning Years and the First FDA-Cleared AI Devices

Between the expert-system bust and the deep learning boom, AI in healthcare didn’t disappear, it just went quiet and shifted toward statistical machine learning and digitized data. The FDA’s involvement started earlier than most people assume: the agency’s own device list shows the first AI/ML-enabled device was cleared in 1995, an AI-enabled cervical smear rescreening tool, but approvals stayed rare for two decades. Only 33 AI/ML devices had been authorized by 2016, before the pace of clearances began accelerating sharply. This period also saw broader digitization of health records and biomedical literature (including PubMed) that later became the raw material for machine learning models. For a look at how those foundational tools evolved into today’s diagnostic software, see our breakdown of what AI in healthcare diagnosis can and can’t do yet.

The Deep Learning Breakthrough That Reshaped Medical Imaging AI

The modern era of AI in healthcare really begins outside medicine entirely, with a computer vision contest. In 2012, a convolutional neural network called AlexNet won the ImageNet Large Scale Visual Recognition Challenge with a 15.3% error rate, nearly halving the previous year’s results. That result convinced researchers across fields, including radiology and pathology, that neural networks trained on large image datasets could outperform older rule-based computer vision. A gastrointestinal endoscopy history review notes that many of the limitations that held back AI in medicine were overcome in the early 2000s by the advent of deep learning, setting up the imaging AI wave that followed through the 2010s.

Not every deep learning-era experiment succeeded. Google Flu Trends, launched in 2008 to predict flu outbreaks from search queries, was an early example of big data optimism in public health before it was quietly discontinued in 2015 after its predictions drifted from actual case data. The lesson repeated a decade later with predictive clinical models, discussed below.

IBM Watson and the Costliest Lesson in AI Hype for Healthcare

No history of AI in healthcare is complete without IBM Watson for Oncology, still the most-cited example of AI hype outrunning clinical reality. After Watson’s 2011 Jeopardy! win, IBM pivoted the technology toward cancer care, partnering with MD Anderson Cancer Center in 2012 to build the Oncology Expert Advisor. MD Anderson canceled the project in 2016 after spending $62 million on it, and physicians who tested the prototype found the software struggled to fit the messy reality of real oncology workflows. A later audit found the deeper problem: Watson was never successfully integrated into MD Anderson’s Epic electronic health record system, forcing manual data entry and making the tool unworkable in routine care. It’s a pattern worth remembering any time a vendor promises an AI system will “revolutionize” a service line before workflow integration has been proven. We cover more of these patterns in our roundup of real risks of AI in healthcare, from bias to deployment errors.

FDA Autonomous AI Clearance: The 2018 Turning Point

A genuine regulatory milestone arrived in April 2018, when the FDA authorized IDx-DR (now LumineticsCore), a system that screens for diabetic retinopathy. On April 11, 2018, the FDA permitted marketing of the first device to use AI autonomously to detect a medical condition, meaning its results didn’t require review by a specialist before a primary care clinician could act on them. In its pivotal trial, the system demonstrated 87% sensitivity and 90% specificity in detecting referable diabetic retinopathy. That single clearance opened a regulatory pathway that other autonomous imaging tools have since followed.

The growth since then has been steep. As of July 2025, the FDA’s public database listed over 1,250 AI-enabled medical devices, up from 950 a year earlier, and industry trackers report the cumulative total reaching roughly 1,451 devices by the end of 2025, with radiology accounting for about three-quarters of listings. That volume can create a false sense of maturity, though: most of these are narrow predictive or image-analysis tools, not the conversational systems patients now associate with “AI.”

Predictive AI’s Reality Check: What the Epic Sepsis Model Taught Hospitals

Not every widely deployed clinical AI tool held up under scrutiny. The Epic Sepsis Model, built into EHRs at hundreds of US hospitals, promised early warning of sepsis onset. An external validation at Michigan Medicine covering nearly 38,500 hospitalizations found a hospitalization-level sensitivity of just 33%, alongside a positive predictive value of only 12%, far below Epic’s internally reported performance. Researchers also found the model missed 67% of sepsis patients despite alerting on 18% of all hospitalizations, creating substantial alert fatigue. It’s a concrete, recent reminder that a tool being widely deployed is not the same as a tool being independently validated, a theme we explore further in our piece on where the real gap sits between operational efficiency claims and clinical AI performance.

The Generative AI Era: LLMs, USMLE Scores, and Where Things Stand Now

The current chapter in the history of AI in healthcare began with large language models trained on medical text. Google’s Med-PaLM, described in a peer-reviewed Nature paper, was the first AI system to obtain a passing score on USMLE-style questions from the MedQA dataset, with an accuracy of 67.6%. Its successor pushed further: Med-PaLM 2 improved on this with a state-of-the-art performance of 86.5%, and independent benchmarking found GPT-4 with a specialized prompting strategy reaching 90.2% on the same style of questions. These are impressive test-taking numbers, but passing a licensing exam is a different task from managing an actual patient, and Google researchers themselves have been careful to frame early testing as limited to low-stakes administrative tasks rather than clinical or patient-facing use.

Generative AI’s clinical footprint today is real but narrower than the headlines suggest. It shows up in ambient documentation tools, patient messaging drafts, and prior authorization letters, areas we cover in depth in our guides to generative AI in healthcare tools and adoption and AI medical scribes and what can go wrong. Regulators were notably cautious about the technology’s highest-risk uses for years, but that started to shift: in December 2025, the FDA cleared UpDoc’s insulin-management platform, the first authorized medical device built on patient-facing large language models. Still, that remains the exception rather than the rule. The vast majority of the FDA’s 1,400-plus AI-enabled devices are narrower predictive or image-analysis tools, not conversational systems. The gap between generative AI’s visibility and its regulatory footprint is narrowing, but slowly, and that gap is arguably the defining tension of this current era.

What Five Decades of AI in Healthcare History Actually Teaches Us

Look across MYCIN, Watson for Oncology, the Epic Sepsis Model, and today’s LLMs, and a pattern holds steady: technical capability has never been the limiting factor. Workflow fit, independent validation, liability clarity, and honest performance reporting have been the actual bottlenecks every single time. That’s worth remembering the next time a vendor pitch sounds like it’s promising something entirely unprecedented. For a broader, balanced look at where AI genuinely helps and where it still falls short, our guide to AI in healthcare pros and cons and our roundup of 12 real applications of AI in healthcare are good next reads.

FAQ

What was the first AI system used in healthcare?

MYCIN, built at Stanford in the early 1970s, is widely considered the founding expert system in medicine, designed to recommend antibiotic therapy for infections. It performed well in testing but was never adopted into routine clinical practice because of liability concerns and workflow integration problems.

When did the FDA first clear an AI-based medical device?

The FDA cleared its first AI/ML-enabled medical device, a cervical smear rescreening tool, in 1995, though approvals stayed rare for two decades. The pace accelerated sharply after 2016, and by mid-2025 the FDA’s public list topped 1,250 authorized AI-enabled devices.

Why did IBM Watson for Oncology fail?

MD Anderson Cancer Center spent roughly $62 million on its IBM Watson oncology project before shelving it, largely because the system never integrated cleanly with the hospital’s electronic health record and struggled with real-world clinical workflows. It became a widely cited cautionary tale about the gap between AI demos and deployable clinical tools.

Has generative AI been cleared by the FDA for clinical use?

For years, no FDA-authorized medical device relied on generative AI or large language models. That changed in December 2025, when the FDA cleared UpDoc’s diabetes management platform, the first authorized device built on patient-facing large language models. Most current clearances still cover predictive and image-analysis models rather than generative tools.