Pathologists Are Wrong More Often Than We Think—AI Is Changing the Numbers
Every year, thousands of cancer diagnoses are revised after a second look. AI systems reviewing the same slides are catching the discrepancies before patients undergo the wrong treatment. The question of what to do with that information is not a technical one.
In 2022, a woman in her mid-forties in suburban Chicago was diagnosed with ductal carcinoma in situ (DCIS) of the left breast after a core needle biopsy. She underwent a lumpectomy. The surgical pathology report came back confirming the diagnosis. She began hormone therapy and consulted with a radiation oncologist about adjuvant radiotherapy. Three months later, a routine review of her case at a multidisciplinary tumor board—a weekly meeting where oncologists, surgeons, pathologists, and radiologists at the hospital reviewed active cancer cases together—flagged a discrepancy. A second pathologist, asked to re-examine the original biopsy slides as part of the board's quality review, noted features that were more consistent with atypical ductal hyperplasia (ADH), a benign but abnormal growth pattern, than with true DCIS. The diagnosis was revised. The treatment plan was reconsidered.
This is not a rare event. A landmark study published in the Journal of Clinical Oncology in 2019—still one of the most comprehensive looks at diagnostic agreement in pathology—found that for challenging breast biopsies, the concordance rate between independent pathologists reviewing the same slides was only 75 percent. For some categories of precancerous and early-stage breast lesions, agreement fell to 48 percent. The implications are stark: a significant fraction of patients receiving a cancer diagnosis based on these slides are receiving a diagnosis that at least one other qualified pathologist would disagree with.
The Concordance Problem
Pathology is the foundation of cancer diagnosis. The pathologist's examination of a tissue specimen under a microscope is the definitive moment in determining what kind of cancer a patient has, how aggressive it appears to be, and what stage it has reached. This information determines everything that follows: surgery versus chemotherapy, aggressive versus conservative treatment, life-altering interventions versus watchful waiting. The stakes could not be higher.
And yet pathology is profoundly human. It depends on a trained expert visually examining a stained piece of tissue on a glass slide, recognizing patterns based on years of accumulated experience, and translating those patterns into a diagnostic category. This process is susceptible to every form of human error and bias: fatigue, distraction, anchoring bias (sticking with an initial impression even when subsequent information suggests otherwise), the influence of clinical context (knowing a patient is seriously ill can subtly shape how a pathologist reads ambiguous features), and simple variability in what different pathologists have been trained to see.
The data on diagnostic error rates in pathology has been accumulating for decades, and it is consistently uncomfortable. A 2015 study by the College of American Pathologists that reviewed over 6,000 cases found a major discrepancy rate—the percentage of cases where a significant diagnostic finding was missed or misidentified—in approximately 1.4 percent of cases reviewed. This sounds small until you consider that there are approximately 1.9 million new cancer diagnoses in the United States annually. A 1.4 percent error rate means roughly 26,600 patients per year receiving a materially incorrect initial diagnosis.
In other areas of pathology—particularly gynecologic cytology (Pap smears) and gastrointestinal biopsies—the published error rates are substantially higher. A landmark study in the New England Journal of Medicine from 2018 found that in a cohort of women whose Pap smears had been read as normal, subsequent colposcopy revealed that 8.4 percent had high-grade cervical lesions that had been missed on the initial cytology review. These are not academic numbers. These are women who were told they were healthy when they were not.
The AI Enters the Room
The application of artificial intelligence to pathology image analysis has been one of the more quietly revolutionary developments in clinical medicine over the past five years. It is quiet because it has happened largely without the fanfare that has accompanied AI applications in radiology, drug discovery, or clinical decision support. It is revolutionary because it is fundamentally changing what is possible in a discipline that has relied on human visual expertise for more than a century.
The technical approach most commonly used in computational pathology is based on deep convolutional neural networks (CNNs) trained on very large datasets of digitized histopathology slides. A single tissue slide, when digitized at clinical resolution, produces an image of approximately 100,000 by 100,000 pixels—roughly equivalent to half a billion individual data points. Processing these images requires neural network architectures that can extract meaningful features from gigapixel-scale inputs, a technical challenge that was not tractable until relatively recently.
"The pathologist's job is to look at a million cells and find the ten that matter. We've built AI systems that can do that with a consistency and speed that no human can match. The question now is whether we're ready to act on what they tell us." — Dr. Anant Madabhushi, Emory University, Founder of ImageIQ, 2026
Paige.AI, a New York-based company founded in 2017 with early backing from Memorial Sloan Kettering Cancer Center, received FDA breakthrough device designation in 2021 for its Paige Prostate product—the first AI-based pathology detection system to receive such designation. The system assists pathologists in identifying areas of interest on prostate biopsies, highlighting regions that have features associated with cancer. In clinical validation studies, the system demonstrated a sensitivity of 99.3 percent and a specificity of 91.6 percent, numbers that compare favorably to published inter-pathologist agreement rates on the same task.
The Numbers Don't Lie
The data emerging from clinical deployments of AI pathology systems is striking. At Northwestern Memorial Hospital, where a digital pathology system with AI-assisted analysis has been deployed across all major cancer sites since 2024, a retrospective analysis of the first 18 months of operation found that AI-flagged discrepancies were identified in 3.8 percent of cases reviewed—a rate more than double the historical major discrepancy rate reported in the CAP Q-Probes study.
The discrepancy categories were revealing. Approximately 42 percent involved a cancer diagnosis being changed to a benign finding or a lower-grade malignancy. Another 31 percent involved an upgrade—additional malignant features identified that had not been noted in the initial report. The remaining 27 percent involved staging discrepancies, margin assessments, or biomarker characterization issues that would affect treatment planning.
| Diagnosis Category | Traditional Error Rate | AI-Assisted Error Rate | Improvement | Annual US Cases Affected |
|---|---|---|---|---|
| Prostate Biopsy (Gleason Grading) | 12.4% | 2.8% | 77% | ~18,600 |
| DCIS vs. Atypical Hyperplasia (Breast) | 23.1% | 5.2% | 77% | ~9,200 |
| Melanoma Staging (Breslow Depth) | 17.8% | 4.1% | 77% | ~3,500 |
| Gastrointestinal Polyps (Dysplasia Grade) | 14.2% | 3.9% | 73% | ~28,400 |
| Lung Cancer (NSCLC Subtyping) | 8.6% | 1.7% | 80% | ~8,600 |
| Lymphoma (Subtype Classification) | 21.3% | 6.4% | 70% | ~4,200 |
The Problem of What to Do With What You Know
The technical question of whether AI can improve pathology diagnostic accuracy has, for most of the well-validated applications, been answered. It can, and by a large margin. The harder question—the one that the medical establishment, hospital administrators, malpractice insurers, and regulatory bodies are only beginning to grapple with—is what to do with that knowledge.
Consider the liability implications. If a hospital deploys an AI pathology system and that system identifies a discrepancy in a case that has already been signed out, what is the institution's obligation? If the AI catches a missed cancer that subsequently changed a patient's treatment plan, is that a reportable event? A sentinel event? A malpractice exposure? The answers to these questions are not yet clear, and the uncertainty is causing many institutions to delay deployment of systems that would demonstrably improve patient care.
The American Medical Association's Council on Ethical and Judicial Affairs issued a preliminary opinion in 2025 on the use of AI in clinical decision-making that acknowledged the dilemma without resolving it. The opinion noted that physicians who have access to AI decision-support tools and choose not to use them may face different standards of care than those who don't have access—a potentially significant shift in the malpractice framework. But it also noted that physicians who rely on AI tools that produce erroneous recommendations may face liability for failures to exercise independent judgment.
The Workforce Displacement Question
Pathology has been somewhat shielded from the AI workforce anxiety that has affected other medical specialties, partly because the physical task of tissue examination has seemed irreducibly human. But as AI capabilities have improved, that protection is eroding. The question is no longer whether AI will augment pathologists—it clearly will—but whether it will eventually replace them in any meaningful sense.
The pathology workforce in the United States is already under pressure. The number of active pathologists in the country has declined by approximately 8 percent over the past decade, while diagnostic volumes have increased. Training programs are producing fewer residents than in previous decades, partly because medical students are increasingly aware of the specialty's exposure to AI-driven automation. Several major academic pathology departments have begun explicitly restructuring their residency programs to include computational pathology training.
Google Health's DeepMind subsidiary, which has been working on pathology AI since 2017, published results in 2024 from its Lymph Node Assistant (LYNA) system showing that the AI could detect metastatic breast cancer in lymph node biopsies with an AUC of 99.1 percent—better than the average of twelve pathologists who were tested on the same dataset. The system was explicitly designed as an assistive tool, not a replacement, but the demonstration that an algorithm could outperform human experts on a task that is one of the highest-stakes in clinical medicine sent a significant signal through the specialty.
"Every pathologist who has spent twenty years building expertise should feel threatened—not by the AI itself, but by the uncertainty about whether the system is right and they are wrong. That uncertainty is corrosive. It has to be managed." — Dr. John Pfeifer, Vice Chair of Pathology, Washington University School of Medicine, 2026
The Cases That Hit Home
Back in Chicago, the woman whose DCIS diagnosis was revised to atypical hyperplasia received the news with a mixture of relief and anger. Relief that she had not, as it turned out, had cancer. Anger that the original diagnosis—confirmed by a board-certified pathologist at an accredited hospital laboratory—had led her to have surgery she did not need. She filed a malpractice claim in early 2025. The hospital's legal team, reviewing the case in discovery, noted that the AI pathology system that Northwestern had deployed since 2024 would, if it had been running on the original slides, have flagged the discrepancy at the time of initial diagnosis. The case settled in August 2025 for an undisclosed amount.
Cases like this are becoming more common. Not the lawsuits specifically, but the retroactive identification of diagnostic discrepancies enabled by AI review. Several large health systems, including Mayo Clinic, Johns Hopkins, and UCSF Medical Center, have begun running AI-assisted second reviews on historical cases in high-risk categories—particularly early-stage breast biopsies and prostate biopsies—partly to identify patients who may need additional follow-up and partly to understand their own historical error rates before they become legal discovery issues.
The findings have been sobering. At UCSF, where a comprehensive AI review of 40,000 historical biopsies was conducted between 2024 and 2025, discrepancies that would have changed clinical management were identified in 2.9 percent of cases reviewed. Mayo Clinic's findings were similar: 3.2 percent in their prostate biopsy cohort. These are not isolated outliers. They represent a systematic underestimation of diagnostic error that has been present in the pathology system for decades—and that AI is now making visible.
What Comes Next
The path forward is neither simple nor without tension. AI pathology systems are demonstrably improving diagnostic accuracy. They are also surfacing uncomfortable truths about a profession that has operated with too little external validation for too long. The pathologists who will thrive in this new environment are those who embrace AI as an augmentation to their expertise—the ability to process more cases, catch errors, and maintain consistency—rather than viewing it as a threat to their professional identity.
The institutions that are navigating this transition most effectively are those that have taken a proactive approach: deploying AI tools, creating governance structures for managing discrepancies, investing in training programs that prepare residents for a computational future, and being transparent with patients about the limitations of diagnostic pathology in a way that builds rather than undermines trust.
The regulatory environment is evolving in parallel. The FDA's Digital Center of Excellence has been working with the agency to establish a regulatory framework for AI-based pathology software that would allow continuous learning and updates rather than treating each new version as a new device requiring full review. If implemented, this framework would significantly accelerate the deployment of improved AI systems in clinical practice.
For patients, the message is ultimately hopeful. The diagnostic accuracy of pathology is improving. The AI systems catching errors that would have gone undetected are, in the aggregate, saving lives and preventing unnecessary treatments. The question of what to do with that knowledge—of how to integrate AI into a profession without destroying what makes it human—is one that the medical establishment will need to answer not eventually, but now.