Patients Advocating
for Responsible AI
in Oncology
AICANcer Patient Advocacy Group
Cancer Organizations & Foundations
How to Use This Guide
Crafted from two years of thorough research, this reference guide compiles peer-reviewed analysis, expert reports, and deep-dive studies with insights from webinars, podcasts, and industry summits. It reflects a collaborative effort between healthcare and technology leaders, alongside dialogues with cancer patients and their advocacy organizations.
We are sharing this resource with patient communities, cancer organizations, clinical professionals, and anyone navigating the rapid integration of artificial intelligence into oncology.
It is not an academic paper. It is not a policy brief written for regulators. This document serves as a practical, living reference. Something you can navigate by section, share with peers, and return to when you need a fact, a definition, or a clear explanation of what is actually happening in this space, and how we can together ensure that AI in cancer care is done responsibly and ethically with the patient always at the forefront.
AICANcer Patient Advocacy Group is currently in soft launch. Our website at aicancerpatient.org is live with a few functions — including a contact form to reach us directly — and will continue to grow significantly as we build out our programs and resources. The AI-Literate Patient Advocate Program is active — interested patients can apply by reaching out via the contact form and we will be in touch directly. AI in Cancer Explained is in development and launching Q3–Q4 2026 on the website. We look forward to connecting with you.
What you will find here:
- Part 1 — Who We Are: AICANcer's mission, vision, our solutions, our Board of Directors, and our founder's personal story as a patient and tech leader
- Part 2 — The Landscape: A clear-eyed picture of AI in oncology today — the promise, the risks, and the numbers that define this moment
- Part 3 — The AICANcer Patient Guidelines: Our ten-point benchmark for responsible AI in cancer care, explained in plain language across three domains
- Part 4 — This Is How We Solve It: What is being done right now, worldwide, and by whom — with links to the organizations driving change
- Part 5 — Organizations: A directory of the mentioned AI companies, cancer centers, industry, government bodies, and global research initiatives
- Part 6 — The Evidence Base: Key findings, expert voices, and statistics with full source attribution — every claim traceable independently
- Part 7 — Terminology: Terms you need to learn explained, including an honest critique of patient experience data collection
- Part 8 — What You Can Do: Specific actions for patient advocates, cancer organizations, clinical communities, and AI companies
A patient-led non-profit dedicated to ensuring that AI in cancer is developed with principles of equality, integrity, empathy, and a patient-centered focus. Together we are defining the future of AI in oncology.
1.1 About AICANcer Patient Advocacy Group
The rapid adoption of AI in oncology brings an unexpected hero in the fight to save millions of lives. AI holds the potential to revolutionize cancer care, accelerating drug discovery, enhancing diagnosis, and personalizing treatment — for real. However, this power comes with inherent risks: algorithmic bias, erosion of patient control and security, incomplete data, and the continuation of health inequities.
Without the active, informed voice of the patient community, AI development risks creating a critical gap between innovation and the patient: with solutions that are technically brilliant but ethically flawed, and clinically unequal.
AICANcer Patient Advocacy Group is a newly founded patient-led non-profit organization, dedicated to ensuring that AI in cancer is developed with principles of equality, integrity, empathy, and a patient-centered focus. Together we are defining the future of AI in oncology.
Our Vision
A future where every cancer patient benefits from equitable, transparent, and patient-centered AI solutions, driven by integrity and the active, informed partnership of a unified, diverse patient community.
Our Solutions
AICANcer de-risks innovation for industry and builds trust for patients through four core solutions — each designed to put the patient voice inside the AI development lifecycle, not at the end of it. Together they form a complete system: from education and guidelines to partnerships and policy.
- AI-Literate Patient Advocates: We train a highly diverse, carefully selected cohort of patient advocates. The certified advocates gain a deep, non-technical understanding of AI in general, AI development lifecycles, AI in oncology, and patient data policy and regulation. The training prepares them to challenge, inform, and partner with companies, cancer centers, and policy makers — actively shaping the future of AI in oncology from the ground up. The advocates receive an honorarium for the work AICANcer engages them in.
- AI in Cancer Explained: A knowledge library on our website that demystifies AI for patients and caregivers, giving every patient the understanding they need to ask the right questions about their care, their data, and their rights. Launching Q3–Q4 2026 at aicancerpatient.org.
- The AICANcer Patient Guidelines: Our ten-point benchmark, recommended for adoption by AI companies, cancer centers, and research institutes — built on equality, integrity, empathy, and patient-centered focus. See Part 3 of this guide for the full framework.
- Cross-Sector Partnership Programs: Securing formal MOUs with AI companies and cancer centers to move from abstract ethical commitments to demonstrable integrity. The patient is the expert in their experience, and their voice is the guide for responsible progress.
- Policy and Regulation: Representing cancer patients by ensuring we have a standing seat at the table where patient data policy and regulation in regards to AI is being discussed.
Governance
AICANcer is newly founded and led on a volunteer basis by cancer patient and tech leader, Adiba Zeito. Our Board of Directors brings together expertise in AI, entrepreneurship, venture capital, healthcare policy, law and data privacy.
- Adiba Zeito, President — Founder, AICANcer Patient Advocacy Group LinkedIn
- Deborah Magid, Chair — Co-Founder and Managing Director, NextStar Venture Partners LinkedIn
- Ann Moravick, Treasurer — Founder and CEO/President, Rx4good LinkedIn
- Curtis Mo, Secretary — Partner, DLA Piper LinkedIn
- Andrew Serwin, Member at Large — Co-Chair, Global Data Protection, Privacy and Security Practice, DLA Piper LinkedIn
1.2 Our Founder's Story
A life shaped by passion, resilience, and reinvention — and every chapter of it led here.
From War-Torn Lebanon to Sweden
Adiba was born in Beirut, Lebanon — during the war. When she was seven years old, her family fled. They left everything they knew and started a whole new life in Sweden.
She grew up advocating for her independence in a conservative household. She learned early that you do not wait for a seat at the table — you build the table yourself. That lesson shaped everything that came after.
She built a career she loved with a fierce passion at the centre of the technology and startup ecosystem in Stockholm. In 2011, Adiba initiated and launched Innovative Sweden — a world tour exhibition that brought her to Silicon Valley where she decided to stay. She had the great fortune of being recruited as the youngest CEO in Silicon Valley Forum's over thirty-year history and she loved every minute of it.
The First Cancer — and the Second
In 2005, at 27, she was diagnosed with early-stage breast cancer. A lumpectomy, lots of radiation, antihormonal therapy, and the quiet hope that it was behind her. It wasn't. Three years later, two tiny new tumors appeared in the scarring from her lumpectomy — so they removed the whole breast and treated her with high-dose chemotherapy. Five years of clear mammograms followed, and she thought she was finally done with cancer. That she was now a survivor.
The Day Everything Changed
In 2015, Adiba was 37 and planning for IVF — finally trying to have the baby she had dreamed of for years. Before starting the IVF process the fertility clinic required a mammogram. It showed that Stage 0 cancer had developed in her other, previously healthy breast. A mastectomy was recommended. While recovering from that surgery, she was rushed to the ER with shortness of breath — where a CT scan revealed something the mammograms for years had not: spots on her spine, pelvis, and femur. Her cancer had quietly spread to her bones. Stage 4. Metastatic. Terminal. The doctors said two to three years.
"That was eleven years ago. I am on my 8th line of treatment. I have had 15 surgeries since cancer came into my life, and so much radiation I might just glow in the dark. I've even done cryoablation twice. I have an implanted machine in my spine and metal in my leg. I have grieved too many friends whose lives were taken way too soon. Cancer has taken a lot from me — but it has not taken my love for living."
Alex — The Reason She Keeps Going
Even after her Stage 4 diagnosis, Adiba still wanted to become a mother. She and her ex-husband had saved three embryos before her ovaries were removed. There was hope sitting in a fertility clinic's freezer. On October 26th, 2018 — thanks to an incredible, selfless woman who carried her boy — Adiba became a mamma.
"My son Alex is my heart and soul. He is seven years old. He is the reason I keep going every single day. I want to be here for all his precious moments in life."
Two Worlds. One Mission.
Adiba started working in the tech and startup world when it started becoming a thing — at the end of the 90s. For over two decades, she had a front-row seat to some of the most extraordinary technological innovations in history — in Stockholm and in Silicon Valley. She worked alongside visionaries and entrepreneurs who changed the world with innovation and relentless ambition.
And for over twenty years, she has also been a cancer patient. She knows what it feels like to be on the receiving end of a system that is brilliant in some places and deeply broken in others. She has also been a patient advocate, created awareness campaigns, raised funds for research, and sat on boards of cancer organizations such as Susan G. Komen in San Francisco, METAvivor, and as a member of the MBC Alliance.
When AI began entering oncology with serious momentum, she found herself at the intersection of two worlds that she had the unique opportunity to bridge: the world of the cancer patient who understands the stakes from the inside, and the world of the technology leader who understands how these technologies are built, what incentives drive them, and what can happen when the humans they are supposed to serve are not part of the design conversation.
"I am not here as an outsider looking in. I am a patient. I am an advocate. I am a tech geek. I am a mom who desperately needs this to work. I founded AICANcer because I believe AI has the potential to exponentially transform cancer care. And right now, we have a unique opening to establish the necessary benchmarks that can ensure that this transformation is done ethically, safely and equally — before harm is done. Achieving this requires the integration of diverse patients' voices into every phase of AI in cancer care. We have a responsibility to the next generation of patients as our actions today will define the accessibility and nature of future cancer care."
AI in Oncology Today
The promise is real. So is the risk. A clear-eyed view of where AI is being applied in oncology right now, the numbers that define this moment, and what is at stake if we get it wrong.
2.1 The Promise — What AI Is Already Doing in Oncology
Let us start with what is real and genuinely exciting, because the case for AI in oncology is not exaggerated. It is, if anything, underappreciated by most of the people whose lives it will affect most directly.
Cancer is one of the most complex biological challenges in medicine. It does not behave the same way in every patient, every tumor type, every body, or every environment. For decades, the tools available to oncologists have been blunt instruments applied to a challenge of extraordinary precision. AI changes that — and across every area of oncology, from imaging to drug discovery to treatment planning, it is already doing so.
Imaging & Early Detection
In lung cancer detection — the leading cause of cancer-related mortality worldwide — a 2025 meta-analysis in npj Precision Oncology, Nature Portfolio reviewed 315 studies and found that AI correctly identifies cancer in 86% of cases where cancer is present, and correctly rules it out in 92% of cases where it is not — performing consistently across screening, staging, and prognosis. In plain terms: when AI reviews a lung scan, it gets it right the vast majority of the time. (Yuan et al., npj Precision Oncology, Nature Portfolio, 2025) AI-integrated CT lung cancer screening is now being implemented nationally across multiple countries, with the US requiring insurance coverage for high-risk individuals. (Yuan et al., npj Precision Oncology, Nature Portfolio, 2025; Mulshine & Pyenson, Frontiers in Oncology, 2026)
In pancreatic cancer — projected to become the second leading cause of cancer death in the US by 2030, with 85% of cases currently diagnosed too late for curative treatment — AI is offering genuine and recent hope. Mayo Clinic's REDMOD model, published in Gut, BMJ British Medical Journal Group (April 2026), detected pancreatic cancer on routine CT scans up to three years before clinical diagnosis, identifying 73% of prediagnostic cancers a median 16 months early — nearly doubling specialists' detection rates without AI. (Mukherjee et al., Gut, BMJ British Medical Journal Group, April 2026)
In digital pathology, third-generation AI algorithms analyze whole slide images to identify cancer subtypes and guide treatment selection. In April 2025, Roche received the first-ever FDA, US Food and Drug Administration Breakthrough Device Designation for a computational pathology companion diagnostic — the VENTANA TROP2 RxDx Device for non-small cell lung cancer. This AI analyzes stained tissue slides and computes a quantitative TROP2 protein score that determines whether a patient is likely to benefit from a specific drug. It is currently in accelerated FDA, US Food and Drug Administration review, not yet fully approved for clinical use. Notably, no audited demographic breakdown of the training data has been published.
Drug Discovery & Repurposing
Traditional drug development takes 10–15 years, costs approximately $2 billion, and succeeds only 10% of the time. AI is compressing that timeline dramatically. AlphaFold — Google DeepMind's protein structure prediction AI, awarded the Nobel Prize in Chemistry in 2024 — predicted the 3D shapes of over 200 million proteins, unlocking drug targets previously invisible to researchers. This foundational work, published in Nature in 2021, is now enabling researchers to design molecules that interact with cancer-specific targets with unprecedented precision. (Jumper et al., Nature, 2021)
Beyond new drug discovery, AI is also unlocking existing medicines. EveryCure — a nonprofit co-founded by Dr. David Fajgenbaum at the University of Pennsylvania — uses its AI-powered MATRIX platform to scan over 66 million drug-disease pairs to identify new uses for already-approved medicines. In February 2026, ARPA-H, Advanced Research Projects Agency for Health selected EveryCure to receive up to $76 million in a new three-year phase of funding to advance its most promising drug repurposing opportunities into preclinical and clinical validation — building on an earlier $48.3M ARPA-H, Advanced Research Projects Agency for Health award that funded development of the MATRIX platform itself. EveryCure has already advanced repurposed treatments for rare cancers including angiosarcoma. (everycure.org)
In 2025, AI identified 307 validated synergistic drug combinations for pancreatic cancer alone — one of the most treatment-resistant cancers we know. (Pourmousa et al., Nature Communications, 2025)
Precision Treatment & Dosing
AI models are moving oncology toward what the NCI, National Cancer Institute now defines as the 2026+ standard: the right patient, the right target, the right drug, the right dose, the right treatment sequence. AI is resolving 61% of variant discrepancies between pathologists interpreting the same genomic data. (Suehnholz et al., Cancer Discovery, American Association for Cancer Research, 2024) Digital twins — AI-generated models of individual patients — are achieving 17% radiation dose reductions for glioblastoma patients: less toxicity, same efficacy. (Blasiak et al., npj Precision Oncology, Nature Portfolio, 2025) The SCORPIO model predicts checkpoint inhibitor response across 21 cancer types. (Yoo et al., Nature Medicine, 2025)
Clinical Trial Matching
80% of clinical studies fail to meet their enrollment goals. One in three Phase III trials are terminated due to poor enrollment, and today only 7% of eligible cancer patients enroll in clinical trials. The RECTIFIER study (Unlu et al., JAMA, Journal of the American Medical Association, February 2025) demonstrated that AI-assisted clinical trial prescreening nearly doubled enrollment rates and reduced time-to-eligibility from 50 to 15 days. HopeLLM, an AI clinical trial matching tool, is already in production at City of Hope.
Continuous Monitoring
AI-powered monitoring tools — from wearables to smart pills to implanted biosensors — are generating continuous streams of patient data in real time. These tools can detect treatment complications earlier, track tumor response between appointments, and in theory provide a far richer picture of how a patient is actually doing than a quarterly clinic visit ever could. How that data is collected, consented to, secured, and used for AI training are questions that do not yet have adequate answers — which is precisely why they are addressed directly in the AICANcer Patient Guidelines.
Across every area of oncology, AI tools are demonstrating capabilities that were not possible even five years ago. The pipeline is real, the evidence is accumulating, and the pace of progress is accelerating. These are a few examples among many — and they are why getting the development of AI right matters so much.
2.2 The Numbers That Define This Moment
Every statistic below is attributed to its real source — so you can cite them with confidence and trace them independently.
2.3 The Risk If We Get This Wrong
The same capabilities that make AI so powerful in oncology are the capabilities that make getting it wrong so dangerous. This is not a theoretical challenge. It is already happening.
Bias at Scale
AI is not inherently biased. It does not have prejudices of its own. What it does — with extraordinary efficiency — is mirror the historical biases already embedded in the data it is trained on. Decades of unequal healthcare access, underrepresentation in clinical research, and systemic inequities in medical records do not disappear when data enters an algorithm. They get encoded into it, scaled, and deployed at speed. The challenge is not the technology. It is the data we feed it, and the oversight we apply.
The Testing Gap
The precision medicine data gap does not begin with AI. It begins with who gets tested. Biomarker and genomic testing — the foundation of precision oncology — is significantly underutilized across minority communities, and the data that does exist is overwhelmingly skewed toward patients of European ancestry.
Only 57% of metastatic lung cancer patients with Medicaid coverage had any biomarker testing at all — and testing rates were especially low among patients of color (JTO Clinical Research Reports, 2024). Analysis of GENIE — one of the largest real-world precision oncology data registries in existence — found that Black patients were significantly underrepresented across most cancer types, Hispanic patients were consistently underrepresented across all cancers, and Native American and Pacific Islander populations had zero representation for many cancer types (Kamran et al., npj Precision Oncology, 2023 / Massachusetts General Hospital). Race was reported as a demographic variable in only 37% of genomic sequencing studies — compared to 84% reporting age and 85% reporting gender (National Cancer Institute review, 2019).
This creates a compounding problem: patients from minority communities are less likely to get tested, which means their data is absent from the registries that train AI, which means the AI performs less reliably for those patients, which means those patients benefit less from precision medicine. It is a self-reinforcing cycle — and it can only be broken by actively encouraging biomarker and genomic testing across all cancer communities, not just those who have historically had the easiest access to it.
AICANcer advocates for cancer organizations, foundations, and clinical communities to actively promote biomarker and genomic testing equity — because diverse, representative testing data is not just good for patients today. It is the foundation that AI needs to serve every patient tomorrow.
The evidence that AI amplifies racial disparities in healthcare is no longer a single study — it is a confirmed pattern. In 2019, a landmark study in Science (Obermeyer et al.) demonstrated that a widely-deployed clinical algorithm — used by hospitals and insurers to manage care for about 200 million patients annually — was systematically biased: Black patients assigned the same risk score were considerably sicker than White patients. The algorithm used healthcare costs as a proxy for health need, but because less is historically spent on Black patients, it falsely concluded they were healthier. Remedying the bias would have increased the proportion of Black patients receiving additional care from 17.7% to 46.5% — a reduction of more than half. A 2024 systematic review of 30 studies confirmed a significant association between AI utilization and an exacerbation of racial disparities across health and healthcare outcomes (KFF, Kaiser Family Foundation, 2024). A 2025 University of Michigan study found that medical testing rates for white patients are up to 4.5% higher than for Black patients with identical age, sex, medical complaints and triage scores — bias baked directly into AI training data. And a 2025 Cedars-Sinai study found that most large language models made dramatically different treatment recommendations for the same condition depending on a patient's race. These are not legacy failures. They are happening with the generation of AI now being piloted in oncology.
A comprehensive systematic review — The Algorithmic Divide — published by Mayo Clinic researchers in the Journal of Racial and Ethnic Health Disparities (February 2026) documented AI-driven racial disparities across healthcare, confirming that the diagnostic and treatment gap is largest precisely for the populations who already face the greatest barriers to cancer care.
The Privacy Illusion
The assumption that de-identified health data is private is increasingly unsupported by evidence. Research has demonstrated that 99.98% of Americans could be correctly re-identified in any dataset using just 15 demographic attributes (Rocher, Hendrickx & de Montjoye, Nature Communications, 2019) — and that was before the current era of wearables, continuous biomarker monitoring, and AI-powered data linkage. In 2026, a single medical record sells for 10 to 20 times more than a stolen credit card number on the dark web (Kroll Risk Advisory, 2025), and healthcare was the most breached sector of any industry in 2024, accounting for 23% of all breach cases handled by Kroll — up from 18% in 2023.
Deskilling
A concerning pattern is also emerging in clinical settings: the first documented evidence of AI-related deskilling. When clinicians rely heavily on AI recommendations over time, research shows they can begin to lose independent diagnostic skills (Budzyn, Mori et al., The Lancet Gastroenterology & Hepatology, October 2025). Physicians shown erroneous AI recommendations achieved 73% diagnostic accuracy versus 85% in a control group without AI.
The Consumer AI Trap
A separate and equally urgent challenge is the consumer AI landscape. 230 million people ask ChatGPT a health-related question every week (Brodeur et al., ARISE 2026). Cancer patients are among them — using AI tools to understand their diagnosis, research treatment options, and prepare for appointments. Used thoughtfully, this can be genuinely valuable: AI can help a patient generate a list of questions they never would have thought to ask their oncologist. Used without awareness of its limitations, it carries real risk. Research shows patients cannot distinguish between a doctor's response and an LLM's response, and are equally likely to follow potentially harmful advice as sound advice (Shekar et al., NEJM AI, New England Journal of Medicine, 2025). The solution is not to stop using these tools. It is to understand what they are good for — and what they are not.
AI in oncology is an unexpected hero in the fight against cancer. It holds real, transformative potential to save lives that would otherwise be lost.
It also holds the potential to replicate and amplify every inequality that already exists in cancer care — at scale, at speed, and invisibly.
Both of these things are true simultaneously. The question is not whether to embrace AI in oncology. The question is whether we have the courage and the commitment to ensure it works for every patient — not just the ones whose data happened to be in the training set. That is what the AICANcer Patient Guidelines are built to ensure.
Patient Guidelines
Ten specific, actionable benchmarks recommended for adoption by AI companies, cancer centers, and research institutes worldwide — built on four core principles: Equality, Integrity, Empathy, and Patient-Centered Focus.
The AICANcer Patient Guidelines are the foundational document of our work. They represent two years of deep research, expert consultation, and patient experience distilled into ten specific, actionable benchmarks. These Guidelines move beyond vague ethical language — they name what we recommend, why it matters, and what it looks like in practice.
Each guideline includes the domain it belongs to, the specific focus area, our recommendation, the challenge it addresses, what it means for patients in plain language, and a real-world proof point showing it is already achievable. These are not aspirational wishes. They are operationally achievable recommendations.
AI systems are only as fair as the data they are trained on. These four guidelines address the foundational requirements for building AI that works equally for every patient.
Over 78% of genome-wide association study participants — the foundational data for precision oncology AI — are of European ancestry. AI tools trained on this data have significantly less information about how cancer presents in patients of African, Middle Eastern, East Asian, South Asian, Latin American, and Indigenous ancestry. The gap for Middle Eastern populations is clinically significant: distinct genetic variants relevant to BRCA mutations, colorectal cancer rates, and specific leukemia subtypes are almost entirely absent from Western genomic training datasets. ARISE 2026 found that 95% of FDA, US Food and Drug Administration-cleared AI device summaries omitted demographic data entirely.
When an AI tool recommends a treatment or predicts your risk, you have a right to know whose data it was trained on. If the training set does not include patients like you, the tool's predictions may be far less reliable for you than the headline accuracy figures suggest. This is not a technical footnote — it is a direct patient safety concern.
There is also something every cancer patient can do right now: ask your oncologist about biomarker and genomic testing. Only 57% of metastatic cancer patients with Medicaid coverage received any biomarker testing in a recent study — and patients of color are tested at even lower rates. Every patient who gets tested and contributes their data to research is helping to build the diverse datasets that future AI depends on. Getting tested is not just good for you. It is an act of advocacy for every patient who comes after you.
Pfizer built Ethicara — an internal AI bias review tool that audits all machine learning models for demographic bias at the design stage, before deployment. The NCI, National Cancer Institute launched a Synthetic Data Initiative in 2024 specifically to fill gaps for understudied communities. These prove the recommendation is achievable — it is simply not yet consistently followed.
Source: Sirugo, Williams & Tishkoff, Cell (2019) | Brodeur et al., ARISE 2026 | Pfizer Ethicara | NCI, National Cancer Institute Synthetic Data Initiative (2024)
A 2024 systematic review of 30 studies over ten years confirmed a significant association between AI utilization and an exacerbation of racial disparities in health outcomes (KFF, Kaiser Family Foundation, 2024). A 2025 University of Michigan study found medical testing rates for white patients are up to 4.5% higher than for Black patients with identical presentations — bias baked directly into AI training data. A 2025 Cedars-Sinai study found most large language models made dramatically different treatment recommendations based on patient race alone. These are not legacy failures — they are happening with the generation of AI now being piloted in oncology.
An AI tool that appears accurate overall may be systematically less accurate for you specifically — based on your race, age, geography, or the type of hospital where your records were collected. Without a proactive bias assessment before deployment, you have no way of knowing. In oncology, where AI recommendations can shape treatment decisions, that uncertainty is not acceptable.
The National Kidney Foundation removed race as a variable from kidney function equations in 2021, after evidence confirmed it was producing discriminatory clinical recommendations. This proves that once bias is identified, correction is possible. The EIA creates the process for finding bias before it harms patients rather than after.
Source: Obermeyer et al., Science (2019) | KFF, Kaiser Family Foundation Systematic Review (2024) | Univ. of Michigan PLOS (2025) | Cedars-Sinai (2025) | Haider et al., J Racial Ethn Health Disparities (2026)
44% of the precision oncology market is in North America. The Middle East and Africa — which carry a disproportionate share of the global cancer burden — represent just 3%. Approximately 70% of potential US cancer clinical trial participants live more than two hours from a trial site. AI tools for trial matching, precision diagnostics, and treatment planning are priced and designed almost exclusively for large academic medical centers — leaving the 80% of cancer patients treated in community settings behind.
AI innovation that only reaches well-resourced, well-connected health systems in wealthy countries is not a solution for the global cancer burden. It is a new mechanism for widening the gap. Every AI tool deployed in oncology should document how it performs in low-resource, low-connectivity settings before deployment — not as an afterthought.
The African Research Group for Oncology (ARGO) — an NCI, National Cancer Institute-recognized consortium with 28 institutions across Nigeria — is building Africa-specific oncology AI data infrastructure. AFRICAI, launched at MICCAI 2024, provides publicly available cancer imaging datasets from the African continent. These organizations are building the foundation. AI companies should build toward it.
Source: Research and Markets Precision Oncology Report (2025) | ARGO / Memorial Sloan Kettering Cancer Center Global Program | AFRICAI / MICCAI (2024)
Research has demonstrated that 99.98% of Americans could be correctly re-identified in any dataset using just 15 demographic attributes (Rocher, Hendrickx & de Montjoye, Nature Communications, 2019) — and that was before the current era of wearables, continuous biomarker monitoring, and AI-powered data linkage. In 2026, a single medical record sells for 10 to 20 times more than a stolen credit card number on the dark web (Kroll Risk Advisory, 2025). Healthcare was the most breached sector of any industry, accounting for 23% of breach cases handled by Kroll — up from 18% in 2023 (Kroll Data Breach Outlook, February 2025).
Your cancer data — your genomic sequence, your treatment history, your scan images — is among the most sensitive information that exists about you. It can affect your insurance. It can affect your employment. It can affect your family members. You have a right to know that this data is protected not just by policy, but by technical architecture that makes misuse mathematically difficult. Federated Learning keeps your data in your hospital. Homomorphic Encryption allows AI to learn from it without ever seeing it in readable form.
The Cancer AI Alliance (CAIA) — a $40 million initiative by Dana-Farber, Fred Hutch, Memorial Sloan Kettering, and Johns Hopkins, supported by AWS, Microsoft, NVIDIA, and Google — is building privacy-preserving federated infrastructure for oncology AI. The NCI, National Cancer Institute Federated Learning Pilot expanded in 2024–2025 to include Moffitt Cancer Center and Wake Forest.
Source: Rocher et al. — 99.98% re-identification | Kroll Risk Advisory (2025) | Cancer AI Alliance (CAIA, 2024) | NCI, National Cancer Institute Federated Learning Pilot (2024–2025)
Trust depends on clarity. These three guidelines address what patients and clinicians have a right to understand about any AI that influences their care — and what meaningful consent actually requires.
When an AI recommends a treatment protocol, flags a trial as unsuitable, or informs a prior authorization decision, the patient — and often the clinician — cannot see the reasoning. The 'black box' challenge in healthcare AI is not merely a transparency inconvenience; it is a patient safety risk. Research confirms that patients using consumer AI tools cannot distinguish accurate advice from inaccurate advice — and are equally likely to follow both (Shekar et al., NEJM AI, New England Journal of Medicine, 2025).
If an AI influences a decision about your cancer care, you have the right to a plain-language explanation of why. Not the algorithm's code. Not a technical report. An explanation a person can understand, question, and act on. The Model Card is the nutrition label for AI: what it was trained on, how well it performs, and where its limitations are.
Memorial Sloan Kettering published its Responsible AI (RAI) governance framework in Nature npj Digital Medicine, Nature Portfolio in 2025 — the first published responsible AI governance at scale in oncology. It includes Model Information Sheets, an AI Registry, and a governance committee. The Coalition for Health AI (CHAI, Coalition for Health AI) and the Joint Commission published joint guidance in 2025 requiring bias assessment and model documentation as health system procurement recommendations.
Source: Memorial Sloan Kettering Cancer Center RAI Framework, npj Digital Medicine, Nature Portfolio (2025) | CHAI, Coalition for Health AI / Joint Commission Initial Guidance (2025) | Shekar et al., NEJM AI, New England Journal of Medicine (2025)
42% of oncology trials are now precision oncology trials requiring genomic matching. Patients are consenting to treatments — and data uses — they often cannot fully understand. AI ambient recording, wearable data streams, and genomic sequencing are generating data far beyond the scope of any standard consent form. Genomic data implicates family members who were never part of the consent process.
Consent must mean something. It cannot be a single checkbox at the bottom of a long form that nobody reads. You deserve the right to say yes to some uses of your data and no to others — separately, specifically, and revocably. Consent for your treatment and consent for training an AI model must be different decisions, made at different times, with no clinical penalty for saying no to either.
ASCO, American Society of Clinical Oncology's AI Task Force Principles (May 2025) include explicit consent requirements as one of its six guiding principles for responsible AI in oncology. The Council of Europe's Framework Convention on AI (2024) — the world's first legally binding international AI treaty — includes the right to be informed when AI significantly affects your rights.
Source: ASCO, American Society of Clinical Oncology AI Task Force Principles (May 2025) | Council of Europe Framework Convention on AI, CETS No. 225 (2024)
'Who is liable when AI gets it wrong?' was explicitly named as an unresolved question at an ASCO, American Society of Clinical Oncology-level oncology AI conference in 2025. AI is compressing regulatory submission timelines from 1–2 years to 10–14 months via auto-drafted documents. 65–70 AI-designed drug candidates are currently in human clinical trials. The accountability frameworks must be established before approvals set precedents without patient input.
If an AI tool contributes to a wrong diagnosis, a missed treatment option, or a denied prior authorization — there must be a clear, documented answer to: who is responsible, how do you report it, and how do you appeal it. These are patient protections. And they should be established before a tool is deployed, not discovered after someone is harmed.
Memorial Sloan Kettering's AI Registry and governance committee reviewed 33 live clinical nomograms, sunset 2 models because clinical evidence evolved, and documented the accountability chain at every node. AI project intake at Memorial Sloan Kettering Cancer Center increased 63% between 2023 and 2024 — proof that rigorous governance and rapid innovation are not opposites.
Source: Memorial Sloan Kettering Cancer Center RAI Framework, Nature npj Digital Medicine, Nature Portfolio (2025) | EU AI Act Healthcare Provisions (Aug 2026) | FDA, US Food and Drug Administration AI/ML SaMD Lifecycle Guidance (2025)
AI must augment human care, not replace it. These three guidelines address what it means to build AI that serves the whole patient — and how the patient's voice must remain at the center of every stage of AI development.
AI systems in oncology are predominantly validated against clinical endpoints: tumor reduction, survival rates, disease-free intervals. Patient-reported outcomes — fatigue, pain, cognitive function, emotional distress, financial toxicity — are rarely incorporated into model validation. Beyond this, PROs are typically collected only at scheduled quarterly appointments using numerical scales. A Stage 4 patient who has lived with cancer for ten years rates their pain a '3' against a completely different internal baseline than a newly diagnosed patient. This is called response shift bias, and it means the data AI trains on is inconsistent at its foundation.
An AI tool that optimizes tumor reduction while worsening quality of life is not a patient-centered tool. Real cancer experience is continuous. Current data collection is episodic. Everything that happens between appointments — the bad week after chemotherapy, the night the anxiety was unmanageable — is invisible to the data. AI trained on episodic data will never understand continuous suffering.
AICANcer's Vision: Real-Time Voice-Reported Patient Experience — a voice-based wearable companion provided by the cancer center. Patients talk about their experience continuously, in their own words, without numbers or scales. The AI translates their words into structured data on the backend, compared against that specific patient's own baseline. The data lives in the same secured environment as the patient chart. We call on the technology and healthcare industry to build this on the patient's terms. Contact us via the contact form at aicancerpatient.org
Source: PHQ-9, GAD-7, FACT-G validated instruments | AICANcer Position (2026) | FDA, US Food and Drug Administration Patient-Focused Drug Development Guidance
The first documented evidence of AI-related deskilling has now emerged: when clinicians rely heavily on AI recommendations over time, they can begin to lose independent diagnostic skills (Budzyn, Mori et al., The Lancet Gastroenterology, October 2025). Physicians shown erroneous AI recommendations achieved 73% diagnostic accuracy versus 85% in a control group. The CMS (Centers for Medicare and Medicaid Services) WISeR Model (January 2026) — the first federal deployment of AI in healthcare coverage decisions — does not yet cover oncology, but sets the precedent for AI-assisted prior authorization in Medicare.
Your oncologist must always be able to say no to an AI recommendation — and you must always be able to say no to AI involvement in your care — without clinical penalty. As AI takes on greater responsibility in oncology, the human override is not a courtesy; it is a safety mechanism. The inequity chain does not end at diagnosis — if AI influences both the clinical recommendation and the coverage decision, a patient can face compounded disadvantage at multiple nodes of their care simultaneously.
Memorial Sloan Kettering's RAI governance committee, AI Registry, and clinical validation methodology provide the operational blueprint for institutional AI oversight. Brainomix's stroke AI (The Lancet, 2024) reduced transfer times by 60 minutes by detecting large vessel occlusions on CT — with clear human action triggers built into the oversight model.
Source: Memorial Sloan Kettering Cancer Center RAI Framework (2025) | CMS (Centers for Medicare and Medicaid Services) WISeR Model (January 2026) | Budzyn et al., Lancet Gastroenterology (2025) | Brainomix / The Lancet (2024)
The precision oncology innovation pipeline — from drug targeting to clinical trials to dosing to treatment sequencing — has no mandatory patient feedback mechanism at any node. The patient is the most important voice in the system — and the most consistently excluded. Not because developers are malicious, but because no recommendation has made patient feedback an engineering constraint rather than an optional addition.
You are not just a recipient of AI-generated recommendations. You are the only person in the system who knows what it actually feels like to live through the treatment that AI is optimizing. Your experience — what helped, what made things worse, what side effects the model never predicted — is essential data. Without a formal, structured feedback loop, that data disappears.
Lidia Fonseca, EVP & Chief Digital and Technology Officer, Pfizer, stated at SXSW 2025: 'Patients want to co-create with us and are eager to provide feedback as we develop new products for them.' Pfizer has built patient feedback into its AI development governance. AICANcer is building cross-sector MOUs with AI companies and cancer centers that make the Patient Feedback Loop a commitment, not a voluntary aspiration.
Source: Lidia Fonseca, EVP, Chief Digital & Technology Officer, Pfizer (SXSW 2025) | AICANcer Cross-Sector Partnership Program (2026)
These ten guidelines are not ten separate requirements. They are one integrated vision for what responsible AI in oncology looks like in practice. Guideline #1 ensures the data is representative. #2 ensures bias is found and fixed. #3 ensures the tool reaches every patient. #4 ensures data is protected. #5 ensures AI can be understood. #6 ensures patients genuinely consent. #7 ensures someone is accountable. #8 ensures AI serves the whole patient. #9 ensures humans stay in control. #10 ensures the patient's voice improves the system over time.
We Solve It
Two years of research confirm one critical and encouraging fact: the solutions exist. The challenge is not that responsible AI in oncology is technically impossible. The challenge is that patient-centered recommendations are not widely communicated and not yet consistently followed. That is what AICANcer is working to change.
Three frameworks now converge on the same core recommendations for responsible AI in oncology:
1. The Joint Commission + CHAI, Coalition for Health AI Initial Guidance (2025) — the most comprehensive institutional validation of the AICANcer Patient Guidelines to date, covering risk-based governance, bias assessment, patient privacy, and voluntary safety reporting for US health systems.
2. ASCO, American Society of Clinical Oncology AI Task Force Principles for Responsible AI in Oncology (May 2025) — six guiding principles from the world's leading oncology professional organization: equity, fairness, accountability, transparency, consent, and human oversight.
3. The AICANcer Patient Guidelines — the patient's perspective, translated into ten specific, actionable benchmarks. When you cite all three together, you are not a lone voice. You are the patient dimension of a set of shared recommendations now emerging across institutions.
4.1 Equality & Integrity — Solutions in Action
Diverse Training Data (Guideline #1)
- NCI, National Cancer Institute Synthetic Data Initiative (2024): Computer-generated patient data specifically designed to fill demographic gaps for understudied communities — nci.nih.gov
- International Health Cohorts Consortium (IHCC): 69 cohorts, 34 million+ people from Africa, Asia, Australia, North and South America, and Europe — ihccglobal.org
- AFRICAI Repository (MICCAI 2024): Publicly available cancer imaging datasets from the African continent — miccai.org
- iOncology.ai (AIIMS New Delhi + CDAC Pune): Deep learning cancer image analysis trained on 500,000 images from Indian patient populations, validated in district hospitals across India — aiims.edu
Proactive De-Biasing (Guideline #2)
- Pfizer Ethicara: Internal AI bias review tool that audits all machine learning models for demographic bias at the design stage — before deployment, not after — pfizer.com
- STANDING Together Consensus (The Lancet Digital Health, 2025): International consensus recommendations for tackling algorithmic bias and promoting transparency in health datasets
- FUTURE-AI Framework (BMJ 2025): 30-criteria international consensus framework for fair and unbiased clinical AI — bmj.com
Access Equality (Guideline #3)
- Tempus AI: AI-driven clinical trial matching now deployed at community oncology centers, not only academic medical centers — tempus.com
- ARGO (African Research Group for Oncology): NCI, National Cancer Institute-recognized consortium with 28 institutions in Nigeria building Africa-specific oncology AI data infrastructure — mskcc.org/global
- NIH DSI-Africa / WASHA Takwimu: Hub-and-spoke data science training across Ghana, Nigeria, Tanzania, and Uganda — nih.gov
Data Privacy (Guideline #4)
- Cancer AI Alliance (CAIA): $40M federated AI infrastructure — Dana-Farber, Fred Hutch, Memorial Sloan Kettering Cancer Center, Johns Hopkins; supported by AWS, Microsoft, NVIDIA, Google — canceraialliance.org
- EUCAIM (EU Cancer Images): EU-wide cancer imaging infrastructure with federated architecture — data stays within home institutions — eucaim.eu
- NCI, National Cancer Institute Federated Learning Pilot (2024–2025): Testing AI models predicting chemotherapy response without patient data leaving its institution — nci.nih.gov
4.2 Transparency & Consent — Solutions in Action
Explainability (Guideline #5)
- Memorial Sloan Kettering Cancer Center Responsible AI Governance: Published governance framework including Model Information Sheets, AI Registry, and governance committee — first published RAI governance at scale in oncology (Nature npj Digital Medicine, Nature Portfolio, 2025) — mskcc.org
- CHAI, Coalition for Health AI Responsible AI Framework: Comprehensive guidance for health systems on transparency requirements, model documentation, and patient communication — chai.org
- Ferrum Health — Real-World Validation: Validates AI tools against each hospital's actual patient population: Algorithm output + Baseline characteristics + Clinical ground truth. Over 80% of FDA, US Food and Drug Administration-cleared AI tools underperform in real-world deployment — Ferrum Health catches this gap — ferrumhealth.com
Informed Consent (Guideline #6)
- ASCO, American Society of Clinical Oncology AI Task Force Principles (May 2025): Six guiding principles including explicit consent requirements — asco.org
- FDA, US Food and Drug Administration Predetermined Change Control Plans (Dec 2024): Finalized guidance establishing that AI training data use must be documented before deployment — fda.gov/medical-devices
- Council of Europe Framework Convention on AI (2024): World's first legally binding international AI treaty — coe.int
Accountability (Guideline #7)
- EU AI Act (Healthcare provisions August 2026): High-risk classification of healthcare AI — data quality, bias analysis, human oversight, post-market monitoring, liability documentation — digital-strategy.ec.europa.eu
- FDA, US Food and Drug Administration AI/ML SaMD Lifecycle Management (2025): Developer accountability for post-market performance monitoring and change control — fda.gov
- GAO AI Accountability Framework: US Government Accountability Office framework addressing governance, data, performance, and monitoring — gao.gov
4.3 Empathy & Human-Centered Outcomes — Solutions in Action
Quality of Life & PROs (Guideline #8)
- AICANcer's Vision: Real-Time Voice-Reported Patient Experience: A voice-based wearable AI companion that lets cancer patients talk about their experience continuously — no numbers, no scales, no clipboard surveys. Data stays in the clinical record. We call on the technology and healthcare industry to build this on the patient's terms — aicancerpatient.org
- ASCO, American Society of Clinical Oncology CancerLinQ: Collecting patient-reported data alongside clinical data to build the infrastructure for PRO-validated clinical AI — cancerlinq.org
- FDA, US Food and Drug Administration Patient-Focused Drug Development Guidance: Requires patient experience data in medical product development — fda.gov/pfdd
Human Oversight (Guideline #9)
- Memorial Sloan Kettering Cancer Center AI Governance Committee + AI Registry: Reviewed 33 live clinical nomograms, sunset 2 due to evolved evidence, documented override protocols. AI project intake up 63% in 2024 — mskcc.org
- CMS (Centers for Medicare and Medicaid Services) WISeR Model (January 2026): Six-year pilot using AI for prior authorization across 17 outpatient services in 6 states. Does not yet cover oncology, but sets the precedent AICANcer is tracking closely. We recommend a patient appeal framework before AI enters any oncology coverage decision — cms.gov
Patient Feedback (Guideline #10)
- AICANcer Cross-Sector MOU Program: Formal partnership agreements with AI companies and cancer centers that make the Patient Feedback Loop a practical commitment — aicancerpatient.org
- NCCN, National Comprehensive Cancer Network Policy Summit on DEI in Cancer Workforce (Sept 2024): Establishing policy framework for community-inclusive AI development — nccn.org
4.4 The International Governance Landscape
Responsible AI in oncology does not operate in a regulatory vacuum. These frameworks are actively shaping what AI companies, health systems, and governments are doing right now.
United States
- FDA, US Food and Drug Administration AI/ML Software as Medical Device (SaMD) Action Plan (2023–2025): Framework for AI medical device lifecycle management. Demographic diversity reporting still not universally required as of 2026 — fda.gov
- CMS (Centers for Medicare and Medicaid Services) WISeR Model (January 2026): First federal AI deployment in healthcare coverage decisions — cms.gov
- Coalition for Health AI (CHAI, Coalition for Health AI): Primary US institutional reference for responsible healthcare AI — chai.org
- Joint Commission + CHAI, Coalition for Health AI Guidance (2025): Most comprehensive institutional validation of the AICANcer Patient Guidelines framework — jointcommission.org
European Union
- EU AI Act (Healthcare provisions fully apply August 2026): High-risk classification of healthcare AI — the most powerful legislative anchor for all ten AICANcer Patient Guidelines — digital-strategy.ec.europa.eu
- EU Health Data Space: Cross-European health data infrastructure with FHIR/OMOP standards and consent architecture built in — health.ec.europa.eu
International
- Council of Europe Framework Convention on AI (CETS No. 225, 2024): World's first legally binding international AI treaty — coe.int
- UN General Assembly Resolution on AI (March 2024 — unanimous, 193 states): First UN consensus on AI governance with human rights as the foundation — un.org
- ASCO, American Society of Clinical Oncology AI Task Force Principles (May 2025): Six guiding principles from the world's leading oncology professional organization — asco.org
Driving Change
A navigable directory of AI companies, cancer centers, industry, government bodies, governance organizations, and global research initiatives mentioned in this reference guide — with roles, websites, and guideline mapping.
| Organization | Role / What They Do | Website | Guidelines |
|---|---|---|---|
| Tempus AI | Precision medicine, AI-driven oncology data, clinical trial matching, chart summarization — deployed at community oncology centers | tempus.com | #1 #3 #10 |
| Flatiron Health | Oncology-specific real-world evidence and EHR data — feeds real-world data into AI training pipelines | flatiron.com | #1 #2 |
| PathAI | AI-powered pathology diagnostics for cancer — training data demographics are EIA-critical | pathai.com | #1 #2 #5 |
| Paige AI (Tempus 2025) | Computational pathology AI — world's largest oncology AI foundation model post-merger | paige.ai | #1 #2 #5 |
| ConcertAI | Real-world oncology data and clinical AI — RWE for oncology trials | concertai.com | #1 #4 #10 |
| Lunit | AI biomarker analysis for oncology trials — pathology and radiology AI | lunit.io | #1 #2 #5 |
| EveryCure | Nonprofit — MATRIX AI platform scanning 66M+ drug-disease pairs. $76M ARPA-H, Advanced Research Projects Agency for Health funding (Feb 2026). Breast cancer program active. | everycure.org | #3 #8 |
| Organization | Role / What They Do | Website | Guidelines |
|---|---|---|---|
| Memorial Sloan Kettering | Published Responsible AI governance framework — AI Registry, Model Info Sheets, governance committee (npj Digital Medicine, Nature Portfolio, 2025) | mskcc.org | #7 #9 #10 |
| Dana-Farber Cancer Institute | Co-founder Cancer AI Alliance (CAIA) — federated learning infrastructure | dana-farber.org | #4 #9 |
| Mayo Clinic | REDMOD — AI detecting pancreatic cancer up to 3 years early (Gut, BMJ British Medical Journal Group, April 2026); Algorithmic Divide systematic review (2026) | mayoclinic.org | #1 #2 #5 |
| Johns Hopkins / Sidney Kimmel | CAIA co-founder — precision oncology and AI governance research | hopkinsmedicine.org | #1 #4 #7 |
| Moffitt Cancer Center | NCI, National Cancer Institute Federated Learning Pilot — AI predicting chemotherapy response without data leaving institution | moffitt.org | #4 |
| Organization | Role / What They Do | Website | Guidelines |
|---|---|---|---|
| Pfizer | Ethicara — internal AI bias review tool auditing all ML models at design stage before deployment | pfizer.com | #1 #2 |
| AstraZeneca | AI integration in oncology R&D — computational pathology data at ASCO, American Society of Clinical Oncology 2025 | astrazeneca.com | #1 #5 |
| Roche / Genentech | VENTANA TROP2 RxDx Device — first-ever Breakthrough Device Designation for computational pathology companion diagnostic for NSCLC (April 2025). In accelerated review, not yet fully approved. | roche.com | #1 #5 #7 |
| Epic Systems | Cosmos Foundation Model — 350M real-world clinical records. AICANcer recommends a published demographic audit. | epic.com | #1 #2 #4 |
| Organization | Role / What They Do | Website | Guidelines |
|---|---|---|---|
| Coalition for Health AI (CHAI, Coalition for Health AI) | Primary US institutional reference for responsible healthcare AI — co-authored Joint Commission guidance | chai.org | #1–#10 |
| The Joint Commission | 2025 Initial Guidance for Responsible AI in US Health Systems — most comprehensive institutional validation of the AICANcer Patient Guidelines framework | jointcommission.org | #1–#10 |
| ASCO, American Society of Clinical Oncology | AI Task Force Principles for Responsible AI in Oncology (May 2025) — CancerLinQ PRO data infrastructure | asco.org | #2 #5 #7 #9 |
| NCCN, National Comprehensive Cancer Network | Policy Summit on DEI in Cancer Workforce (Sept 2024) — framework for community-inclusive AI development | nccn.org | #1 #3 #10 |
| Council of Europe | Framework Convention on AI (CETS No. 225, 2024) — world's first legally binding international AI treaty | coe.int | #4 #5 #6 #7 |
| Organization | Role / What They Do | Website | Guidelines |
|---|---|---|---|
| Ferrum Health | Platform for hospitals to continuously validate and monitor clinical AI performance against their own patient population — vendor-neutral ground truth on any FDA-cleared, CE-marked, or homegrown model. Founded by Pelu Tran after a personal family tragedy involving undeployed cancer detection AI. 43% of FDA-approved AI devices validated without real patient data. | ferrumhealth.com | #2 #7 #9 #10 |
| Cancer AI Alliance (CAIA) | $40M federated AI infrastructure — Dana-Farber, Fred Hutch, Memorial Sloan Kettering Cancer Center, Johns Hopkins; AWS, Microsoft, NVIDIA, Google | canceraialliance.org | #4 #9 |
| Organization | Role / What They Do | Website | Guidelines |
|---|---|---|---|
| EUCAIM | EU-wide cancer imaging infrastructure — federated architecture, data stays within home institutions | eucaim.eu | #1 #3 #4 |
| IHCC | International Health Cohorts Consortium — 69 cohorts, 34M+ people from 5 continents | ihccglobal.org | #1 #3 |
| ARGO | African Research Group for Oncology — NCI, National Cancer Institute-recognized, 28 institutions in Nigeria building Africa-specific oncology AI data | mskcc.org/global | #1 #3 |
| AFRICAI Repository | Publicly available cancer imaging datasets from the African continent (launched MICCAI 2024) | miccai.org | #1 #3 |
Key statistics, expert voices, and further reading — all fully attributed so you can cite with confidence and trace every claim independently.
Every statistic below includes its source. Peer-reviewed studies are marked (PR). Practitioner-reported figures are marked (EX) — expert estimates from named practitioners. Always cite the original source, not this document.
6.1 ARISE State of Clinical AI 2026 — Key Findings
The ARISE State of Clinical AI 2026 (Brodeur et al., Stanford/Harvard, January 2026) is the most comprehensive annual synthesis of clinical AI evidence, reviewing 131 studies across six domains. Published by Dr. Peter Brodeur (Beth Israel Deaconess Medical Center / Harvard Medical School), Dr. Ethan Goh (Stanford University), Dr. Adam Rodman, and Dr. Jonathan H. Chen — with contributions from a multidisciplinary group across Stanford, Harvard, and affiliated health systems.
- 95% of FDA, US Food and Drug Administration-cleared AI device summaries omitted demographic data (PR) — Brodeur et al., ARISE 2026
- 91% of FDA, US Food and Drug Administration-cleared AI devices lacked bias assessments entirely (PR) — Brodeur et al., ARISE 2026
- 10–22% of real clinical AI cases had potentially severely harmful recommendations — 77% were errors of omission, the invisible failure mode (PR) — Brodeur et al., ARISE 2026
- 230 million people ask ChatGPT a health question every week. Patients cannot distinguish AI advice from physician advice and are equally likely to follow both regardless of accuracy (PR) — ARISE 2026 / Shekar et al., NEJM AI, New England Journal of Medicine, May 2025
- Deskilling: First documented evidence that clinicians over-relying on AI begin to lose independent diagnostic skills. AI-trained physicians shown erroneous recommendations: 73% accuracy vs 85% control (PR) — Budzyn, Mori et al., Lancet Gastroenterology, October 2025
6.2 The Scale of the Opportunity
- $341B — Precision oncology market projected by 2035 (PR) — Research and Markets, 2025
- $106B — Precision oncology clinical trial market, projected to reach $159B by 2029 (PR) — Novotech CRO / Research and Markets, 2024–25
- 1,200+ — AI-enabled medical tools cleared by FDA, US Food and Drug Administration as of 2026 — a tenfold increase from 2020 (PR) — FDA, US Food and Drug Administration data / ARISE 2026
- 65–70 — AI-designed drug candidates currently in human trials — zero have received full FDA, US Food and Drug Administration approval yet (PR) — Wilczok & Zhavoronkov, Clinical Pharmacology and Therapeutics, 2025
- 2× — AI clinical trial prescreening doubled enrollment rates and cut time-to-eligibility from 50 to 15 days (PR) — Unlu et al. RECTIFIER, JAMA, Journal of the American Medical Association, February 2025
6.3 The Equity Gap
- 78% of genome-wide association study participants are of European ancestry (PR) — Sirugo, Williams & Tishkoff, Cell 177(1), 2019
- 80% of cancer patients receive care in community oncology settings, not academic centres where most AI is built (EX) — Lidia Fonseca, Chief Digital & Technology Officer, Pfizer; Jeff Legos, Chief Oncology Officer Pfizer, 2025
- 3% — Middle East and Africa's share of the precision oncology market, despite disproportionate global cancer burden — Research and Markets, 2025
- 85% of pancreatic cancer cases diagnosed too late for curative treatment — projected #2 US cancer killer by 2030 (PR) — Mukherjee et al., Gut, BMJ British Medical Journal Group, 2026
- 99.98% of Americans can be correctly re-identified in any dataset using 15 demographic attributes (PR) — Rocher, Hendrickx & de Montjoye, Nature Communications, 2019
- 23% — Healthcare accounted for 23% of breaches handled by Kroll in 2024 — the most breached sector, up from 18% in 2023 (PR) — Kroll Data Breach Outlook, February 2025
6.4 AI Promise — The Evidence
- 86% / 92% — AI correctly identifies lung cancer in 86% of cases where it is present, and correctly rules it out in 92% of cases where it is not — across 315 studies (PR) — Yuan et al., npj Precision Oncology, Nature Portfolio, 2025
- 73% of prediagnostic pancreatic cancers identified by REDMOD a median 16 months early — nearly doubling specialists' detection rates (PR) — Mukherjee et al. / Mayo Clinic, Gut, BMJ British Medical Journal Group, April 2026
- 17.6% higher cancer detection rate when AI supported radiologists in nationwide German mammography study — 463,094 women, 119 radiologists (PR) — Eisemann et al. (PRAIM), Nature Medicine, January 2025
- 61% of genomic variant discrepancies between pathologists resolved by AI (PR) — Suehnholz et al., Cancer Discovery, American Association for Cancer Research, 2024
- 17% radiation dose reduction achieved by digital twin AI for glioblastoma patients (PR) — Blasiak et al., npj Precision Oncology, Nature Portfolio, 2025
- 307 validated synergistic drug combinations identified by AI for pancreatic cancer alone in 2025 (PR) — Pourmousa et al., Nature Communications, 2025
- 66M+ drug-disease pairs scanned by EveryCure's MATRIX AI platform (PR) — EveryCure / ARPA-H, Advanced Research Projects Agency for Health, 2026
6.5 Expert Voices
On the State of Clinical AI — ARISE 2026
"The field has studied performance before safety — which is kind of the opposite of the way the pharmaceutical industry works. You start with proof of concept and safety, then eventually get into effectiveness. It's a little bit inverted."
"Patients can't tell the difference between a doctor and an LLM. They can't tell the difference between good and bad advice. They're just as likely to follow good advice as bad advice. The onus is not on them and it can't be on them."
"230 million people on a weekly basis ask ChatGPT a question related to healthcare. One of our predictions: patients will get more advice, more coaching, more often from AI tools than from an actual human. That's probably already the case."
On Bias & Equity
"Patients in Appalachian communities are entirely absent from most AI training datasets. When AI has never seen your patient population, it cannot serve them."
"More than 80% of FDA, US Food and Drug Administration-cleared AI tools perform significantly worse in real-world hospital deployment than their clearance benchmarks suggested."
On Innovation & Promise
"Patients want to co-create with us and are eager to provide feedback as we develop new products for them."
"The greatest barrier to saving lives from pancreatic cancer has been our inability to see the disease when it is still curable. This AI can now identify the signature of cancer from a normal-appearing pancreas, and it can do so reliably over time and across diverse clinical settings."
"No patient should suffer while a cure hides in plain sight."
6.6 Key Sources & Further Reading
ARISE & Most Recent Evidence
- Brodeur et al. / ARISE State of Clinical AI 2026 — The most comprehensive annual synthesis of clinical AI evidence — 131 studies, six domains. arise-ai.org
- Mukherjee et al. / REDMOD (Gut, BMJ British Medical Journal Group, April 2026) — AI detecting pancreatic cancer up to 3 years before clinical diagnosis. gut.bmj.com
- Yuan et al. (npj Precision Oncology, Nature Portfolio, 2025) — Meta-analysis of AI for lung cancer image-based diagnosis — 315 studies
- Eisemann et al. / PRAIM (Nature Medicine, Jan 2025) — Nationwide real-world AI mammography screening study
- Unlu et al. / RECTIFIER (JAMA, Journal of the American Medical Association, February 2025) — AI clinical trial prescreening doubles enrollment rates
- EveryCure / ARPA-H, Advanced Research Projects Agency for Health (February 2026) — AI drug repurposing platform — everycure.org
Governance & Policy
- ASCO, American Society of Clinical Oncology AI Task Force Principles (May 2025) — asco.org
- CHAI, Coalition for Health AI / Joint Commission Initial Guidance (2025) — chai.org / jointcommission.org
- EU AI Act Healthcare Provisions (2024) — digital-strategy.ec.europa.eu
- Council of Europe Framework Convention on AI (2024) — coe.int
- FDA, US Food and Drug Administration AI/ML SaMD Lifecycle Management (2025) — fda.gov/medical-devices
AICANcer's Own Work
- AICANcer Patient Guidelines — available via the contact form at aicancerpatient.org
- AI-Literate Patient Advocate Program — active now. Interested patients can reach out via the contact form at aicancerpatient.org
- AI in Cancer Explained — Free knowledge library for patients and caregivers, launching Q3–Q4 2026 — aicancerpatient.org
Directory
Every term used across AICANcer's work, explained in plain language. No technical background required. Organized by category — AI & Technology, Clinical, Governance & Policy, and Patient Experience.
You do not need to memorize these terms. You need to recognize them when they come up — in a vendor meeting, a conference panel, a policy hearing — and know what question to ask. Each entry includes what the term means and why it matters for patients and advocates.
AI & Technology Terms
Clinical Terms
Governance & Policy Terms
Patient Experience & Quality of Life Metrics
The following tools are the standardized instruments currently used to measure patient experience and quality of life in oncology. They are clinically validated and represent decades of research. AICANcer does not dismiss them — we critique how they are collected and why that makes them insufficient as the primary source of patient experience data for AI training.
Response Shift Bias: A patient who has lived with Stage 4 cancer for eleven years rates their pain a '3' against a completely different internal baseline than a newly diagnosed patient. The scale looks standardized. The human experience underneath it is not. When AI trains on numerical PRO data, it inherits this inconsistency invisibly.
Collection Method Inconsistency: A nurse asking verbally in triage, a printed clipboard survey with smiley faces, a 1–10 scale on a tablet — these are not equivalent data collection methods. Context affects the data.
The Frequency Gap: PROs are collected at scheduled appointments — monthly or quarterly at best. The bad week after chemotherapy, the night the anxiety was unmanageable — all invisible. Real cancer experience is continuous. Current collection is episodic. AI trained on episodic data will never understand continuous suffering.
This critique comes directly from lived experience. After eleven years with Stage 4 cancer, AICANcer's founder Adiba Zeito has completed hundreds of PRO surveys. The gap between what those surveys capture and what the experience of cancer actually feels like is not a small methodological detail.
A voice-based wearable companion provided by the cancer center. Patients talk about their experience — their pain, their fatigue, their fear, their good days, what is helping, what is not. No numbers. No scales. No clipboard. Just conversation.
The AI listens and translates the patient's words into structured data on the backend — continuously, in real time. That data is compared against that specific patient's own baseline, not a population average. It lives in the same secured environment as the patient chart — not in a consumer app. Part of the clinical record. Part of the AI training data with proper governance.
AICANcer is a patient advocacy organization — we will not build this vision. We are defining what it must be, and calling on companies with the technical capability to build it with these non-negotiables: patient data security equivalent to clinical records, clinical integration not consumer app architecture, and reduced patient burden not increased.
If your organization is working on something like this, we would be happy to provide input and share the patient perspective. Reach out via the contact form at aicancerpatient.org
Can Do
The solutions exist. The governance frameworks are being built. The patient voice is the missing ingredient that transforms technically capable AI into equitable AI. Here is what you can do right now.
For Patient Advocates
- Become an AI-Literate Patient Advocate: AICANcer's training program gives you the knowledge to challenge, inform, and partner with companies, cancer centers, and policymakers on behalf of every cancer patient. No technical background required — you bring what no researcher or vendor can: lived experience. Our cohort is deliberately diverse across cancer type, cancer stage, race, ethnicity, geography, age, and background. We are cancer type and stage agnostic. Advocates receive an honorarium for the work AICANcer engages them in. The program is active — reach out via the contact form at aicancerpatient.org and we will be in touch directly. We share the application link with those we believe would be a strong fit.
- Use AI tools wisely — and help others do the same: 230 million people ask ChatGPT a health question every week. AI can be genuinely useful for generating questions to bring to your oncologist, understanding medical terms, or finding out what clinical trials exist. What it cannot do is replace clinical judgment. Use AI to prepare for conversations with your care team. Not to replace them. AICANcer's AI in Cancer Explained knowledge library helps you navigate exactly this — launching Q3–Q4 2026 at aicancerpatient.org
- Ask your oncology team: Which AI tools are being used in your care, what were they trained on, and who is responsible if they get something wrong. These are your rights. You are allowed to ask.
- Ask about biomarker and genomic testing: If you have not had biomarker or genomic testing, ask your oncologist whether it is appropriate for your cancer type. Precision medicine — and the AI being built to power it — depends on diverse, representative data. Patients from minority communities are tested at significantly lower rates, which means their data is missing from the registries that train AI. Every patient who gets tested and contributes to research is helping to build a more equitable foundation for future cancer care. Getting tested can open doors to targeted therapies and clinical trials — and it makes the system better for everyone who comes after you.
- Share your experience: Your lived experience is the most powerful evidence for why better data collection tools are needed. When our website is fully live, you can share your story at aicancerpatient.org — or reach out to us now via the contact form
For Cancer Organizations & Foundations
- Endorse the AICANcer Patient Guidelines as part of your AI governance stance — and signal to vendors that patient-centered standards matter to you.
- Partner with us — reach out via the contact form at aicancerpatient.org
For Clinical Communities & Cancer Centers
The clinical AI market is intensely competitive — thousands of vendors need your deployment credibility. That is your leverage. Use it to set the terms.
Before signing any AI contract, consider asking for: an audited demographic breakdown of training data, an Equality Impact Assessment across patient subgroups, a Model Card, documented liability at every decision node, and evidence of validation against patient-reported outcomes — not just clinical endpoints. Validate every tool against your own patient population, not just the academic datasets used for FDA clearance.
Build an AI Registry. Establish override protocols. Engage CHAI and cite the AICANcer Patient Guidelines alongside ASCO's AI Task Force Principles. The patient voice belongs at your procurement table — and we are here to help put it there. Partner with us via the contact form at aicancerpatient.org
For AI Companies & Technology Partners
The patients your tools are designed to serve are not at the end of your development process. They should be at the beginning of it.
- How many of our Guidelines are you applying? Are you addressing Guideline #3 on access equality, #5 on explainability, #7 on accountability — or are you working toward all ten? The AICANcer Patient Guidelines are not a compliance checklist. They are a blueprint for building AI that patients can trust. Which leads us to our ask:
- Bring us in from day one — not as a checkbox, but as a design partner. The patient voice de-risks your development and builds the trust your product needs to succeed. Partner with us via the contact form at aicancerpatient.org
Ensuring that AI in oncology is developed with principles of
Equality, Integrity, Empathy, and a Patient-Centered Focus.
Prepared by Adiba Zeito, Founder · May 2026 · A Reference Guide
© 2026 AICANcer Patient Advocacy Group