If you’ve used a generative AI (GenAI) tool—like ChatGPT, Gemini, or Claude—to research a cancer treatment, or to evaluate a program you’re considering, you are not alone. Wanting to understand what you’re walking into is a sign of an engaged, thoughtful patient. But before allowing an AI summary to shape a decision about your health, there are important things you need to understand about how these tools work, and where they fail.
GenAI Relies Heavily on What’s Been Published
Most widely used GenAI systems are trained primarily on text: published medical and scientific literature, regulatory databases, mainstream health websites, and clinical guidelines. They are, in essence, very sophisticated readers of what has been formally written down and made publicly available.
That may sound comprehensive. However, it is not.
A vast body of clinical knowledge never makes it into peer-reviewed journals. This silence is structural, not scientific. Research into treatments that are off-patent, low-cost, or practiced at specialized centers operating outside major academic networks faces layered barriers to publication: limited institutional funding, journal gatekeeping, and the simple reality that independent treatment centers rarely have the infrastructure to generate the large randomized trial data that flagship journals demand.
But this is only one part of a larger truth. Medicine is far wider than what gets studied, funded, and published by its most dominant institutions. There is a vast body of clinical knowledge, accumulated through careful observation, refined over decades of practice, and validated in the lives of real patients, that exists entirely outside the canon GenAI was trained on. It isn’t fringe. It isn’t unproven. It is simply unnoticed by the systems that decide what counts as evidence. This is the knowledge that we often perceive as medical intuition, experience, or acumen. And GenAI, trained on the same systems, inherits that blind spot.
When a GenAI tool tells you a treatment “lacks evidence” or is “not supported by the literature,” what it is actually telling you is: “this treatment has not been extensively published in the sources I was trained on.” That is a statement about the limits of its training data, not a statement about the treatment’s clinical value.
This skew has a practical consequence: GenAI tends to favor treatments that are well-funded, well-published, and institutionally mainstream — not because they are necessarily more effective, but because they are more represented in the data. And because GenAI is fluent and confident by design, it can construct a compelling, well-reasoned case for that skewed picture without any indication that the picture is incomplete. The gap doesn’t read as a gap. It reads as a verdict.
GenAI Also Underestimates Treatments It Does Know About
GenAI doesn’t only lack data on unfamiliar or emerging therapies. It also tends to underestimate the effectiveness of well-recognized treatments whose full clinical potential has outpaced what’s been formally studied and published.
Take photodynamic therapy or hyperthermia, treatments with real scientific backing that GenAI will acknowledge as legitimate. What AI doesn’t know is how much more effective these therapies can be in the hands of experienced clinicians, applied at the right intensity, sequenced with complementary treatments, and individualized to the patient.
The published literature captures what was studied in controlled trials: isolated conditions, standardized protocols, and academic center settings. It does not capture what happens when a physician with decades of clinical experience applies these tools within a multi-modal protocol refined over thousands of patient encounters. This means a treatment GenAI rates as modestly promising may, in the right clinical context, be something considerably more. That distinction never appears in the summary a patient reads.
GenAI Has a Jurisdictional Blind Spot
Medical practice is not uniform across the world, and neither is the research literature GenAI learns from. A GenAI tool trained predominantly on content from one country will naturally reflect the clinical norms, approved treatment pathways, and regulatory context of that country, and may have limited exposure to legitimate practices, approved therapies, and clinical evidence from other healthcare systems.
This matters for patients exploring treatment options across borders. A therapy that is standard practice in one country may simply not appear, or may appear with caveats and warnings, in a GenAI tool anchored to a different regulatory and clinical context. That gap is a limitation of the tool’s training, not a reflection of the therapy’s validity.
AI Cannot See Clinical Experience
Perhaps the most important limitation of all: GenAI generally has no access to what physicians have observed at the bedside over years and decades of practice. Clinical intuition — the pattern recognition that develops when a physician has treated thousands of patients, seen what works, adapted protocols in real time, and built up institutional knowledge that has never been formally published – is largely invisible to GenAI. It isn’t written in a form that GenAI could learn from, so, as far as most GenAI tools are concerned, it doesn’t exist.
This is not a minor gap. Clinical experience is one of the most important inputs in medicine. It is how the field has advanced for most of human history, and it remains indispensable even in the age of evidence-based, data-driven decision-making.
It is worth noting that this limitation applies equally across all of medicine, not just in integrative or non-institutional settings. The most respected oncologists at the most prestigious academic centers routinely make clinical judgments that go beyond what any published protocol specifies. They adjust dosing based on patient response, sequence treatments in ways that trial design never tested, and recognize patterns in their patients that no dataset has formally captured. Clinical experience is not the domain of any particular school of medicine. It is how all good physicians practice, everywhere.
What This Means for You
When AI renders a verdict on a treatment, which it does confidently and in authoritative language, it is generally doing so without access to outcomes that weren’t published, without a full understanding of jurisdictional differences in clinical practice, without knowledge of how treatments perform in combination rather than isolation, and without any of the lived experience your physician brings to the table.
It is not evaluating your situation. It is pattern-matching against a dataset that was never designed to represent the full landscape of medicine as it is actually practiced.
If you are a patient or representing one, bring your questions, including what AI told you, to an experienced physician. GenAI can be a useful starting point for building knowledge and framing questions. But your physician can tell you what GenAI cannot: what they have actually observed across years of clinical experience, using protocols that extend far beyond what any database could possibly contain.
That conversation is where real medical judgment lives. No GenAI has access to it.
Author
