Who Should Read This
- Oncologists and clinicians considering AI adoption in diagnostic or treatment pathways
- Medical technology decision-makers evaluating AI platforms for institutional use
- Healthcare administrators weighing investment in AI-enabled oncology systems
- Anyone professionally responsible for integrating emerging technology into high-stakes environments where judgment errors have clinical consequences
Why You Should Read This
- Because AI enthusiasm in medicine currently runs ahead of AI discipline — this piece recalibrates the conversation
- Because most discussions treat AI as categorically different from prior tools when the accurate frame is amplification, not transcendence
- Because blind reliance creates false confidence precisely where skepticism belongs — in diagnostic reasoning, treatment selection, and outcome prediction
- Because this analysis comes from someone who uses AI extensively across multiple domains yet remains anchored to epistemic rigor rather than technological fascination
The Mythological Inflation Around AI
A recent review of ASCO publications included discussion of artificial intelligence in oncology. The narrative felt inflated — bordering on the mythological. AI was presented almost as a demi-god rather than what it actually is: a human-made tool constrained by human knowledge, human data, and human assumptions.
This prompted me to draft a response outlining, with balance, what AI can and cannot do in oncology. Not because I oppose AI — I use it daily, across medicine, hospitality operations, land management, even circuit diagnostics. Because I’ve seen what happens when technological enthusiasm bypasses epistemic discipline. The fascination with capability obscures the question of governance.
AI Does Not Transcend — It Amplifies
Before calculators existed, a subset of humans could perform complex mathematics manually. The process was slower and restricted to those trained in it. Calculators did not change the nature of mathematics. They accelerated outcomes and democratized access. The capability already existed — the scale and speed changed.
We are witnessing a similar paradigm shift with AI, amplified by cross-domain reach. What was previously inaccessible to many is now available through algorithmic inference rather than direct expertise.
Two personal examples. I could not program the timer switch for a fountain and its lighting system — a task involving hardware I’d never encountered. By uploading a photo, AI provided step-by-step instructions within seconds. In another instance, I shared an image of a gate operator’s motherboard. AI analyzed the circuitry and guided fault identification. These were tasks I could not have accomplished independently, yet they required no transcendent intelligence. The pattern recognition architecture was simply trained on enough examples to infer likely solutions from visual input.
The Limits: Hallucination, Confidence Without Verification, and Ego Reinforcement
Despite being a powerful enabler, AI has serious structural limitations — especially when used uncritically in high-stakes domains such as medicine.
First: AI hallucinates. It generates plausible-sounding outputs that have no correspondence to reality. In oncology, this could manifest as diagnostic suggestions based on pattern similarity rather than mechanistic validity, or treatment recommendations drawn from correlations in training data that do not survive clinical scrutiny. The system cannot distinguish what it actually knows from what it probabilistically generates.
Second: AI presents output with uniform confidence regardless of evidentiary strength. A clinician reviewing AI-generated differential diagnoses receives no calibrated uncertainty signal. The system does not flag when it is operating at the edge of its training distribution or when multiple plausible interpretations exist with equivalent probability. This creates the illusion of certainty where ambiguity should trigger further investigation.
Third — and this is particularly insidious — AI can be addictive. It subtly reinforces ego by making us feel unusually “ahead of the curve.” The psychological reward system activates not because we’ve acquired deeper understanding but because we’ve accessed a capability shortcut. Over time, this can erode the discipline of verification. If the tool usually produces acceptable results, the impulse to independently validate weakens. Blind reliance can trap users in a cul-de-sac of false confidence, harming outcomes rather than improving them.
Oncology as the Test Case for Epistemic Rigor
Oncology is unforgiving. A missed diagnosis, an incorrect staging assessment, a treatment recommendation based on flawed pattern recognition — these have irreversible consequences. The margin for error is narrower than in domains where mistakes are iteratively correctable.
This makes oncology the ideal domain for testing whether AI adoption can coexist with epistemic discipline. Can clinicians integrate AI’s pattern recognition capability while maintaining the diagnostic reasoning that catches what algorithms miss? Can institutions deploy AI-enabled platforms without deskilling the workforce or creating dependency on systems whose failure modes remain incompletely characterized?
The answer depends on governance architecture — not just the technology itself. If AI is introduced as an adjunct to clinical judgment, with explicit protocols for when algorithmic output requires independent verification, it can genuinely enhance capability. If it is introduced as a substitute for diagnostic reasoning, it will degrade it. The difference is not in the AI; the difference is in how the institution governs its use.
Technology Fascination Must Not Replace Judgment
We do not use calculators because they are mesmerizing. We use them when calculation is required, and we verify results when precision matters. The same principle applies to AI.
A simple analogy holds: we do not drive a car merely because we own one — we drive it when there is a need to reach a destination the car enables. AI, like any tool, must be governed by intent, restraint, and discernment.
In oncology specifically, this translates to clear operational discipline:
- Use AI for pattern recognition tasks where it demonstrably outperforms manual review — radiology screening, genomic variant annotation, survival curve generation from large datasets
- Do not use AI as the sole arbiter of diagnostic or treatment decisions where mechanistic understanding is required
- Require independent verification when AI output influences irreversible clinical pathways
- Monitor for deskilling — if clinicians begin deferring to AI in situations where their own diagnostic reasoning should be exercised, the institution has a governance failure, not a technology success
The Central Question: Will AI Make Us Better Oncologists or Just Faster Ones?
The defining question is not whether AI works in oncology. It demonstrably does, in specific applications where pattern recognition at scale exceeds human capacity. The question is whether AI adoption will make oncologists better diagnosticians or merely faster ones.
Better means deeper mechanistic understanding, sharper differential reasoning, more precise risk stratification. Faster without better is operationally valuable but intellectually reductive. If AI compresses time-to-decision without compressing cognitive quality, it is a genuine advancement. If it compresses time-to-decision by bypassing the cognitive steps that catch edge cases, it is a regression disguised as progress.
Used wisely, AI expands human capability by handling computational burden, freeing clinicians to focus on what machines cannot replicate: causal reasoning, contextual judgment, and the integration of patient-specific factors that fall outside algorithmic training distributions. Used indiscriminately, it dulls those very faculties by making shortcuts irresistible.
The Responsibility Remains Human
AI is an amplification tool. It does not absolve oncologists of diagnostic responsibility. It does not replace the need for mechanistic understanding. It does not eliminate the requirement for independent verification in high-stakes decisions.
The responsibility, ultimately, remains human. That is not a limitation of AI. That is the appropriate distribution of accountability in a domain where errors have irreversible consequences.
If we approach AI with epistemic rigor — deploying it where it demonstrably enhances capability, constraining it where it creates false confidence, and maintaining the diagnostic discipline that catches what algorithms miss — it becomes a genuine asset. If we approach it with technological enthusiasm unchecked by governance, it becomes a liability.
The choice is ours.
