Oncologists Can Experiment With Generative AI Now: A Safe, Practical Roadmap

Why Oncologists Should Begin Testing AI—Cautiously

Artificial intelligence is flooding healthcare, and oncology is no exception. While fully integrating generative AI or retrieval-augmented generation (RAG) into cancer care remains premature, experts are urging clinicians not to sit on the sidelines. Matthew Matasar, MD, recently outlined a pragmatic, low-risk approach for oncologists who want to test these tools in their own workflows right now—long before they touch a patient chart.

The core message: start small, stay far from clinical decision-making, and use the time to learn what AI can and cannot do in the complex world of cancer care.

Start With Narrow, Well-Defined Use Cases

Matasar emphasizes that the first forays into generative AI should never involve patient diagnosis, treatment planning, or direct communication. Instead, oncologists and their teams should identify tightly scoped, non-clinical tasks where errors carry minimal risk and human oversight is complete.

Examples of safe starting points include:

• Drafting summaries of internal research papers or clinical trial protocols for educational purposes
• Generating plain-language explanations of complex oncology terms for staff training
• Helping to organize or rephrase meeting notes and administrative memos
• Suggesting structured templates for tumor board documentation that are then manually reviewed

By confining AI to these rings of low consequence, clinicians build familiarity with the technology’s behavior, its tendency toward hallucination, and the critical need for traceability in every output.

How “Safe Testing” Actually Looks in Practice

A major concern for any oncology practice is patient safety and data privacy. Matasar’s framework for safe experimentation relies on a few non-negotiable principles:

• De-identified or synthetic data only – No real patient information enters the AI tool at this stage. Teams should use dummy cases or publicly available datasets stripped of identifiers.
• Non-patient-facing tasks – The output must never reach a patient or influence a patient’s care path without a human gatekeeper. The AI is a lab experiment, not a clinical assistant.
• Mandatory human review – Every AI-generated suggestion, summary, or draft must be checked by a qualified clinician who remains fully accountable for the content.
• Clear limits on clinical responsibility – The experimenting clinician must define upfront that the AI has no authority over medical decisions. This boundary is both ethical and regulatory.

These guardrails echo broader institutional concerns that are already part of oncology’s compliance culture, but they must be explicitly applied to generative AI sandboxes.

The Distinction Between Administrative Help and Clinical Aid

One of the most important lines Matasar draws is between AI for support tasks and AI that touches the clinical pathway. Current experimentation should focus overwhelmingly on the former—reducing administrative burden, enhancing efficiency, and streamlining communication workflows that already require human certification. The moment an oncologist asks an AI “what chemotherapy regimen should we try next?” the experiment has crossed into dangerous territory.

Generative models, including those enhanced with retrieval-augmented generation (RAG) that can pull from knowledge bases, still lack the nuanced understanding of an individual patient’s comorbidities, preferences, and the nuanced biological behavior of their cancer. RAG may reduce hallucinations by grounding responses in specific documents, but it does not eliminate them, nor does it understand context the way a board-certified oncologist does.

Practical Barriers Oncologists Will Face

Even in low-risk pilot projects, real-world hurdles emerge quickly. Matasar’s advice acknowledges several obstacles that practices must address before opening up a test environment:

• Data privacy and HIPAA – Even de-identified data can trigger compliance concerns if not handled strictly within approved systems. Institutional IT and privacy officers must be looped in early.
• Hallucinations and reliability – Generative AI can produce plausible but dangerously wrong medical statements. Oncologists must verify every fact against established guidelines and primary literature.
• Bias in training data – AI models may reflect disparities already present in cancer research datasets, requiring ongoing vigilance to avoid skewed outputs.
• Institutional policy – Many hospitals and cancer centers have not yet published clear guidance on the use of large language models. Gaining leadership buy-in and aligning with IT governance is a prerequisite.

Overcoming these barriers is part of the learning process. The goal of current experimentation is not to replace human judgment, but to understand exactly where AI can be trusted—and where it cannot—in the oncology workflow.

Evaluating AI Tools for Oncology-Specific Reliability

Matasar suggests that oncologists apply the same rigorous skepticism they use with new clinical data when assessing a generative AI tool. Key questions to ask:

• Can the tool show its sources and make its reasoning traceable?
• How often does it produce incorrect or misleading information in oncology-specific queries?
• Does it respect the complexity of cancer guidelines, or does it oversimplify?
• Is the model transparent about its limitations and training data cutoff?

For RAG systems, the quality of the knowledge base is everything. If a RAG tool is connected to outdated or incomplete oncology references, its outputs will be too.

A Call to Experiment—Responsibly

The overarching theme from Matasar’s guidance is that oncologists should not wait for a perfect, fully validated AI product. Instead, they can begin today by carving out safe, small-scale experiments that respect the immense responsibility of cancer care. The lessons learned now—about documentation workflows, data pipelines, and the real-world behavior of language models—will pay dividends when more robust clinical AI tools eventually arrive.

For oncology professionals, the message is clear: you don’t need to roll out AI across your practice to start benefiting from it. You just need a sandbox, a dose of curiosity, and an unwavering commitment to keeping patient care squarely in human hands.