AI Cancer Treatment Planning: Useful Support Requires Strong Safety Checks

Why a treatment plan is harder than a search result

New software can read records, compare guidelines, and suggest treatment options in seconds. Speed is the easy part.

A cancer treatment plan pulls together the diagnosis, the stage, biomarker results, other health conditions, past treatment, side effects, current evidence, and what matters to the person. NCI's guidance on biomarker testing makes one piece of that clear: each person's cancer has its own pattern of biomarkers, and some treatments only work when a particular biomarker is present. Miss that detail and the rest of the plan is wrong.

This page explains federal sources. It is not medical advice and does not suggest a test or treatment.

What NCI says about AI in cancer care

NCI supports AI research across the whole span of cancer work, including health care delivery. Its framing sets a condition: these tools have promise when they are built and used in an ethical and scientifically rigorous way.

NCI gives examples of what that looks like in practice. Researchers are using large language models to pull knowledge out of the scientific literature. That is a research task with a human reader at the end of it, not a substitute for a clinician.

What a confident answer can hide

A well-written answer is not the same as a correct one. Software can sound certain while working from a record that is incomplete, out of date, or wrong.

A report on a planning tool should be able to say:

• What information the system read, and what it could not see.
• Whether a clinician can trace each suggestion back to current evidence.
• How often experts disagreed with the output, and why.
• Who signs the final plan and carries responsibility for it.

If a story cannot answer those four questions, it is describing a demonstration, not a tested tool.

Test the workflow, not just the model

Cancer care is not one decision. It moves as new scans, laboratory results, side effects, and goals come in. A single output cannot stand in for that.

So the useful test is of the whole workflow. Researchers should feed the system messy records and unusual cases, not tidy examples. They should record how often clinicians reject a suggestion, whether the tool adds delay, and whether any error reached a patient.

When the software updates, the evidence may not

This is the quiet problem. Checks must continue after each update, because a new version may not behave like the version that was studied.

FDA keeps a public list of AI-enabled medical devices cleared for marketing in the United States. Devices on it have gone through a review of safety and effectiveness for their stated use. FDA also says plainly that the list is not a complete record of every AI-enabled device, and it is working on ways to flag devices that use large language models. A product being marketed is not the same as a product being proven for your situation.

What software cannot take over

• A language model is not a cancer specialist and cannot carry responsibility for care.
• A polished answer does not prove every source or patient fact was weighed.
• Doing well on test cases does not establish safety in a real clinic.
• AI may help a qualified team. It does not replace professional review or a second opinion when one would help.

The shape of cancer treatment itself is unchanged by these tools. Surgery, radiation, drug treatments, and supportive care are still chosen by people who know the case.

Questions to raise with your team

• What information did the system use, and what might be missing?
• Can you trace each suggestion to current evidence?
• Where do you disagree with it, and why?
• Who checks for errors and stays responsible for the decision?

How this article was prepared

An AI-assisted editorial system helped prepare this page. No named medical reviewer has reviewed it unless one is listed.

The National Cancer Information Foundation publishes Cancer Explained. This page is for learning. It is not medical advice and does not suggest a test or treatment.

See an error, old source, or unclear wording? Tell us.