A pathology lab at a mid-sized cancer center gets a new AI tool on a Tuesday. By Friday, attendings are quietly double-checking its slide reads before signing off, not because they were told to, but because the model flagged a handful of cases they would have called differently. Nobody logged those disagreements anywhere. That gap, between a tool entering clinical work and a system for tracking what it does once it’s there, is where the real story of AI in oncology sits right now.
The technology itself has moved quickly. Machine learning models now assist with tumor detection on imaging, triage pathology slides, mine electronic health records for eligible trial participants, and model drug responses before a single patient is enrolled. CancerNetwork’s recent look at the field captures a shift that clinicians have been sensing for a while: AI is no longer a demo in a conference hall. It’s showing up in trial design, in radiology queues, and in the unglamorous administrative layers that determine whether a study finishes on time or stalls out.
The trial angle is the one worth watching closely. Patient matching has long been one of oncology’s quiet bottlenecks. A promising protocol can sit half-full for months because the right patients never got screened, or got screened too late. AI tools that scan unstructured clinical notes for eligibility criteria promise to compress that timeline. That’s a genuine operational win, and it’s also where things get murky. If a model decides a patient is a poor fit and that decision never surfaces to a human reviewer, you’ve automated a judgment call that regulators, ethicists, and frankly most oncologists would want a person to own.
That tension is why the regulatory conversation is heating up. Groups like RAPS have been pushing for a framework built specifically around clinical AI models rather than trying to stretch existing device rules to cover software that updates, learns, or behaves differently across patient populations. The core problem is familiar to anyone who’s followed AI governance generally: a model cleared on one dataset doesn’t necessarily perform the same way in a hospital with different imaging equipment, different coding practices, and a different patient mix. Static approval doesn’t map cleanly onto software that drifts.
There’s a legal dimension too. The Florida Bar’s proposed advisory opinion on AI tools is a reminder that professional responsibility rules were written for a world where a human made every judgment call. Medicine is heading toward the same reckoning. If an AI-assisted diagnosis leads to a missed cancer, who carries the liability? The vendor, the hospital, the attending who clicked past a warning? Nobody has a clean answer, and the ambiguity is already shaping how cautiously institutions deploy these systems.
The cybersecurity thread running through all of this deserves more attention than it usually gets. Anthropic researchers have warned that connected devices, the kind that now populate hospital wards and even homes, can become attack surfaces for AI-driven threats. In oncology, where imaging systems, infusion pumps, and records platforms all talk to each other, that’s not an abstract concern. A model that ingests patient data is also a model that can be poisoned, probed, or exfiltrated. The same AI wave transforming diagnostics is transforming the threat model around them.
None of this argues for slowing down. The case for AI in oncology is strong, and the field’s enthusiasm is earned. Trial enrollment gets faster. Radiologists get triage support. Researchers get patterns they’d never spot by hand. But the current moment feels less like a breakthrough and more like an awkward adolescence. The tools work well enough to matter and poorly enough to require constant human oversight, which is exactly the phase where institutions tend to underinvest in monitoring.
What’s missing isn’t more models. It’s the boring infrastructure around them: logging when a clinician overrides a recommendation, tracking performance drift across demographics, defining who’s accountable when the system is wrong. Some health systems are building that. Many aren’t, because it’s expensive and invisible until something goes wrong.
The regulatory frameworks being debated now will determine whether that infrastructure becomes standard or stays optional. If history is any guide, it’ll take a high-profile failure to force the issue. The question is whether oncology wants to wait for that, or build the guardrails while the tools are still new enough that nobody’s defensive about them yet.