UB Researchers’ Push to Detect AI-Written Radiology Reports Opens a New Integrity Front
Researchers at the University at Buffalo are developing a tool to identify AI-generated radiology reports, signaling growing concern over provenance in clinical documentation. The effort reflects a broader shift from asking whether generative AI can draft reports to whether health systems can verify what was human-authored, machine-assisted, or fully machine-generated.
Generative AI in radiology has mostly been discussed as a productivity tool, but provenance is fast becoming the harder problem. A detection tool for AI-generated radiology reports suggests that institutions are beginning to worry less about whether LLMs can produce plausible text and more about how to govern clinical documents once machine authorship becomes hard to distinguish.
That concern is not merely academic. Radiology reports are clinical, legal, billing, and communication artifacts all at once. If AI can draft them fluently, the risks are not only hallucinations or omissions, but also uncertainty about responsibility, supervision, and auditability. A detection layer could become part of compliance infrastructure, especially in settings where report generation is semi-automated.
There is also a strategic irony here: healthcare may soon need AI to police the outputs of other AI systems. That does not solve the core governance issue, but it does indicate where the market may go next. Report provenance, signature workflows, metadata standards, and disclosure policies could become just as important as the generation models themselves.
The larger lesson is that clinical AI adoption is entering a second phase. First came experimentation with drafting and summarization; now comes the institutional need to verify, monitor, and document machine involvement. In that sense, tools that detect AI-authored radiology text are not side projects. They may be early components of the trust stack healthcare will require for routine generative AI use.