I’m headed to HLTH to participate in a panel titled “Human in the Loop or Human in the Way?” The title says it all.
Nowadays, healthcare AI offerings often come with an assurance: don’t worry, we’ve ensured that humans are still in the loop at every step along the way. However, as models demonstrate increasingly sophisticated reasoning capabilities, would having a human in the loop at every step stifle innovation?
The answer, as you might imagine, is not as binary, and not as simple. There are three discrete scenarios I want to touch on, and each may warrant a different degree of human involvement.
Human as a detective
Let’s start with the problem: today, radiologists are under an immense amount of stress. Estimates suggest that while their workload has increased by 80% over the last decade, the 10% workforce growth hasn’t kept pace.1
Here’s where AI could help by freeing up radiologists’ time, so that they can focus on the cases that really matter.
However, we should be careful about introducing “pixel” AI. If it just adds more alerts, more clicks, and more worklists, it makes the day harder. Pixel AI tools that flag nodules or highlight lesions can backfire when false positives force radiologists to chase noise.
What are our customers telling us?
They say that a substantial opportunity lies in the work surrounding interpretation. These are the workstreams that offer opportunities for automation. AI can summarize a chart, bring relevant history into view and draft a report for review. Studies of particular applications report productivity gains in the 30–50% range, although the gains depend on the task and how the technology fits into the workflow.2
Radiologists can then spend their time doing what motivated them to join the profession in the first place: solve difficult problems, see the person behind the image and make a human connection.
Human as a boss
No one likes a boss who micromanages. Ideally, you want a human to get involved only when there’s sufficient reason.
The level of oversight should reflect the intended use, degree of automation, and potential consequences of an error.
Consider an AI offering assembling a draft report from existing findings. Its job may be to organize information already available rather thancreating any new clinical findings, and its output can be reviewed and verified. Findings can be linked to their underlying sources to supportclinician review and verification. Showing sources for every piece of information contained in an output can help the clinician navigate the report easily and validate the AI’s findings.
And where does human judgment come in? In circumstances that are more consequential.
Take the example of an AI offering that helps hospital operations cope with unexpected events like a surge in demand. A model might analyze patient data over time to estimate when individual patients are likely to be discharged, updating those estimates hourly as new information becomes available. It could recommend actions and prioritize operational workflows to help teams respond to pressure on capacity.
However, sending a patient home is a more serious matter. Ultimately, a clinical team has to be responsible for that determination. Here’s when you absolutely need to have a human in the loop.
Human as a teacher
Humans have a vital role to play in making sure that systems improve over a period of time and become more helpful in helping them cope with the many stresses acting on the healthcare system.
Teams at GE HealthCare are exploring connected AI workflows that could support clinicians across the imaging pipeline, with human checkpoints at key steps. As an illustrative future workflow, a radiologist might ask the system to perform a coronary review. Bounded AI agents would retrieve the coronary CT angiography images, identify anatomy or lesions, and create views that follow the course of the vessels. The system could then assess whether those vessels are open or narrowed and assemble a draft structured report.
Along the way, it would surface uncertainty and require explicit clinician confirmation before proceeding at key steps. The radiologist could examine a finding, question an interpretation and correct something the system had missed.
Those interactions could also provide feedback for evaluating and improving future versions of the system. A correction should be captured and assessed through a controlled process before it informs a change in clinical use.
That feedback is especially valuable in specialized fields. As new findings continue to emerge, their implications are not always straightforward. Domain experts help determine how that knowledge should be applied and where a system’s interpretation falls short.
Ultimately, there’s no one solution when it comes to humans in the loop. Be it as a detective, a boss or a teacher, clinicians can play varying roles to help their peers cope with ever-increasing workloads and free them up to make human connections.
- Harvey L. Neiman Health Policy Institute / Journal of the American College of Radiology, 2024 ↩︎
- Friebe M. “AI in radiology and interventions: a structured narrative review of workflow automation, accuracy, and efficiency gains of today and what’s coming.” International Journal of Computer Assisted Radiology and Surgery, published online November 17, 2025; issue year 2026. https://link.springer.com/article/10.1007/s11548-025-03547-2 ↩︎
