The two black boxes and the centaur
By Josh Ewell, Founder, Frontier Radiology
John Searle asked us to imagine a man locked in a room with a rulebook. Chinese characters come in through a slot in the wall. He looks them up, follows the instructions, and pushes characters back out. To everyone outside, the room appears to understand Chinese. The man inside knows no Chinese at all. He is transforming symbols.
Searle’s question was about machines and the meaning of consciousness, but it has its own application in our field. He asked whether the room understands anything. The question I hear hospital administrators asking is the more practical version: if the product is the same, what difference does it make whether a machine or a person produced it?
In teleradiology we have built our own version of that room. Images come in through the proverbial slot, a report comes back out, and somewhere in the middle is an invisible physician with thirteen years of training that nobody outside the wall will interact with. All the hospital sees is the report and the turnaround time, or in computer vernacular, processing time.
A black box is a system you cannot see into and can therefore only be judged by its output. If that describes AI, it also describes what we have taught hospitals to expect from us. Our knowledge is real and our contribution to patient care is significant, but none of that is visible from the outside. And if from the hospital’s side of the room there are two black boxes, no detectable difference between them, and one is considerably more expensive, we have done ourselves a great disservice.
Being replaced is the loud version of this fear, and I do not think it is the version that gets us. There is a quieter arrangement already taking shape in teleradiology, and it does not require anyone to fire a single radiologist.
Cory Doctorow gives us a useful pair of terms for it in his 2026 book The Reverse Centaur’s Guide to Life After AI. A centaur is a person assisted by a machine. A reverse centaur is a person conscripted to assist a machine. You keep your job, but the work becomes checking the model’s output at the pace the model sets, and when it misses something, the signature on the report is still yours. Doctorow borrows Dan Davies’ term for that role: an accountability sink. The institution gets the speed of automation and keeps a physician on hand to absorb the liability when it fails.
It is a worse arrangement than it appears, because it erodes the judgment it claims to be protecting. Sustained review of output that is almost always correct degrades the reviewer. It is the same vigilance problem seen in airport screeners, who get worse at spotting the rare positive precisely because the system is right nearly every time. Anyone who has worked through a long list of drafted normals knows the feeling. The arrangement makes us less accurate, then holds us responsible for the misses.
And notice that none of it works unless we are already a black box. An available and much needed consultant in teleradiology is much harder to replace than a name on a paper you have never interacted with. Or put simply, it is easier to replace someone you have never met.
The way out of that arrangement is to be worth more than the signature at the bottom of the report and to do more than to rubber stamp a high volume of reports. This raises a question about what we intend to do with the time this technology gives back to us.
Assume the technology works as well as the marketing claims. Assume we recover meaningful time by minimizing friction, drafting reports, and offloading busy work.
The default answer in teleradiology is more volume, and it is not wrong. Go from a hundred studies a day to a hundred and fifty. Given how many patients are waiting on a backlog of reporting right now, more capacity is a real win. But if throughput is the only place the time goes, we have improved the processing speed of the black box and nothing else.
And here is the part that should worry us. The hospitals will be happy. Short turnaround and solid accuracy are precisely what they have asked for, and they will build their practices around it, designing workflows that assume a written report is our only value. They will learn to function without us, and once they have, they will have no reason to distinguish our output from software that does the same thing.
The alternative is to send those gains somewhere other than volume. The first place is the report itself. Sensitivity and specificity are not fixed properties of a radiologist, they move with attention and fatigue. And likewise, even the best multimodal tools hallucinate and err. Time returned to a case is the comparison you would otherwise skim, the prior you might take on faith, or the time spent looking up a finding you were not quite sure about.
The second thing is what we quietly gave away. Radiology used to be a consultative specialty. The referring physician called, or came down to the reading room, and the conversation changed what happened to the patient. The report was the medical record of that conversation, not a substitute for it. Searle’s room had no door, only a slot. Ours started with a door, and somewhere in the move to distributed high-volume work we boarded it up and put up a sign that said please use the slot for all requests.
Ask a hospital why they are unhappy with their teleradiology group, and you will infrequently hear that the reports are wrong, though I have heard that complaint as well. What I most often hear is that no one at the group can be reached, that administrators are not responsive, that radiologists do not take calls. That no one communicated the critical finding. That reports came over with delays. That the report was technically accurate but did not address the clinical question. Or all too often, it is a combination of all of the above.
Notice what these complaints imply. Referring physicians still want to reach a radiologist. The appetite for the conversations and contributions our specialty was once known for has not gone anywhere, and neither has the clinical need for it. What is missing is our availability, not their interest. The door is boarded up, but there are still people on the other side of it knocking.
The recovered time has somewhere to go. Some of it goes to productivity, but the rest has to go into being reachable: knowing our hospitals well enough to understand what a referrer is actually worried about, and picking up the phone when the case warrants it. On the conventional remote radiology pay-per-click model, none of this represents compensated work. We cannot bill for the tech’s question or the ER doctor’s consultation call. But increasing our efficiency, and with it our pay, is what buys us the headroom to be better, to do better for our hospital partners and our patients.
There is a decade–old argument that AI will end radiology for radiologists. The likelier failure, if failure is an outcome, is not from AI, but from us. It is not that the technology will surpass us; it is that we will have redefined our job as something less than it was always meant to be.
The way out is not to read faster or better than AI, which will ultimately achieve parity and superhuman performance at generating reports. I made this argument at greater length in my John Henry Generation essays for AuntMinnie if you’re interested in another read. The way out is to be the centaur rather than the reverse centaur, a specialized and clinically relevant physician that the machine serves instead of one conscripted to serve the machine.
Take the boards off the door. Welcome our colleagues back to use us as specialists and consultants whose value extends well beyond the words on the report we generate. Let the AI do what it is good at and then let it get out of the way, so that what the hospital sees is not a second black box but a physician they can reach. That is what using AI the right way looks like.
If this is the kind of radiology you want to practice, we should talk.
