I build imaging models and I read scans for a living, which means I get asked the anxious version of the AI question a lot. Is it coming for the job. The honest answer is more interesting than yes or no, and it starts with being specific about what these tools are actually good at right now.
What AI is genuinely good at
The strongest current use is not diagnosis. It is attention and consistency. A model does not get tired on the two hundredth scan of the day, does not anchor on the first striking finding and skim the rest, and does not quietly lower its standard at the end of a long list. Those are real human failure modes, and a well-built model that flags regions worth a second look is a genuine help against them.
The second strong use is the unglamorous work around the read. Structuring reports, keeping terminology consistent, pulling the right prior study, transcribing dictation without mangling dental terms. None of this is diagnosis, and all of it eats time that should go to actually looking at images. This is where I have gotten the most concrete value, and it is where I would tell any department to start.
What it is not good at yet
It is not good at the parts of the read that depend on judgment under uncertainty. Weighing a subtle finding against the clinical context, deciding whether an ambiguous change is real or an artifact, knowing when the study does not answer the question and saying so. These require a model to be calibrated about what it does not know, and calibration is exactly where current systems are weakest. A model that is confidently wrong is more dangerous than one that is uncertain and honest, and confidently wrong is the easy failure mode to build by accident.
It is also only as good as the agreement of the people who taught it. In areas where expert readers genuinely disagree, and there are many, the labels carry that disagreement as noise. A model cannot be more consistent than the ground truth it learned from, and pretending otherwise is how you get a tool that launders human uncertainty into a number that looks authoritative.
Why I still read the scan
I treat a model as a second reader with a different set of failure modes than mine. That framing is the whole thing. My mistakes are fatigue, anchoring, and the occasional skipped view. A model’s mistakes are shortcuts in the data, overconfidence out of distribution, and blindness to context it was never shown. Because our errors are different, we catch different things, and the combination is better than either alone. But that only works if I can still open the volume and check the model’s claim on the merits. The moment I cannot do that, the model has not reduced the uncertainty in the read. It has just moved it somewhere I can no longer see.
So I read the scan myself, and I build tools that make my read faster and more consistent rather than tools that try to replace it. That is not caution for its own sake. It is the arrangement that actually produces better reports today, and it is the one I would trust with my own imaging.
The short version for a worried resident
The job is not going away. The job is changing, and the people who do well will be the ones who understand both sides well enough to know when to trust the tool and when to overrule it. A radiologist who can read and who understands how the models work is not competing with AI. They are the person the whole thing depends on.