Over the past few weeks, two of the most prominent voices in artificial intelligence have quietly revised a prediction. A year ago, Sam Altman warned that entry-level work was at serious risk, and Dario Amodei said AI could eliminate half of white-collar jobs. Both have since stepped back. Altman now says he was “pretty wrong” about the near-term economic damage, and Amodei has begun describing AI as something that expands the work people do rather than erases it. The displacement they forecast has not arrived in the form they expected.
Economists reach for the Jevons paradox to explain it. When something becomes much cheaper, we tend to use more of it, not less. Cheaper computing did not shrink the software industry, it grew it. Whether that holds for cognitive work across the whole economy is an open question. In oncology, where I practice, the reversal is not surprising at all. The help these tools have actually provided was never aimed at the part of the job people were afraid of losing.
The fear assumes AI will do the doctor’s work: read the scan, make the call, have the difficult conversation. That is not where it has been useful. It has been useful for the work that surrounds those decisions, the reading and the arithmetic that no clinician has ever had enough time to finish.
What actually limits a doctor
What limits me in clinic is not a shortage of doctors willing to decide. It is that there is more to know than any one person can hold.
For a single patient, the evidence that should inform her care is spread across dozens of trials, long-term follow-up analyses, genomic findings, guideline revisions, and toxicity data, and it changes faster than anyone can track while seeing a full day of patients. So I do what most oncologists do. I keep the major trials in mind, rely on patterns I trust, and when a patient asks what a treatment really means for her, I give her my considered judgment rather than a precise figure, because the precise figure would take hours the schedule does not contain.
That is a limit of arithmetic, not of effort. There is more reading, and more calculation, than a clinical day allows. And that is the kind of work these tools handle well. Not the judgment, and not the conversation, but the reading and the keeping-up that was going undone alongside them.
What the help actually is
Most of the public debate concerns what AI removes. Less attention goes to the work it makes possible.
At Kesis & Sisters, the clearest example is Chloe, the part of our work devoted to keeping current with the literature. She does the reading the schedule never allowed: working through dense oncology papers, extracting the figures that matter, and turning unstructured clinical reports into something organized and usable. No clinician can keep a complete view of the evidence current across even one cancer while seeing patients all week. With the right support, that becomes feasible, and the guidance we build rests on evidence that stays current rather than evidence that settled the last time someone had a free evening.
It shows up in specifics. In our work on gastrointestinal stromal tumours, the literature is monitored continuously rather than loaded once and left to age. When something meaningful changes, the guidance can change with it. In a disease where a single mutation result can point toward a different treatment entirely, the distance between current evidence and evidence eighteen months old is the distance between helping a patient and missing something that mattered.
The part I care about most is the conversation itself. Much of what a trial establishes gets flattened by the time it reaches a patient, until a careful result becomes “yes, this works, we’ll use it.” The aim is to give that precision back: take the actual profile of the patient in front of me, set it against what the evidence shows for someone like her, and say something honest about what she can expect. That is a conversation I always wanted to have and rarely could, less for lack of knowledge than for lack of time. It is becoming possible, and it is a better conversation for her.
None of this replaces the clinician. It clears away the work that crowds out the human part, leaving more room for the judgment and care only a person provides.
Where the line is
There is a limit worth being clear about. I would not let a language model do the math for a patient.
These tools are impressive with language and unreliable with calculation, and the unreliability is the dangerous kind: confident, and wrong often enough that the number cannot be trusted on its own. I would no more ask one to compute a survival benefit than to choose a chemotherapy dose. That is not what it is for.
Where it helps is in building something steadier. The model does the reading and helps assemble a proper calculator, one that produces the same result every time, traceable to its source, with nothing hidden in the middle. The numbers a patient receives come from logic we can inspect and stand behind, not from a model’s guess in the moment. Chloe does the reading it is suited to. The calculation runs separately and reliably, the same on Tuesday as on Monday, the same for one patient as for another.
The common mistake in this field is to point a chatbot at a clinical question and trust the answer. That is the wrong tool for the task. The sounder approach is less visible and more durable: use AI to help build the tool, then rely on the tool, not the chatbot, at the point of decision.
More room for the part that matters
This is why the reversal reads as unsurprising from inside the work.
What cancer care is short on is not doctors. It is time, set against a volume of evidence no one can fully absorb, in a clinic that stays busy. Taking some of that weight off does not push the doctor out of the decision. It gives the decision the grounding it always needed. More patients get an honest conversation with real numbers in it. More of what a treatment can offer reaches the person who needs it. More of the good work that was always possible, but rarely reachable, gets done.
There is a commercial version of this as well. When a good drug reaches fewer patients than forecast, it is usually not because oncologists are ignoring it, but because applying it confidently to a specific patient was not feasible in the moment of decision. That was a stubborn problem for years. It is now solvable, at a fraction of the cost of generating the evidence in the first place.
None of that is the part of the work I would want to keep. The part worth keeping is sitting with a frightened person, holding the numbers steady, and helping her choose well with honesty and care. That is what these tools should protect. Not a future that empties the clinic, but one that returns the time to practice medicine well.
If this resonates with something you’re working on, I’d like to hear about it. Contact.
Dr. Henry Conter is a Medical Oncologist and Hematologist at William Osler Health System and the founder of Kesis & Sisters. He trained in Medical Oncology at MD Anderson Cancer Center and spent six years at Hoffmann-La Roche in progressively senior roles spanning oncology clinical development, portfolio strategy, and medical and regulatory affairs, across both national and global functions.