Last Tuesday in clinic, a patient asked me whether AI would replace radiologists before her next annual follow-up, and I caught myself pausing longer than I wanted to. I had just left a hospital committee meeting where a vendor had promised “predictive workflow intelligence” for everything from inbox triage to sepsis alerts, and the contrast was almost comic. In one room, a human being wanted clarity about her care. In the other, a dashboard was trying to sell certainty.
The next five to ten years of medicine cannot be forecast with the confidence many people pretend to have, because AI progress is shaped by compounding technical, economic, and institutional forces that do not move in a straight line. The right response is to build hospitals that can absorb surprise, regulate vendors carefully, and protect clinical judgment while the curve keeps bending.
I used to think the smartest way to talk about AI in medicine was to estimate timelines, map likely product categories, and sketch a neat five-year plan. Then I watched benchmarks improve, saw clinical workflows get rearranged by tools nobody had previously taken seriously, and read scenarios like AI 2027 alongside more cautious work on long-horizon limits and recursive self-improvement. Now I think the real skill is not prediction. It is institutional weatherproofing.
Why forecasting breaks down fast
The singularity metaphor matters because it describes a point of curvature, not a line item in a strategic plan. Once systems begin to accelerate their own research, product cycles, or model improvement, the error bars widen faster than board slides can keep up. The public conversation often assumes a smooth slope, but medicine lives inside a messier reality: procurement delays, regulatory review, liability fears, union contracts, EHR integration, and the ordinary human friction that slows even excellent tools.
AI 2027 is useful because it refuses the lazy habit of speaking in vibes. Its authors explicitly frame superhuman AI as something to reason about with concrete scenarios rather than slogans, and they note that their forecast is a best guess, not a recommendation. That posture should make clinicians humble. If even the people studying frontier capability say they cannot know the exact path, hospital leaders should stop pretending their vendor roadmap is destiny.
One reason I take this seriously is that I have seen how quickly confidence outruns validation. A model that looks persuasive in a demo can still stumble in actual clinical flow, where a wrong summary lands in the wrong inbox or a “helpful” draft displaces the one sentence a physician needed to preserve nuance. The failure mode is usually not cinematic. It is bureaucratic, slow, and expensive.
What the evidence actually supports
Several recent papers help frame the problem. In Faith in AI can narrow the futures individuals consider, Naito and Shirado report that greater trust in AI can narrow the set of futures people imagine, a useful warning for health systems that let automation become a substitute for judgment. In parallel, SAHOO: Safeguarded Alignment for High-Order Optimization Objectives in Recursive Self-Improvement shows why recursive self-improvement raises alignment risks that do not fit neatly into conventional model QA. In a different vein, Khanh and Hoa’s Dynamic Intelligence Ceilings argues that long-horizon planning and creativity remain bounded in ways that should temper overconfident claims about near-term omniscience.
These papers point in the same direction. We should expect uneven progress, not a smooth glide path. We should expect narrow failures in tasks that look broad from a distance. And we should expect organizations to misread partial success as proof of general reliability.
That is why I pay attention to operational numbers, not hype. The U.S. Food and Drug Administration’s device pathways, including 510(k), De Novo, and PMA, exist because regulators know that evidentiary burden has to match risk. The same logic applies to clinical AI governance. A model that writes a draft message for a dermatology teleconsultation is one thing. A model that influences triage, imaging urgency, or discharge decisions is another. In practice, the question is whether a hospital can demonstrate traceability, fallback, and accountability before deployment, not after the first adverse event.
For context on one healthcare use case, the review A Review on Teledermatology in Saudi Arabia describes both the promise and the barriers, including access, workflow fit, and implementation constraints. That is the part many AI discussions omit. A tool can be technically impressive and operationally awkward at the same time. Medicine experiences both.
The physician-executive lesson
When I evaluate an AI vendor, the first question I ask is embarrassingly simple: what happens when this tool is wrong at 2 a.m. and the covering clinician has never seen it fail before? I have learned that the best systems are not the ones with the slickest demo, but the ones that leave a clean audit trail, degrade gracefully, and make it easy for a clinician to override them without a fight. If a product cannot do that, I do not want it in a patient-facing workflow.
What I would not do: I would not build hospital strategy around a single forecasted date for AGI, nor would I budget as if every current task category will remain stable long enough to be managed on a five-year capital plan. I would not let “the model said so” become a proxy for clinical authority. I would not let a vendor define success in terms that cannot be independently audited by our quality and safety teams.
I am also skeptical of the idea that more faith in AI always improves decision-making. The research on faith narrowing futures is a warning for clinicians too. If we become overconfident, we stop asking what else could happen. That is a bad habit in medicine, because surprise is where harm lives.
There is a useful counterpoint in engineering work on energy-frugal or safeguarded recursive architectures, including S-AI-Recursive, which suggests that efficiency and introspection can be designed into systems rather than added later as afterthoughts. I like that framing because hospitals are already constrained by staffing, power, compute, and budget. Any AI future that ignores resource costs will fail first in the places with the least slack.
Clinical vulnerability matters here. I have been wrong before about which tools would matter. I expected some categories to stay niche. Instead, a few humble utilities, especially those that save minutes across thousands of encounters, became more important than the flashier systems everyone wanted to discuss. That taught me to respect boring adoption curves. They often outlive the keynote.
How hospitals should plan for the unknown
Hospitals do not need prophecy. They need adaptive governance. That means scenario planning with multiple time horizons, a formal review path for AI systems, drift monitoring after deployment, clinician opt-out mechanisms, and an explicit rule that high-risk systems do not go live without human-centered fallback. It also means treating AI like infrastructure, with uptime, maintenance, and failure modes, rather than like a novelty feature bolted onto the EHR.
I think the most honest plan is to prepare for three futures at once. In one, progress is steady and useful but bounded. In another, capability jumps create rapid workflow reorganization. In a third, the bottleneck is not intelligence but governance, litigation, and patient trust. All three can be true in different departments at the same time. Radiology may move faster than pharmacy. Documentation may move faster than diagnosis. The board will want one answer. Reality will not provide one.
For physicians, the practical question is not whether the singularity arrives on schedule. It is whether our institutions can absorb discontinuity without abandoning safety. I think about that every time a patient asks me if AI will change her care. The honest answer is still: probably, yes. Exactly how? I do not know, and anyone who says they do is selling a story more than a forecast.
Back in clinic
After that Tuesday visit, I went back to the patient and answered more plainly. I told her that the useful question is not whether a machine will someday think faster than we do, but whether her care team will still be able to explain decisions, catch errors, and keep her safe while the tools change around them. She nodded and said, “So basically, you want the machine to help, but not run the room.” That was exactly right.
That is where I land now. I used to think the main task was to predict the curve. Now I think the main task is to build clinical systems that stay humane when the curve stops being predictable.
For more on my background, you can see Dr. Sina Bari, Stanford-trained physician and clinician-educator. I keep coming back to the same point because it matters at the bedside and in the boardroom alike: uncertainty is not a planning failure. It is the plan.
FAQ
How should a hospital board think about AI that may improve too quickly to forecast?
Assume the forecast will be wrong and design for reversibility, oversight, and auditability. A board should ask for escalation paths, drift monitoring, and a clear clinical owner for every deployment. If those pieces are missing, the product is not ready for high-stakes use.
What happens if a clinical AI tool works in a demo but fails in real workflows?
The failure usually shows up as delay, confusion, or silent error rather than a dramatic crash. In medicine, that can mean extra chart review, missed nuance in a note, or an alert that no one trusts. I care less about whether the demo was impressive and more about whether the tool survives a night shift.
What is Dr. Sina Bari's approach to AI governance in hospitals?
I favor conservative deployment, explicit accountability, and strong fallback behavior. If the system cannot be explained to frontline clinicians and audited by quality teams, I do not think it belongs in a patient-facing workflow. The standard should be safety first, novelty second.
Can hospitals plan around recursive self-improving AI at all?
Yes, but only at the level of scenarios and safeguards, not exact dates. The sensible move is to prepare governance for multiple trajectories, including rapid capability jumps and slower institutional adoption. Planning for the shape of uncertainty is more useful than guessing the winning year.
Why do AI predictions narrow the future clinicians consider?
Because confident narratives can make people stop imagining alternatives. That matters in healthcare, where overconfidence leads to brittle systems and missed second-order effects. The better habit is to keep multiple futures in play until the evidence settles them.