Some Insights from the MIT AI Conference
At last week’s MIT AI Conference, a panel of scientists and industry leaders gathered to tackle a deceptively simple question: how do we build artificial intelligence that clinicians can trust?
More specifically, how to we get to AI precise enough to handle life-or-death decisions yet transparent enough for a physician to confidently use?
This wasn’t a debate about flashy demos or futuristic promises. Instead, the chat hinged on the rigorous engineering that lies beneath the surface of medicine’s AI revolution.
The panel featuring Andy Beck (CEO of Path.ai), Sadegh Salehi (Principal ML Scientist at Overjet), Sam Sinai (Co-founder of Dyno Therapeutics), and moderated by Manolis Kellis of MIT and the Broad Institute united around one critical theme: medical AI needs determinism.
In practical terms, this means the same input must always yield precisely the same output, every single time, without exception.
After all, medicine has little patience for improvisation. As Beck explained, generative models like ChatGPT, despite their popularity and versatility, won’t suffice in clinical contexts. They’re remarkable conversationalists, but their occasional flights of fancy and unpredictable outputs clash dramatically with healthcare’s unyielding standards of clarity and reliability.
Take pathology, Beck’s own specialty. Many patients assume that modern pathology already relies on advanced, digitized systems, but Beck described a reality far less futuristic. Human pathologists still interpret thousands of tissue samples manually, examining thin slices under a good old fashioned microscope.
It’s painstaking, exhausting work and prone to human error. What AI offers isn’t simply automation, but precision, those deterministic systems trained on immense, diverse datasets. Crucially, Beck noted, human experts will remain at the core of this process, guiding and validating AI’s output rather than being replaced by it. This hybrid approach, he argues, promises far lower error rates, dramatically reducing misdiagnoses and enhancing patient outcomes.
The demand for deterministic precision extends across medicine. Dentistry even came up. Sadegh Salehi of Overjet described the complex processes dentists face (diagnosing tooth decay from ambiguous images, clinical notes, and patient history for example). Salehi argued forcefully for transparency. AI recommendations, he insisted, must clearly state their rationale, linking each diagnosis to visual evidence and historical records.
The goal isn’t blind trust, but evidence-driven confidence. Dentists need not just an accurate AI partner, but one whose reasoning they can fully understand and verify.
But deterministic precision alone isn't sufficient.
Medical AI faces another daunting hurdle in the validation department.
Unlike purely digital applications where outputs can be instantly and cheaply confirmed, biology demands costly, time-consuming real-world experiments.
Sam Sinai from Dyno Therapeutics knows this intimately. He designs viruses to deliver therapeutic genes precisely to target cells, an elegant concept whose biological validation comes with a staggering price tag. Each tiny error or inconsistency in AI predictions costs months and millions, he explains. Deterministic, thoroughly validated AI is not a luxury, but an imperative.
Underlying this shared urgency for determinism is biology’s sheer complexity, a point passionately underscored by moderator Manolis Kellis (MIT/Broad Institute).
Life itself, Kellis reminded the audience, has spent billions of years refining staggeringly sophisticated biological mechanisms. Humans, even aided by AI, remain just at the edges of fully understanding these intricacies. Yet paradoxically, biological complexity isn't an obstacle, it's what drives the necessity for deterministic AI.
The more complicated and nuanced biology becomes, the greater our need for rigorous, transparent, and consistently accurate technologies.
This deterministic approach isn't just engineering precision, it's rebuilding medicine’s fragile trust. It rejects the black-box mysticism of earlier AI experiments in favor of openness and reliability. Precision and transparency become intertwined, each reinforcing the other.
Clinicians, naturally skeptical and protective of their patients, will embrace only those technologies that demonstrate clear benefit and total transparency. As Sinai said, the burden of proof rests entirely on innovators, if the technology works reliably and demonstrably improves patient care, adoption becomes inevitable.
The panel’s collective vision isn’t of robotic automation replacing doctors, but of deterministic AI deeply embedded within clinical workflows, offering meticulous clarity and augmenting human judgment.
Far from diminishing human roles, this deterministic revolution enhances them profoundly. Physicians remain central but now armed with tools previously unimaginable in their precision and reliability.
One other theme all panelists could agree on was that precision medicine is at a striking crossroads. MIT AI Conference panelists made it abundantly clear. For AI to genuinely revolutionize medicine, it must earn trust through uncompromising determinism, transparency, and validation.
