Uncertainty is not the enemy of intelligence. Uncertainty is where intelligence is born. MK

The Collapsed Signal

A month before surgery, the patient’s hip X-rays told one story — the femoral head was arthritic but intact, the acetabulum thin but holding.

At the time of surgery, the story had changed. The femoral head had collapsed into itself, the socket now cratered and jagged, the architecture of motion erased.

In the operating room, the anatomy no longer matched the map. Landmarks I had memorized were gone; what should have been smooth cortical bone was now friable dust. Each move of the reamer sent new information through my fingertips — weaker density, softer than expected, no resistance — all data contradicting the plan.

At that moment, surgery ceased to be an execution and became learning. The procedure transformed from following a model to retraining it, one cut at a time. The X-ray had been a dataset; the body, the updated ground truth.

The Feedback Loop of Surgery

Every feedback loop in surgery is a live experiment in adaptation — a human form of gradient descent.

In Machine Learning, a gradient is the direction of steepest improvement; gradient descent means taking small, repeated steps downhill, each adjustment reducing error until the mathematical system stabilizes. To guide that descent, a model needs a loss function — a scorecard of wrongness measuring how far prediction strays from truth. Every step is an attempt to reduce that loss — to move closer to reality.

Now translate that mathematics into motion. In surgery, we begin with parameters: incision line, retractor placement, angle of approach. Each is a weight — an influence shaped by experience and context. If tissue tension is high, we ease our force; if exposure is poor, we reorient our retractors. Each correction tightens the loop between intent and outcome, pushing us toward understanding.

In machine learning, the process is numerical — predict, compare, compute loss, and propagate error to update weights. In surgery, the propagation is physical and cognitive. The “weights” are tactile and visual priors refined across thousands of cases. When an osteotomy is too shallow or a graft sits too proud, the loss is not abstract; it’s embodied — in the hand, in the view, in the patient’s recovery.

Early in training, I feared those corrections — each one an admission that my plan was imperfect. Over time, I learned to trust the loss signal. In both surgery and AI, progress depends not on avoiding error, but on sensing it with precision. The sharper your gradient — the clearer your awareness of deviation — the faster and more gracefully you adapt.

Expert surgeons don’t operate by memory; they operate by attention.
The novice asks, “Am I right?”
The expert asks, “What is the system telling me now?”

Just like Computer Scientists, based on instant feedback, medical experts constantly update their models in a Bayesian type of way. 

The Wisdom Waste Paradox

And yet, for all our sophistication, our tools remain silent.
Why can my car tell me when I’m tired and show me a coffee cup, but my surgical drill cannot?
Why can every athlete wear a smartwatch that coaches them in real time, while my hammer, chisel, and knife — tools unchanged for ten thousand years — offer no guidance at all?
And when I retire, why does forty years of tacit knowledge — every pattern, every near-miss, every learned instinct — vanish into the garbage can?

That is the wisdom waste paradox — the greatest opportunity lost in modern medicine.
Other fields have solved this through tech collisions: where computer science meets physics, where autonomous vehicles meet sensor fusion, where machine learning meets mechanical design.

Surgery has remained largely analog. Our instruments are slow to evolve, even as the world around us hums forward with sensors, feedback loops, and adaptive intelligence.

Embodied Surgical Intelligence: When Data Learns to Feel

Embodied Surgical Intelligence (ESI) is the next step: the fusion of human sensory skill with machine perception. It is not a device but a framework — a learning system grounded in the body itself. Sensors on the patient, the instrument, and the surgeon feed data into a shared feedback loop — not as separate streams, but as one integrated awareness.

Intelligence begins in the body. Long before cognition, there is sensation — pressure on the fingertip, torque through the wrist, resistance from living tissue. A surgeon reads these signals in real time, integrating touch, sight, and sound into a coherent mental model.

In the operating room, feedback is multisensory. The pitch of a drill, the color of synovium, the vibration of bone — each signal alone is partial; only through fusion does meaning emerge. Modern machine learning is discovering the same truth. The most advanced models learn through multimodal integration — combining vision, sound, text, and force into unified understanding.

When I first encountered sensor fusion in robotics — LIDAR, visual–inertial odometry, force feedback — I recognized it instantly. It was the digital equivalent of surgery. The robot was learning to feel.

But embodiment carries responsibility. A purely computational model can act without consequence; a physically situated one must confront it. Every cut in surgery has moral weight. As AI moves from screen to skin, designers inherit that same moral gravity: to build systems that sense not only data, but ethical impact.

The Bias of the Hand

Every hand carries its own bias. Over time, patterns of repetition harden into reflexes — the way a surgeon grips a scalpel, anticipates a curve, or sets an implant. Experience breeds efficiency, but it also narrows perception.

The surgeon’s bias is the biological analog of overfitting — learning the past too well and mistaking it for the future.

In Artificial Intelligence, overfitting happens when a model memorizes its training data so precisely it cannot generalize to the real world. Humans do the same. What once was mastery can quietly become rigidity.

The remedy is regularization — the art of tempering certainty. In code, it’s a mathematical penalty that discourages overconfidence. In surgery, it’s humility. Every unexpected bleed, every subtle error is a weight-decay term — a reminder to stay adaptive.

Bias also arises from imbalance. A career spent operating on one population breeds blind spots, just as a dataset built from narrow experience creates brittle algorithms. True intelligence — human or machine — requires diversity: of anatomy, of context, of culture, of failure.

The goal is not to erase bias, but to illuminate it — to understand its contours and prevent it from becoming invisible. Bias, after all, is memory. It tells us where we’ve looked most often. The challenge is ensuring it doesn’t define what we see next.

The Loss Function of Care

Every discipline optimizes toward something. In machine learning, it’s a loss function — a measure of how far prediction strays from truth. In surgery, that loss is measured in pain, infection, disability, or regret. Every incision carries potential for loss; every act of repair is an attempt to drive that function toward zero.

Early in my career, attending mostly to trauma, I thought the goal was survival. Later, as I entered the field of Orthopedic Surgery, the variables expanded — range of motion, pain scores, efficiency, cost. Over time, I realized the true loss function of surgery isn’t numerical. It’s embodied in the waiting family, the patient’s trust, the fatigue after the thousandth repetition. These aren’t metrics, but they shape the model all the same.

Machine learning taught me that what you choose to optimize defines the intelligence you build. Optimize for speed, and you get throughput. Optimize for empathy, and you get wisdom. Surgery is no different. If we optimize only for efficiency, we risk the erosion of deep human meaning. The better model must include dignity, compassion, and recovery as terms in the loss equation.

Loss is not failure — it’s feedback. Each deviation sends a signal. In code, it’s backpropagation; in surgery, reflection. After a complication, we replay each step, trace the chain of causality, and update our internal weights. Both processes demand vulnerability — the courage to face error not as shame, but as information.

And yet, in medicine, the loss function can never reach zero. Every patient we mend will one day fail; our task is not to defeat mortality, but to give it meaning. Empathy becomes the ultimate optimizer — redefining loss as the distance between what is and what could have been for another human being.

AI reminds us that loss is data. Surgery reminds us that loss is love.
Between those truths lies the future of systems that heal.

When the Model Learns You

At first, I thought machine learning was about teaching the system — training it to recognize patterns and correct its mistakes. But over time, I realized it was training me. Each misclassification revealed my own blind spots. Every error forced me to see probability where I once saw certainty.

Surgery is the same. A surgeon is not a static expert mastering a fixed technique, but a continuously updating model — shaped by each patient’s anatomy, each unexpected finding, each failure. The tissue learns you as much as you learn it.

Now, as intelligent systems enter the operating room — cameras, sensors, haptics, edge processors — this feedback loop tightens. We are no longer teaching machines to mimic us; we are co-evolving. The model learns from our experience, and we learn to see as the model does — in probability, in signal, in pattern.

This reciprocity forces a deeper question: what remains uniquely human when machines begin to mirror our judgment? Perhaps it is the capacity to care about loss, to feel the gravity of error, to hold meaning where code holds only math. Machines may minimize loss; only humans can assign it value.

The goal is not automation, but alignment — systems that learn with us, not merely from us. The best models won’t replace intuition; they’ll amplify it, surface bias, and expand awareness. The operating room and the neural network are, in the end, chapters of the same book — both refining themselves through feedback, error, and care.

The Learning Body

Bias and loss are not defects of intelligence — they are its conditions.
Every learning system, biological or artificial, must confront uncertainty.
Progress is not the absence of error, but the ability to adapt to it.

What surgery taught me about machine learning — and what machine learning taught me about surgery — is that intelligence is not a possession; it’s a process.
It lives in motion, in feedback, in the uncertain space between intent and reality.
The scalpel, the sensor, the neuron, and the algorithm are all extensions of the same truth: learning is embodied.

To operate — on tissue or on data — is to allow oneself to open a dialogue with uncertainty. Every action alters the system that teaches you. Every iteration redraws the line between control and discovery.

After thirty years at the table, and years more studying Artificial Intelligence, I’ve come to see machine learning as one of humanity’s most powerful tools — if guided wisely.
My fear is not that AI will replace us, but that we’ll misuse it — chasing superficial revenue efficiency while overlooking the deeper human signals that matter most: trust, understanding, and growth.

If we use data only to accelerate throughput, we’ll miss its higher purpose — to study ourselves.
To illuminate bias.
To preserve tacit wisdom.
To augment the human hand rather than automate it.

The promise of AI in medicine is not perfection. It is deeper understanding — a fuller model of what it means to heal.

And that is the hope of Embodied Surgical Intelligence:
that our tools might finally learn to feel,
that our knowledge might no longer die in silence,
and that our systems might evolve ethically — with us, through us, because of us.

Because in the end, the most profound act of intelligence — in life, in code, or in surgery — is not to predict the future perfectly,
but to be changed by what you encounter.

By Michael M. Karch, MD, FAAOS. Mammoth Orthopedic Institute | Harvard Business Analytics Program | MIT Executive Program in Machine Learning