The Proof Economy

Across my career, from medicine and life sciences into structural biology, higher education and later semiconductor verification, I have kept meeting the same fundamental challenge in different forms.

How do we know something is true?

Scientists speak of reproducibility. Physicians of evidence. Educators of understanding. Engineers of verification. Beneath the terminology lies a common objective: establishing confidence before action.

Why this question matters now

As a founder, I use AI daily. It drafts, summarises, explains, and occasionally surprises me with the elegance of what it produces. It has changed the pace at which I think and act as an executive.

Yet the more I use it, the more I notice a quiet friction: The output looks finished. It reads confidently. But before every decision that matters, I still have to ask: do I trust this enough to act on it?

That question, once asked mainly in the laboratory or the verification lab, is now asked in the boardroom, the classroom and the consulting room. The urgency is not that AI has become dangerous, but that it has become good enough to be believed without being verified.

What follows is not a survey. It is my own reading of a shift I have watched from four vantage points: the wet lab, the university, the boardroom and the verification lab. The disciplines look nothing alike from the outside. From the inside, they have been converging on the same question for some time. AI has only made it louder.

Science: Prediction is not discovery

The tension between prediction and proof is perhaps most visible in science.

Proteins form the molecular machines that run every cell, and their three-dimensional shape determines what they do: how a pathogen injects DNA into a host, how an antibiotic binds its target, how an enzyme catalyses a reaction. For half a century, discovering that shape was led by laborious experimental work. AlphaFold, released by DeepMind in 2020, changed that. Given only a protein sequence, it predicts the folded structure in minutes, at accuracies that in many cases rival experiment. The leap is real. It is also easy to misread.

My own training was as an experimental structural biologist, first using X-ray crystallography and later single-particle cryo-electron microscopy to solve the architectures of large bacterial machines: the type IV secretion system that Gram-negative pathogens use to inject DNA and toxins, and the RNA polymerase complexes that read genes into RNA. Neither project started with the structure. They started with a biological question: how does this machine open, gate, translocate, switch state?

The structure was a means to that question, not an end in itself.

Everything downstream depended on how honestly the atomic model reflected the experimental evidence, the diffraction data in a crystallographic map or the density in a cryo-EM reconstruction. A side chain placed one rotamer too confidently, a loop built where the density did not support it, and the mechanistic story built on top would quietly drift away from reality. The discipline was as much about restraint as ambition: build only what the data can carry.

AlphaFold accelerated the early stages of that work dramatically. Starting models that once took months of molecular replacement or manual building could be generated in hours. Yet the underlying discipline did not change.

An independent assessment in Nature Methods concluded that AlphaFold’s predictions are best treated as “exceptionally useful hypotheses”, highly accurate on average, yet with roughly 10% of the most confident predictions still deviating from experimental structures by more than 2 Å (Terwilliger et al., 2024).

A prediction, however elegant, still had to be reconciled with the data.

That principle that a model must correspond to evidence rather than merely appear plausible is the same one that governs verification in engineering. The language differs. The discipline does not.

Medicine: Confidence is a clinical requirement

I came to structural biology through medicine. Twenty-five years ago, I sat on the other side of this question, as a medical student at King Edward Memorial Hospital in Mumbai, learning how a diagnosis is defended. The tools were textbooks, cadavers, bedside teaching and long viva examinations in front of consultants who did not settle for the right answer if you could not defend the reasoning behind it.

Diagnostic confidence was not an intuition. It was a chain of evidence: a differential considered and ruled out, a treatment justified against alternatives.

That discipline scaled poorly, bounded by the memory of individual clinicians and the time they had for each patient. AI is now genuinely helping where it was hardest to sustain, matching or exceeding specialists on narrow, well-scoped tasks as I hear from my friends. Where it has landed responsibly, it has extended access rather than replaced judgement.

Yet, clinical adoption still depends on prospective validation and accountability, not algorithmic performance alone. As a person on the other side of the table now rather than a clinician, I notice the shift in myself. I welcome the speed of an AI-assisted read. I still want to know a human took responsibility for the decision that acts on it.

Patients do not seek the most probable answer, but the most reliable.

Education: An improvement challenge, not a detection problem

For generations, producing an answer was accepted as evidence of learning. Generative AI has made that assumption fragile. Essays, code and sophisticated analyses can now be produced in seconds. Chasing this with ever-better detectors misreads the problem. Detection is a policing frame. What has actually changed is the assessment target: understanding can no longer be inferred from a finished artefact alone. Education has been handed an improvement challenge, not a plagiarism one.

I spent more than nine years as an educator at Queen Mary University of London, teaching biochemistry and structural biology to undergraduates and postgraduates, and became a Fellow of the Higher Education Academy along the way. Long before generative AI, the honest question in a classroom was always the same one: has this student understood, or have they reproduced? A well-written essay could hide either. So could a correctly worked problem. The teachers I learned most from asked follow-up questions. They watched students draw a mechanism on the board. They ran lab reports as conversations, not as marking exercises.

Assessment, done well, has always been a dialogue.

What generative AI has done is remove the last comfortable hiding places for that lazier version of assessment, the one that graded the artefact and inferred the mind behind it.

Universities that respond by tightening detection are trying to preserve a shortcut. Universities that respond by rebuilding around dialogue, oral examinations, staged submissions, in-class problem solving and process-graded work, are finally being pushed to do what good teachers already knew they should. The tools have caught up with the standard, not undermined it.

The classroom I remember at Queen Mary and the training rooms that Axiomise now runs for practicing engineers are asking the same question in different vocabulary. Can the learner walk through the reasoning, defend the choice, explain why one approach holds and another does not, and reproduce the argument on something they have not seen before? AI can accelerate every part of that preparation. It cannot answer for the learner.

The challenge is no longer proof of completion. It is proof of comprehension.

Semiconductors and autonomous systems: The cost of uncertainty

Testing can show that a system works under specific conditions; it cannot show that failures are absent under all conditions. The semiconductor industry learned this the hard way. As designs grew to billions of transistors and moved into safety-critical systems from pacemakers to avionics, simulation alone could not deliver assurance. Formal verification emerged as the mathematical discipline that reasons about all possible behaviours of a design, not only those a testbench happens to exercise.

Autonomous vehicles are confronting the same lesson publicly. In July 2026, the U.S. NHTSA issued a call to action after documenting cases of autonomous vehicles driving into active emergency scenes, noting that “emergency scenes are not rare or extreme edge cases” (TheStreet, 2026). Regulators are moving in step: the EU AI Act’s conformity assessment regime for high-risk AI systems begins applying in August 2026 (EU AI Act Article 43).

The chip inside the car and the software driving it face the question formal verification has answered for decades: can we show absence of failure, not merely absence of observed failure?

The emergence of the proof economy

Every technological revolution creates a new bottleneck. The Industrial Revolution reduced the scarcity of physical production; the Information Revolution, the scarcity of information; the AI Revolution is reducing the scarcity of creation itself. As creation becomes abundant, validation becomes scarce.

Whether validating a protein structure, a medical diagnosis, a student’s understanding or the behaviour of an autonomous system, the question is the same:

Not whether something appears to work, but whether we can demonstrate why it should be trusted.

A reflection from a conference floor

This year I attended various conferences from Silicon Valley to India, and I walked the floor, spoke to many, and heard the chatter in the halls. In the exhibition hall around us, half the conversations were about what AI can now generate. On the panel in front of me, the conversation was almost entirely about what it should have to prove before it is trusted in a car, a train, in healthcare, the sensors around us. Different rooms, different disciplines, the same underlying question.

Walking out of the venues afterwards, I thought about how far the road to Axiomise for me had run. From the corridors of King Edward Memorial Hospital in Mumbai, to a Cambridge crystallography lab, to a Queen Mary lecture theatre, to the audience of industry events where the question on the table was the one that had followed me from the wards. How do we know?

The age of productivity is already here.

The question is whether we can build the discipline, the training and the institutions to make the age of proof match it.

The post The Proof Economy appeared first on Semiconductor Engineering.