AI can't prove anything that it's doing
AI and decentralised tech are opposites. That’s why they belong together.
5 min read. In this piece:
Why "how do we prove what the AI did?" has no technical answer
The vending-machine-and-doctor split at the heart of it
What regulators will ask for before most orgs are ready
I’ve been building data infrastructure for governments for the past few years, and there’s a conversation I keep having that I can’t shake. It usually starts when someone senior in a ministry or a large enterprise leans forward and says something like: “We want to use AI for this. But how do we prove what the AI did?”
That question sounds simple. It isn’t.
It took me a long time to understand why. I kept thinking the answer was technical, that we needed better systems, cleaner data, smarter models. But after years of sitting across from compliance officers and procurement teams, I realised the problem isn’t technical at all. It’s architectural. AI and proof operate on completely different principles, and most organisations are trying to solve both with the same infrastructure.
They can’t.
The vending machine and the doctor
There are two different ways a system can work, and the easiest way I’ve found to explain the difference is with two objects you already understand.
A vending machine is deterministic. You press B4, you get the chips. If you press B4 again tomorrow, you get the chips again. The machine doesn’t have a bad day. It doesn’t interpret. It executes - same input, same output, every time, without exception.
A doctor is probabilistic. Same symptoms, different patient, different history. The doctor weighs the evidence and gives you the most likely answer: “I’m fairly confident this is X, but let’s run some tests.” That’s not a failure of medicine. That’s how sophisticated reasoning works. You don’t compute certainty. You estimate, update, and decide.
AI is a doctor. It was built this way on purpose. When a language model responds to your question, it generates the most probable continuation of your input based on patterns learned from billions of examples.
When a fraud detection system flags a transaction, it doesn’t execute a fixed rule. It says: “Based on everything I’ve seen, there’s an 89% chance this is suspicious.” When an image classifier identifies a tumour, it gives you a confidence score, not a verdict.
This is powerful, but it creates a problem that most people building with AI haven’t fully confronted.
If the system generating your output is probabilistic, how do you know what happened? Imagine an AI system approves a mortgage application, or flags a citizen as a fraud risk, or classifies a shipment as compliant with sustainability standards.
The decision was made. But was it made correctly? By whom? On what input? Can you prove it to a regulator three years from now? AI makes decisions it can’t prove. Proving them requires a different kind of system.
It requires a vending machine.
What does ‘proof’ mean when the stakes are real?
Proof means something very specific in these environments.
It doesn’t mean a PDF report. It doesn’t mean a database entry that someone could edit. It means a record that exists independently of any single party, cannot be revised after the fact, and that every participant in the network can verify on their own terms.
A centralised audit trail is only as trustworthy as the organisation that maintains it. When five agencies need to agree on what happened, a record controlled by one of them isn’t proof. It’s a claim.
That’s what deterministic infrastructure gives you. When you record a transaction on a decentralised network, the outcome is fixed. The same input produces the same output, for every participant, forever. The rule executes, and the record exists. There’s no interpretation, no probability, no “we think this happened.” It happened, and the proof is timestamped and permanent.
The Dutch Tax Office understood this early. In their pilot for real-time VAT collection, when a purchase is made at a point of sale, the system applies the correct tax rules, splits the payment, and transfers the VAT amount to the appropriate authority within milliseconds.
Every step is recorded on a decentralised network that every party involved can independently verify. mintBlue ran a live demo at the Dutch Ministry of Finance’s innovation event earlier this year and won the innovation award. That’s a vending machine: deterministic rules, deterministic records, no single party in control.
Now imagine adding a doctor to that infrastructure. A fraud detection system spots a pattern across government databases that no human analyst would catch. It flags an entity. If that flag is logged in the same kind of shared, verifiable record, every authority involved sees what triggered it, when, and on what basis. The chain of evidence is intact and auditable. The deterministic layer turns the AI’s output into something that can survive a legal challenge.
The same logic applies beyond government. An insurer using AI to assess claims, or a supply chain classifying shipments for sustainability compliance. Wherever a probabilistic decision needs to survive an audit, you need the vending machine underneath the doctor.
The layer that’s still missing
We’re entering a world where probabilistic AI systems make more decisions. The efficiency gains are too large for that to change. But accountability keeps growing with them. The EU AI Act already requires traceability of decisions. Regulators are moving faster than most organisations realise.
The vending machine and the doctor each cover ground that the other can’t. AI without a deterministic anchor can’t pass an audit, and deterministic systems without AI can’t adapt to new patterns. Together, they cover different zones of the decision-making process, and the organisations building this accountability layer now will be the ones ready when a regulator asks them to prove what their AI decided.
That layer is still being built. I think the question worth asking is what infrastructure we need underneath AI so that its decisions can be trusted.
Where in your organisation is a probabilistic system making decisions without a deterministic audit trail?