Code doesn’t lie. But government spreadsheets do.
Last week, the UK National Audit Office (NAO) publicly demanded that His Majesty’s Treasury verify its claim that AI will save the public purse £45 billion annually. Independent analysts already peg the real number at half that. This isn’t just a fiscal spat. It’s a textbook case of the disconnect between narrative and data — a gap blockchain protocols have been wrestling with for years.
Context: The Claim and Its Counterpart
The UK government’s AI savings figure is a headline number, not a cryptographic proof. It rests on assumptions about automation of administrative tasks, reduction in headcount, and improved service delivery. No on-chain attestation, no merkle tree of verified outcomes. Just a press release and a budget line. The NAO’s intervention signals that blind trust in these projections is no longer acceptable.
In crypto, we see the same pattern daily: a Layer-2 promises 100,000 TPS; a DeFi protocol claims “unstoppable” security; an AI-crypto project boasts “verifiable inference.” Yet when you crack open the code, the proofs are incomplete, the sequencer is centralized, or the oracle is a single node. The UK case is a mirror: a massive claim with no audit trail.
Core: Why ZK-Proofs Are the Only Way to Close the Gap
From my years auditing ZK-rollups, I know that verifiability isn’t optional — it’s the whole point. The NAO’s demand is a primitive version of what zero-knowledge proofs (ZKPs) have already solved in the blockchain world. Consider: a government department deploying an AI system to process benefit claims. It claims a 30% cost reduction. How do you verify that? Traditional audit digs into logs, interviews staff, and samples cases. That’s slow, expensive, and open to interpretation.
Now imagine a ZK-based audit pipeline: each AI decision is accompanied by a succinct proof that the model followed the prescribed rules, that the training data wasn’t tampered with, and that the output matches the claimed efficiency metric. The verification circuit is published on-chain. Any third party can re-run the proof. The government’s £45 billion claim becomes a set of publicly verifiable statements, not a PowerPoint slide.
This is exactly what we do in blockchain infrastructure. I’ve spent months optimizing Celestia’s blob-sidecar parameters to ensure data availability proofs are both correct and performant. The same engineering mindset applies here: you don’t trust the claim; you trust the proof. The UK government’s mistake is treating AI savings as an accounting entry rather than a cryptographic assertion.
But the challenge is scaling. Verifying every government AI output in zero-knowledge would require massive prover hardware and standardized circuits for each use case. The rollup ecosystem taught us that trade-offs are real: proving time, gas costs, and developer overhead. A one-size-fits-all ZK-audit would be impractical. However, a selective verification layer — sampling critical decisions, generating proofs for high-value transactions — could reduce audit costs by orders of magnitude while maintaining truthfulness.
Contrarian: The Real Blind Spot Isn’t Tech — It’s Incentives
The crypto community will instinctively say, “Just put it on-chain.” But the UK case reveals a deeper problem: governments don’t want verifiability. If the NAO confirms the £45 billion figure is inflated, the Treasury loses credibility, and budget plans unravel. There’s a perverse incentive to keep the numbers fuzzy. The same applies to many blockchain projects: founders claim high TVL or low fees, but when you try to verify the sequencer’s revenue model, the data is siloed.
During the bear market of 2022, I reverse-engineered the exploit of a lending platform that had audited contracts but no on-chain verification of its liquidation logic. The code was correct in isolation; the assumptions about price feed latency were wrong. That’s the same gap here: the UK’s £45 billion saving assumes AI models perform perfectly in production, that user adoption is instant, and that no new costs arise from managing the AI itself. None of these assumptions are verified.
The contrarian take is that ZK-proofs alone won’t fix this. What’s needed is a cultural shift toward mandatory, verifiable disclosures. In crypto, we call it “trust but verify.” In government, it’s called a NAO audit. The difference is that on-chain verification is continuous; government audit is periodic. Until the UK mandates real-time, cryptographically-backed reporting for its AI deployments, the £45 billion claim will remain a political number, not a technical one.
Takeaway: From Fiscal Fiction to Cryptographic Fact
The NAO’s demand is a warning shot for every blockchain project making unverified claims. Your 100,000 TPS is the government’s £45 billion savings. Both sound great in a keynote. Both fall apart under scrutiny. The only way to build lasting trust is to embed verifiability at the protocol level — not as an afterthought, but as the primary design principle.
Will the UK be the first government to adopt ZK-audits for its AI spend? Or will it double down on narrative? The answer will tell us whether the public sector is ready for the same transparency standards we hold blockchain to. Code doesn’t lie. But will the Treasury let the code speak?