The hidden cost of unverifiable AI: fines, disputes, and lost trust

When teams weigh whether to make their AI systems verifiable — able to produce a signed, tamper-evident record of what each model did — the decision usually gets filed under compliance and deferred. It looks like optional overhead. But the cost of unverifiable AI is not zero; it is deferred and compounding, and it lands at the worst possible moment: when something goes wrong and you need proof you do not have.

The first cost is regulatory. The EU AI Act attaches penalties of up to 7% of global turnover for the most serious violations, and its logging and transparency obligations are phasing in through 2026 and beyond. When a regulator asks a provider to demonstrate what a system did and prove the records were not altered, 'we have internal logs' is a weak answer — the logs belong to the party being investigated. An organization that cannot produce a tamper-evident record is not merely non-compliant on paper; it is exposed to the full penalty because it cannot mount a credible defense. The cost of the missing record is measured against turnover, not against the modest expense of having produced it.

The second cost is disputes. Every consequential AI decision is a potential dispute — a denied claim, a rejected application, a flagged transaction, an automated trade. When the affected party challenges it, the question is always the same: what exactly did the system do, and can you prove the record is genuine? Without a verifiable attestation, these disputes become swearing contests the operator often loses, because the burden of showing the record is trustworthy falls on the party that controls it. With a signed record anyone can verify, the dispute collapses into a check: the operator shows the proof, the challenger verifies it, and the facts are settled. Unverifiable AI turns every dispute into an expensive, open-ended argument; verifiable AI turns it into a lookup.

The third cost is the quietest and largest: trust. Counterparties, customers, auditors, and partners are increasingly unwilling to take 'trust our AI' on faith, and the ones who demand proof will route business to providers who can give it. In markets where AI decisions carry weight — finance, healthcare, public services, autonomous agents transacting with each other — verifiability is becoming a precondition for doing business at all, not a differentiator. The organization that cannot prove its AI's behavior slowly loses access to the counterparties who now require it. Against all three costs, the honest accounting is simple: producing a signed attestation at the moment of each decision is cheap and scales to millions of decisions; reconstructing trust, winning a dispute without evidence, or defending a fine without a record is expensive or impossible. Sign those attestations with a post-quantum scheme (ML-DSA, FIPS 204) so they hold up for the years such records must survive — resistant to known classical and quantum attacks per NIST, not unbreakable. Verifiability is not the cost; unverifiability is.