Your Next AI System May Need a Failure Passport.
Companies scrutinise what AI systems promise to do. A harder question is emerging: should their history of consequential failures follow them to the next customer?
A refinery evaluating an AI system today can examine its accuracy, cybersecurity controls, architecture and compliance documentation.
What it may not know is whether essentially the same system failed somewhere else six months earlier.
That information gap could become one of the next major questions in AI assurance.
Gleb Tsipursky recently proposed a portable record of consequential AI failures, corrective actions and unresolved risks that could accompany systems across deployments and jurisdictions.
The principle has precedent.
Aircraft carry maintenance histories. Contractors arrive with safety-performance records. Critical equipment has inspection histories.
None guarantees future performance, but each gives the next decision-maker something valuable: memory.
AI procurement may eventually need the same.
Imagine discovering that a system being considered for a hospital, refinery or transport network previously failed under comparable conditions.
The important question wouldn’t simply be whether it failed.
What caused it?
What authority had been delegated? Did human intervention arrive in time? What changed afterwards?
Was the corrective action independently verified?
Suddenly, failure history becomes decision intelligence.
It also creates a counterintuitive possibility.
A vendor disclosing three well-investigated failures, verified corrective actions and residual risks could be more governable than a competitor reporting none.
Because zero reported failures and zero observed failures are not necessarily the same thing.
Confidentiality, liability, verification, model versioning and common reporting thresholds make implementation difficult.
But the underlying principle is powerful:
AI risk should have memory.
Today, an organisation can pay the price of discovering an AI failure while another organisation later deploys related technology without inheriting that learning.
That is not scalable risk management.
As AI moves between companies, industries and jurisdictions, perhaps its consequential history should move with it.
Not another compliance document.
Portable failure intelligence.