Behavioral scientist Gleb Tsipursky argues that AI-assisted inspection can produce a precise measurement and still reach the wrong conclusion when the result is tied to an obsolete requirement. In an article titled “AI Inspection Needs a Version Passport for Every Measurement,” he describes a part checked against an outdated drawing revision: the instrument records the dimension faithfully, the software generates a polished report, and the wrong conclusion travels quickly.
Connected metrology makes this failure mode more consequential, Tsipursky writes. Quality Information Framework data can preserve the identity of a measured characteristic across product definitions, inspection plans and results. The technical promise is traceability, but the management challenge is ensuring that each decision follows the correct version of the requirement.
His proposed remedy is a “version passport” for every consequential measurement. It need not be a new software product, he says; it can start as a set of fields that travel with the result: the approved drawing or model revision, characteristic identifier, inspection-plan revision, instrument or method, result timestamp, and the identity of the person or system that approved the disposition.
The extra context matters because AI can summarize an inspection report, point to an apparent deviation or recommend cases for review. If it receives results without version information, it may compare an observation with the wrong specification. The arithmetic can be correct while the conclusion is invalid. Tsipursky notes that the National Institute of Standards and Technology’s 2026 smart-manufacturing roadmap identifies integration of heterogeneous data and trustworthy operation as continuing challenges for industrial AI. A version passport addresses a specific part of that problem by making a measurement’s origin and applicable conditions visible to the next decision-maker.
Quality teams can test the idea before integrating a new platform, he suggests. They should select a small set of completed inspection records, including one legitimate out-of-tolerance result, one conforming part and one example with mismatched revisions. An AI assistant should be asked to summarize the evidence, identify the applicable revision and flag missing provenance. Its responses should be compared with an experienced quality engineer’s determination.
The test should not stop at how often the assistant labels a part good or bad. Tsipursky recommends recording whether it catches a revision mismatch, whether it invents an unstated tolerance and whether it can point to the approved source behind a conclusion. Those are the errors that digital handoffs often conceal, he says.
Clear ownership is also needed when records disagree. A quality engineer may determine whether a measurement is valid, design engineering confirms the governing definition, and manufacturing decides whether production can continue. The escalation route should be settled before an alert appears rather than negotiated during a shipment dispute.
Tsipursky distinguishes the proposal from the familiar demand for more AI-versus-human comparisons. Even perfect agreement between human and automated inspectors can reinforce the wrong decision if both were given an obsolete specification, he argues. The reference itself needs scrutiny.
Automation can make measurement data travel farther and faster. A reliable chain of custody, in his view, ensures that the drawing, feature, method and result still belong together when the decision is made. The next generation of connected inspection, he concludes, should make that connection impossible to ignore.