Article
Why Real-World Safety Cannot Be Crowdsourced to Generic AI
Statutory duty is not a consensus of internet text. It is a specific obligation, held by a named person, testable in court.
11 February 2026 · 6 min read
The argument
Averages Do Not Hold Duty
A general-purpose model reflects the statistical centre of its training data. Statutory compliance is the opposite discipline: it turns on exceptions, on the specific clause, on the one reading that sits outside tolerance.
No amount of general fluency substitutes for knowing that this particular flue arrangement, in this particular block, was altered in 2019 by a contractor whose competency record has since been downgraded.
Operational Memory Is the Differentiator
Thirteen years of decisions, failures, engineering updates, asset intelligence, contractor performance and evidence chains is not a dataset a generic model can approximate. It is the accumulated record of what actually happened in real buildings.
That memory is what allows the platform to spot the anomaly across documents, benchmark a contractor's outputs against their peers, and escalate before the failure rather than after it.
Accountability Cannot Be Delegated to a Model
When something goes wrong, the duty holder is a person. A system that cannot name who decided what, and on what basis, has moved liability onto that person rather than away from them.
Questions this Article Answers
Questions This Article Answers.
Use This Insight
Use This Insight.
If you are reviewing your current systems, ask:
- Does our system retain history, or only the current certificate?
- Can we benchmark contractor output quality against peers?
- Would contradictions between two related documents be detected?
- Is every statutory decision attributable to a named, competent person?