Definition
The cost of establishing whether a machine-generated claim is true, paid by the person the claim was generated for. Generation takes seconds. Verification of plausible prose against its sources takes minutes of high-load scanning — and the sources are frequently the scans, faxes, and unindexed records the system summarized in the first place. A fluent draft that cannot show where it got each claim does not save time; it moves the time onto the reader and hides it inside a metric that only counts generation.
It is the burden-side twin of compensated performance. That term covers labor spent making a system's errors invisible. This one covers labor spent making its outputs usable — and a system can be well built, fast, and accurate, and still tax its readers, because the tax is levied per claim and no metric counts it.
It differs from verification harm, which is the burden of proving eligibility to an institution. This is the burden of trusting a machine's summary of a record the reader cannot cheaply check. And it differs from audit burden ceiling, which caps the labor a respondent spends producing evidence. Here the reader did not choose to verify; the output was presented as if the work were done.
The checkable form is claim-level anchoring: every displayed claim links to its source on demand, and a claim the system cannot anchor is suppressed rather than shown. The explainability evals test it.
Hypothetical example
- An AI summary states a documented penicillin allergy. Confirming it against twelve unindexed outside-hospital scans takes eight minutes; the generation took four seconds and is what the metric counted.