Missing evidence is not negative evidence.
An empty provenance field can mean not observed, not collected, unavailable, withheld or not applicable. Accountable AI systems should preserve that difference before making a decision.
Dr. Benjamin Harvey
Watch the 48-second missing-evidence explanation
Silence can look more certain than it is
AI systems often receive incomplete records. A source document may be unavailable. A retrieval step may fail. An audit log may stop. A model may not know whether a field was never collected, deliberately withheld, judged irrelevant or lost during transfer. Yet the downstream interface may render every condition as the same empty value.
That collapse creates a governance problem. If missing evidence is treated as negative evidence, an attribution system may conclude that a contribution did not occur. A policy engine may interpret an absent approval record as a denial—or, more dangerously, as no restriction. An incident reviewer may assume that an action never happened because the logging mechanism failed to capture it.
The first control is conceptual: absence in the record is not proof of absence in the world.
Standards distinguish records from recording failures
NIST SP 800-53 Revision 5 treats audit content and audit failure as separate control problems. AU-3 defines the content needed to establish what event occurred, when and where it occurred, its source, outcome and associated identity. AU-5 separately requires organizations to alert defined roles when the audit-logging process fails and to define additional response actions.
That distinction matters for AI governance. A negative result is content: the test ran and did not find the condition. Missing evidence is state information about the evidence process itself.
The NIST AI RMF Playbook likewise recommends documentation inventories whose completeness, usability and efficacy are reviewed, and asks organizations to document data provenance, including sources, origins, transformations, dependencies, constraints and metadata. These official materials support disciplined records. They do not define the missingness vocabulary proposed here.
Provenance should carry the gap
The W3C PROV Data Model represents entities, activities, agents and relationships such as generation, use, derivation and responsibility. It gives a system a way to describe how a record came to exist. My synthesis is that the provenance layer should also make decision-relevant gaps explicit instead of leaving reviewers to infer meaning from silence.
A practical evidence state could distinguish: present; expected but missing; not collected; collection failed; unavailable at review time; withheld under an identified rule; not applicable; and superseded by a later record. Each state should carry a source, observation time, responsible process, retry or escalation status, and the decision rule that applies while the evidence is unresolved.
This is especially important in AI IP attribution. A missing source link does not establish independent creation. A failed similarity search does not establish no relationship. An unavailable license record does not establish either permission or prohibition. Those conclusions require their own evidence and legal review.
Test decisions under controlled missingness
I would evaluate this with synthetic or consented workflows in which the underlying events are known. Reviewers would receive the same cases under three conditions: empty fields without explanation; explicit missingness states; and explicit states plus bounded response rules such as retry, escalate, deny, defer or permit under a named exception.
Measure false negative conclusions, false permission conclusions, unnecessary blocks, reviewer agreement, reconstruction time, escalation quality, legitimate task completion and the amount of sensitive information retained. Include logging failure, delayed evidence, conflicting sources, withheld records and genuinely inapplicable fields. Hold the cases, tools, review time and decision thresholds constant.
This is a proposed evaluation and original governance synthesis, not an official NIST or W3C schema and not a report of measured results. The question is narrow but consequential: when evidence is missing, can the system preserve uncertainty without turning silence into a fact?