Missing Is Not False: How Can AI Systems Distinguish Unobserved Evidence from Evidence of Absence?
Artificial-intelligence systems increasingly translate unsuccessful searches, empty query responses, absent alerts, and inaccessible records into statements about the state of the world. A negative observation, however, does not by itself entail that the investigated event or object is absent. The evidentiary meaning of non-detection depends on the specified target, the completeness of the observational domain, the observation mechanism, identity resolution, access rights, temporal validity, and the reliability of the evidence route. This paper asks when an AI system may legitimately promote non-observation into a bounded negative claim. It proposes a claim-relative Absence Qualification Contract (AQC), a methodological specification rather than a new logic or computational primitive, and distinguishes positive observation, qualified absence, unresolved uncertainty, and incapacity to observe. These labels are refined into separate dimensions because epistemic support and observability are not always mutually exclusive categories. A possible-worlds analysis establishes a conditional non-entailment result: if two admissible worlds yield the same recorded observation but differ on the target's existence, no procedure using only that observation can soundly certify absence in both worlds. Constructed cases involving enterprise transactions, clinical records, telemetry, access-controlled data, and AI-agent completion illustrate why query success, missing entries, and confident model language do not discharge the burden of a negative claim. The contribution is an explicit protocol for specifying negative-claim warrants, expiration, refusal, and independent adjudication, alongside a falsifier-first evaluation design against established database-completeness, multi-valued logic, probabilistic detection, provenance, and assurance-case baselines. No live-system experiment, improved accuracy, independent replication, or universal superiority is claimed. A well-specified conventional architecture may implement every proposed distinction, in which case a separate AQC abstraction is unnecessary.
Authors
- Rogério Figurelli (ORCID: https://orcid.org/0009-0002-5726-1765)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-12
- DOI
- https://doi.org/10.5281/zenodo.24160973
- Primary Topic
- Logic, Reasoning, and Knowledge
- Type
- preprint