Mapping what counts as proof, and the conditions, competencies, and capabilities that shape it.
-
Every serious domain has already answered this, and answered it differently. Legal systems have rules of admissibility and weight. Clinical and social research have reporting standards — CONSORT, STROBE, PRISMA and their relatives. Intelligence practice grades sources and information separately. Audit works in assurance levels. Regulators work in conformity assessment.
Machine-generated material does not fit cleanly into any of them. It arrives without a chain of custody, frequently without a stable account of its inputs, and with a confidence signal that is either absent or not comparable to anything the receiving domain recognises. The practical consequence is that the same output is treated as authoritative in one function and inadmissible in another, within the same organisation.
Related programmes: Conditional Ontologies, Governance Frameworks.
-
Mature, well-documented evidential standards already exist within domains, and are better developed than most AI governance frameworks.
Relevance is conditional. The same item of evidence carries different weight depending on domain, purpose, source and the competence of the operator handling it.
A standard that cannot be applied consistently by different people is not functioning as a standard.
-
Whether machine-generated material can meet existing evidential standards, or requires standards of its own.
Whether operator competence is properly a condition of admissibility or a separate question of professional responsibility.
Whether machine capability should be assessed per system, per task, or per deployment context.
-
How much of an evidential standard survives translation between domains.
How to express a standard so that conformance is checked rather than argued.
What happens to a standard, and to work already assessed under it, when the underlying machine capability changes.