Manifold Grace
PRIMARY GLOBAL GOVERNANCE RESEARCH · 5 NOVEMBER 2026

AI Identifiability First; Compliance Second

A cross-industry, cross-jurisdiction research framework for detecting, identifying, documenting and governing artificial intelligence wherever it is used.

Emerging Research Horizon · Government & Society · Interdisciplinary Atlas

My first two global-governance considerations

Question 1 · Ontological and operational

Is an AI instance identifiable?

Identify whether AI exists in a system, device, process, organisation or interaction; establish its function, accountable provider and deployer, version, deployment scope and potentially affected persons. Distinguish knowing a tool is present from uniquely identifying a particular running instance.

Question 2 · Normative and evidentiary

Is that identified AI use compliant?

Assess the laws, standards, controls and duties actually applicable in the specific jurisdiction, sector, role, risk category and deployment setting; examine evidence of governance, transparency, privacy, safety, human oversight and redress.

Order of inquiry: identify the AI → identify its purpose, operator, context and jurisdiction → determine applicable duties → test compliance evidence → document findings, uncertainty, remediation and audit history.

AI identifiability is a necessary input to many forms of accountability, not proof that a particular operator is acting unlawfully. A failure to locate AI is not evidence that none exists; nor does an identifiable model automatically meet legal obligations.

What counts as identification?

LayerQuestionPossible evidence
Existence / presenceIs AI part of the system or interaction?Disclosure, procurement records, system inspection, audit evidence
System / model identityWhat system, model and version?Registry entry, version information, system/model card, supplier attestation
Instance / deploymentWhich running service or embedded component?Deployment IDs, signed configuration, logs, hardware or software inventories
ResponsibilityWho provides, deploys, operates and oversees it?Accountability register, contractual responsibility, contact and oversight routes
Outputs and decisionsWhere did AI materially contribute?Decision logs, provenance and AI-content marking where relevant and feasible
Compliance stateWhich requirements apply, and were they met?Impact assessments, tests, documentation, monitoring, audit, incident and correction records

These layers prevent confusion between a labelled AI-generated image, a disclosed chatbot, a registered high-risk system, and a particular AI component within an industrial device. They are not interchangeable kinds of evidence.

Physical materials, nanoengineering and software disclosure

Working research hypothesis: Certain physically embodied AI systems may be identifiable through their material components, sensors, chips or engineering signatures, potentially including specialised nanomaterials. This could be an appropriate domain-specific identification method, but there is no evidence here that AI generally has a unique material substance or that nanomaterial detection could universally identify it.

AI is a computational capability, not inherently a particular material. It can run on conventional digital hardware, in cloud services, as an embedded component, through an API, or within specialised physical systems. Material-based identification would therefore cover only some embodiments and would require separate feasibility and accuracy testing.

Software and institutional approach: AI can be identified without a distinctive material signature through inventories, deployment registers, provenance or attestations, supplier duties, technical inspection and required disclosures. A legal duty to disclose is a governance mechanism; it does not guarantee that hidden or unauthorised systems become technically detectable.

Research tests to compare identification mechanisms

Evaluate disclosure completeness, false positives and negatives, resistance to evasion, coverage of embedded components, version traceability, proportionality, privacy, accessibility, implementation cost and independent auditability. Compare physical detection only where there is an actual distinctive component to test. Record unknowns explicitly rather than infer presence from an absence of documentation.

Existing regulatory precedents and their limits

Laws differ by country, provider/deployer role, dates of application, use cases, and exemptions. These references are examples, not a global legal determination. The proposed universal identification-first principle goes beyond current legal coverage.

Five research problems following the two primary considerations

  1. Reliable AI evaluation: How can evaluation systems detect incorrect grading, reference labels or benchmarks?
  2. Evidence-grounded AI: Can versioned source evidence distinguish contradiction, changes over time and incompleteness?
  3. Long-horizon agents: How can multi-stage work remain reliable, correctable and accountable?
  4. AI-assisted scientific discovery: How can independently reproduced discoveries be distinguished from hypotheses and previously known work?
  5. Governable AI architecture: Can auditability, privacy and accountability be designed into systems from the beginning?

These questions can be investigated under both primary priorities: what AI is present and whose responsibility it is; then whether its operation, testing, records and effects satisfy applicable requirements.

Frontier research discipline

  1. Unpublished, independent or emerging discoveries may exist beyond a model's represented or retrievable knowledge.
  2. Unfamiliar research deserves critical examination, not automatic dismissal or acceptance.
  3. Openness to discovery must coexist with scientific evidence, uncertainty reporting, rights protection and safety.

Related: Why AI access is not the limit of human knowledge · Research gaps, sources and solutions · LLM Evaluation and Governance.

Research position published 5 November 2026. Proposed global framework, not a claim that all AI systems are currently detectable or subject to identical legal identification duties.