1. Embedded inference
Why it can be missed: AI is a small feature inside a larger product, with no separate user-facing label.
Identification and remediation
Inspect product architecture, supplier declarations, model dependencies and release notes.
How can governments and organisations regulate, monitor or audit AI if they cannot first establish which AI systems and uses exist? A system-of-systems methodology for locating, defining, recording and governing AI across institutions and industries.
← Primary identifiability / compliance principles · Interdisciplinary Atlas · Government & Society
Primary consideration 1 — identification: Determine whether an AI system or use is present and establish its operational identity, purposes, deployment locations, relationships, accountable parties and affected groups. Primary consideration 2 — compliance: identify the applicable duties, then assess evidence of conformity, monitoring, oversight, incident response and redress.
Foundational distinction: an accurate asset inventory is the prerequisite for comprehensive organisational AI governance, but a complete inventory is difficult to prove: unknown applications, vendor-controlled components, short-lived deployments and undocumented use can remain outside the observed population. A public disclosure register is not the same as a complete internal inventory, and neither establishes that every use is compliant.
NIST AI RMF explicitly calls for AI inventories (GOVERN 1.6). The UK ATRS covers specified public-sector algorithmic uses, with scope rules and exemptions. In a written parliamentary answer on 15 July 2026, government stated that DSIT did not maintain a public or internal inventory of automated decision-making systems while explaining ATRS publication. This is evidence of a governance distinction, not a claim that every public-sector organisation lacks records.
Use precise identity layers rather than treating AI as one undifferentiated object. A model may serve multiple products, a product may contain multiple models, and the same deployment may change versions or operators over time.
| Entity | Definition for the proposed register | Identity / evidence examples |
|---|---|---|
| Capability | Type of prediction, generation, ranking, optimisation or agent operation | Capability taxonomy, purpose classification |
| Model | Particular learned artefact or model family | Provider, model ID, version, integrity digest where appropriate |
| AI system | Application or service that uses one or more models and associated components | System ID, purpose, architecture, owner |
| Deployment / runtime instance | Operating implementation in a specified environment and period | Deployment ID, effective dates, region, configuration |
| Use case / workflow | Organisational activity in which the system is used | Process map, function, decision influence and affected people |
| Supplier / operator | Entities developing, providing, deploying or managing the system | Contracts, responsibility records, service providers |
| Data / output / decision | Information processed, produced or used in consequential action | Data flows, decision records, output provenance |
| Requirement / control | Applicable legal, policy or safety obligation and its implementation | Jurisdiction, law, standard, control and evidence |
| Assessment / event | Validation, audit, incident, update, reassignment or retirement | Timestamped evidence, reviewer, corrective action |
The register must express many-to-many relationships: system ↔ models ↔ deployments ↔ uses ↔ suppliers ↔ people affected ↔ controls ↔ evidence. Version and time are required to avoid confusion between obsolete and current systems.
Each scenario is a potential discovery problem; it is not evidence of wrongdoing or undisclosed AI in any particular named organisation. Expand a card for an appropriate control.
Why it can be missed: AI is a small feature inside a larger product, with no separate user-facing label.
Inspect product architecture, supplier declarations, model dependencies and release notes.
Why it can be missed: A familiar software function silently gains predictive or generative capability.
Version-by-version inventory review and vendor change notifications.
Why it can be missed: AI runs only for a subset of users, sessions, territories or trial groups.
Register experiments, deployment toggles, owner and exposure criteria.
Why it can be missed: A service sends data to a third-party model without visible AI branding.
Map API calls, subprocessors, contracts and service dependencies.
Why it can be missed: An AI library arrives as a bundled component within another application.
Software component inventory, package audit and code review.
Why it can be missed: The supplier changes a hosted model or its underlying provider without local release.
Contractual change-control, attestations and version verification.
Why it can be missed: A vendor depends on another platform, model provider, retrieval service or moderation model.
Tiered supplier mapping and responsibility matrices.
Why it can be missed: A historic application includes ML that current staff did not build or document.
Legacy asset discovery, historical procurement, architecture review.
Why it can be missed: Staff adopt unsanctioned AI applications outside approved procurement.
Safe approved-tool route, training, proportionate asset discovery and regular declarations.
Why it can be missed: Staff access public AI through individual subscriptions rather than enterprise accounts.
Policy, training, procurement checks and privacy-respecting risk review.
Why it can be missed: AI functions hide in macros, connectors, automation platforms or extensions.
Workflow inventories, connector reviews and app owner attestations.
Why it can be missed: People see a human interface while AI prioritises cases or applications behind it.
Trace end-to-end workflows, decisions and material model contributions.
Why it can be missed: An operator follows an AI score without acknowledging the model's role.
Decision mapping and evidence of meaningful human review.
Why it can be missed: Background models select, suppress, route or transform content.
Component-level documentation, impact mapping, appropriate disclosure.
Why it can be missed: Rankings or feeds shape exposure without the AI role being obvious.
Disclose material ranking factors, document model use and complaint routes.
Why it can be missed: A model estimates price, risk, access or eligibility in ways the recipient cannot see.
Decision impact assessments, model inventories and review rights.
Why it can be missed: A person encounters the decision of agency B based on AI outputs supplied by agency A.
Cross-organisation flow records, interfaces and responsibilities.
Why it can be missed: A visible assistant launches hidden subagents, API calls or additional models.
Agent identities, scoped permissions, parent-child task logs.
Why it can be missed: AI runs in unattended batches or in response to events without a direct user interaction.
Scheduler, service-account and workflow execution inventories.
Why it can be missed: AI spins up briefly and leaves little visible persistent infrastructure.
Deployment manifests, signed execution records and runtime telemetry.
Why it can be missed: One public service actually uses several specialist models that disagree or vote.
Component graph, routing and model-level version records.
Why it can be missed: Inference occurs locally, without contacting a central service.
Device/vendor documentation, signed component manifests and local audit options.
Why it can be missed: ML modules operate in robotics, monitoring, production or sensor systems.
Operational technology inventories, safety cases and asset inspections.
Why it can be missed: AI support is integrated into diagnostic, imaging or research equipment.
Manufacturer declarations, change logs and clinical/lab governance review.
Why it can be missed: An ordinary consumer product receives AI through firmware or remote updates.
Product lifecycle register, notice of material changes, manufacturer attestations.
Why it can be missed: Local embedded classifiers are distributed across many devices or facilities.
Device identities, fleet inventories and supplier supply-chain maps.
Why it can be missed: Content combines AI and human edits and lacks credible provenance.
Origin records and context-sensitive provenance standards, not watermark alone.
Why it can be missed: An outsourced service returns only recommendations; the ordering organisation cannot inspect internals.
Contractual audit rights, purpose, output and decision traceability.
Why it can be missed: Jurisdiction, service operator and applicable obligations differ across deployment layers.
Record providers, deployment regions, data flows, user locations and legal roles.
Why it can be missed: Small experimental AI deployments are omitted from production-only registers.
Include pilots, tests and planned deployments with status and exposure.
Why it can be missed: An AI model is declared retired but still appears in a forgotten integration.
Decommission verification, live-traffic monitoring and historical records.
Why it can be missed: Complex statistics, expert systems and advanced automation are misclassified as AI or vice versa.
Use documented definitions, classification rationale and human technical review.
Declarative methods: supplier attestations, procurement, contracts, employee declarations, app inventories. These are explainable but can omit hidden subcomponents and unreported use. Technical methods: authorised code/dependency inspection, cloud/service inventories, observability, endpoint agents, runtime records and security monitoring. These depend on permissions and coverage; network observations alone cannot reliably establish AI use. Human methods: staff interviews, practitioner workshops, user complaints and qualitative process mapping can identify otherwise undocumented uses. Reconciliation: require evidence from multiple independent sources, log uncertainty and review unresolved candidates.
Discovery must be lawful, proportionate, privacy-preserving and authorised. Do not equate every statistical tool with AI or use speculative signatures as proof. Avoid indiscriminate employee surveillance. Publish appropriately limited summaries, not credentials, exploit-relevant secrets, confidential model internals or personal data. Some sensitive systems may be documented in restricted oversight channels rather than open public registers.
| Measure | What it answers | Important qualification |
|---|---|---|
| Observed inventory coverage | Registered confirmed systems ÷ confirmed systems observed by a defined independent audit method | This is audit-sample coverage, not proof of total real-world coverage. |
| Discovery precision | Confirmed AI candidates ÷ reviewed candidate detections | Depends on the operational AI definition and reviewer quality. |
| Discovery recall | Known AI instances detected ÷ independently established known AI instances in a test population | Unknown unseen instances make global recall unknowable without assumptions. |
| Owner completeness | Identified deployments with verified accountable owner | Assigning a name without confirmed authority is not enough. |
| Dependency completeness | Verified supplier/model relationships relative to test sample | Vendor non-disclosure and nested outsourcing limit visibility. |
| Regulatory classification coverage | Identified uses with recorded jurisdiction, risk and applicable duties | Classification needs legal review and may change with law. |
| Update latency | Delay from system change to verified inventory update | Consider short-lived or frequently changing instances. |
| Public comprehensibility | Proportion of affected users correctly understanding a notice | Use accessible, user-tested language; not disclosure alone. |
| Compliance evidence coverage | In-scope controls supported by current acceptable evidence | Evidence existence is not automatically evidence of compliance. |
Fundamental research gap: no inventory can prove completeness simply by counting what it already contains. Independent discovery, denominator assumptions, sample-based audits, adversarial testing and uncertainty bounds are essential. A detection tool can miss the very AI systems being sought; multiple methods reduce, but do not eliminate, blind spots.
Faceted navigation, findability, content modelling, cross-references and public-facing maps.
Explicit entity definitions and machine-readable relationships: model, system, instance, use, actor, evidence, control.
Connecting multiple inventories without conflating providers, models, deployments and organisational uses.
Business capability mapping, process flow diagrams, component and data dependency graphs.
Asset discovery, ownership, lifecycle, updates, reconciliation, unknown and retired assets.
Dependency composition, vendor chains, libraries, hosted services and change-of-control.
Authorised verification, integrity, telemetry, incident evidence and nonintrusive discovery.
Purposes, lawful bases where relevant, retention, data flows, DPIAs, accountability.
Measurement of inventory completeness, detection accuracy, recall, evidence quality and false assurance.
Applicable duties, jurisdictions, risk levels, regulatory reporting and control testing.
State responsibilities, contracting, agency boundaries, public reporting and redress.
Recognisable notices, user understanding, informed choice and usable challenge routes.
Shadow adoption, internal silos, procurement incentives, reporting culture and organisational change.
Who can discover system use, who bears effects, transparency asymmetries, procedural justice.
Hazard identification, failure modes, safety cases, validation and life-cycle traceability.
Categories, social construction of expertise, partial knowledge, situated practice and contested boundaries.
Estimating coverage, denominator uncertainty, detection bias and unseen populations.
Cross-border obligations, sector regulators, interoperability, accountability boundaries.
Use the same identity architecture across national and local government, education, employment, finance, healthcare, justice, migration, social care, transport, media, infrastructure, manufacturing, research and consumer services, while recording sector-specific legal duties. The same product may be low-risk in one context and materially consequential in another.
| Governance layer | Lead function | Key question |
|---|---|---|
| Product developer / supplier | Model identity, dependencies and documented capabilities | What precisely is being sold, and what changes? |
| Organisation / deployer | Complete local use register, purpose, roles and technical oversight | Where is AI actually used in this organisation? |
| Sector regulator | Risk classification, compliance evidence and sector audits | Which uses are within statutory remit and obligations? |
| Cross-government coordination | Common vocabulary, registry interoperability and inter-agency links | What is overlooked between departments and contractors? |
| International governance | Definition alignment, cross-border responsibility and safe reporting | Can compatible registries map systems that span borders? |
| Public / affected individuals | Appropriate awareness, explanations, challenge and redress | Can a person tell when AI materially affects them? |
Each trial should preregister its scope, baseline inventory, independent discovery method, reviewer criteria, error rates, risk controls and correction record. No global identifiability solution is claimed to have been validated by this publication.
These sources establish existing work and regulatory precedents, not universal mandatory registration. They offer starting points for expert collaboration and critical comparison.
16 direct sources · Published 5 November 2026 · Curated research agenda with proposed control architecture, not a regulatory opinion, systematic review or proof of completeness.
The earlier two-question governance framework remains intact. This page adds overlooked deployment scenarios, information-architecture ontology, inventory methods and test criteria. Changes should be date-stamped and linked, not silently erase prior proposed hypotheses or externally authored contributions.
Related: Research Atlas · Research Method · Research Gaps · Cross-sector verification.