Spokes.wiki Search About

Synthesis — AI Governance, Policy & Regulation

The evolving thesis. Spun out 2026-06-07 from the hub ai-governance cluster (3 founding sources).

Scope: two axes of "governing AI"

AI governance is the discipline of setting and enforcing the rules AI systems operate under — who is accountable, what risks are controlled, and how trust is demonstrated. The founding sources stake out two complementary axes:

  1. Organizational governance (the firm). nemko-digital's nemko-ai-governance-services is the inside-the-organization view: AI management systems, inventories, lifecycle & risk management, and assurance/certification ("AI Trust Mark"), operationalized against standards — iso-iec-42001 (AI management systems), the nist-ai-rmf, and the eu-ai-act. Governance here is a compliance & trust function.
  2. National governance (the state). Israel supplies both halves of the state view: israel-ai-regulation-overview is the regulatory regime (how AI is governed by law), and israel-ai-strategy-2026 is the national strategy (how AI is strategically steered). Governance here is policy, regulation & industrial direction.

The through-line: standards are the connective tissue — the same frameworks (eu-ai-act, iso-iec-42001, OECD) that a firm certifies against are the ones a nation aligns its regime to. Organizational compliance and national regulation are two ends of one standards pipeline.

Current thesis — a taxonomy of governance instruments

The discipline's sharpest axis is how binding and how broad AI rules should be. The sources have grown that question from a binary (EU↔Israel) into a five-way instrument-philosophy map, plus a set of levers that sit outside the conduct-rule frame entirely.

The conduct-rule map — five jurisdictions, five philosophies

How a state writes the rules AI itself must follow:

The UK and Israel both dodge an EU-style act, but between them they show the "sectoral-soft" corner was hiding two distinct mechanics: top-down delegated principles (UK) vs bottom-up statutory reinterpretation (Israel). All five converge on risk-based-regulation — oversight scaled to impact (credit, medical, autonomous vehicles get the most scrutiny) — so the disagreement is about instrument, not the risk-based principle underneath. The two soft-law cells also share a crack: the UK renamed its AI Safety Institute → AI Security Institute (Feb 2025) and is trending toward a statutory frontier framework (2026) — soft drifting toward binding, the same pressure mooted for Israel's "Framework Law."

Instruments beyond conduct rules

Several sources push past what the rules say into other levers — and the newest shows the lever need not be pulled by a state at all:

The convergence substrate

Across all of this, the oecd-ai-principles definitional layer is the shared substrate — the eu-ai-act, the CoE treaty, US policy, the UK's five principles, and UN guidance all reuse the OECD AI-system definition. Divergence is at the instrument level, agreement at the definitional level. That is the caveat holding the whole map together.

Israel's posture — governance as competitive advantage

israel-ai-strategy-2026 ("Strengthening the Global Leadership of Israeli High-Tech… Where Excellence Outweighs Scale," Israel Innovation Authority, April 2026 draft) is primarily an industrial-competitiveness strategy (four pillars: Applications, AI Enablers, Technology/"Physical AI," Geopolitics/"Pax Silica"), but it treats regulation as an enabler, not a brake: regulatory sandboxes and a "global validation hub," dedicated certification pathways for autonomous-system safety, and active participation in international standard-setting as a five-year target. A nimble, light-touch regime is itself industrial policy.

Thesis lineage: an EU↔Israel binary (06-07) extended into regulation-by-existing-law (06-07), a four-way map adding China + US (06-10), a fifth UK cell that split the soft-law corner (06-12), the export-control instrument (06-14), governance-by-technical-standard at the IETF — non-state venue, input-side lever (06-18), and collective industry defense via Akrites — a second non-state venue, defense-side/operational (06-26).

Open questions

Growth edges

Ranked; each names the kind of source that would close it (see ../QUALITY.md → Growth edges).

  1. Do certifications correlate with anything? iso-iec-42001 certification and Trust Marks are assurance signals with no outcome data behind them — the same question early security certification answered badly. — needs: T1/T2 evidence linking certification to incident or harm rates.
  2. Does a preference signal ever become legally binding? ai-preferences/web-bot-auth are conventions until a jurisdiction gives them force. — needs: a statute, regulator decision or court ruling (T1).
  3. Does soft law hold under pressure? Israel's sectoral bet is untested against a major AI harm. — needs: an enforcement action or a post-incident regulatory review.

Coverage edges (added 2026-08-08, at the curator's request for a wider backlog). These widen what the spoke covers instead of answering an open question above; one ordinary solid source closes any.

  1. The regime already in force — closed 2026-08-14. garante-openai-jurisdiction-ruling (T1, European Law Blog on Tribunale di Roma R.G. 4785/2025) supplies both halves: the regulation (GDPR Art. 56 one-stop-shop, applied here for the first time to an AI provider's establishment timing) and the enforcement decision (the Garante's €15M ChatGPT fine, annulled on jurisdiction — never on the merits). Successor: the ruling creates an "AI enforcement gap" — a non-EEA provider that incorporates in the EU after a DPA opens an investigation can have the case redirected to the new home authority before any decision issues, for conduct that predates the EU entity. — needs: whether Ireland's DPC actually picks up the pre-establishment conduct, or the case simply lapses; and, separately, a ruling that reaches GDPR's substantive rules (lawfulness, transparency) applied to an AI system, which this one explicitly did not.

  2. Copyright and training data. What a model may legally be built from is being decided in court and appears nowhere here. — needs: a decided case or a legal review, T1/T2.

  3. Incident reporting. HALF-CLOSED 2026-08-22 — ai-incident-database (McGregor, IAAI-21, T1) supplies the database and its methodology, and disqualifies it as an outcome measure: coverage follows press attention (report count is "a proxy for interest"), there is no severity scale, and taxonomies are namespaced by design rather than shared. The Act's own Article 73 guidance is still unread, so the corpus cannot say what a "serious incident" legally is. — needs: the Commission's Article 73 guidance or a national implementing text. 6a. Is there any denominator anywhere? New 2026-08-22, the residue of edge 6. Every governance instrument in this spoke is justified by harm, and the field's largest harm collection is explicitly a case-finding tool, not a rate. — needs: a dataset with a population behind it — deployments as well as incidents — or a regulator's own reporting statistics under a mandatory duty. ../defensive-security-wiki hit the identical wall with breach data. HALF CLOSED 2026-08-25 (research pass) by ai-ml-device-adverse-events (Babic, Cohen, Stern, Li & Ouellet, npj Digital Medicine 8:328, T1) — and the half it closes is the diagnosis, not the ask. Medical AI is the field with a mandatory reporting duty, a published roster of 950 FDA-authorised AI/ML devices, and an open harm database; the authors linked 823 devices to 943 adverse-event reports over 2010–2023 and then wrote that "without an overall 'denominator' for the frequency of a medical device's use, it is not possible to talk about the relative safety of one product versus another." So the missing denominator is not a symptom of the AI Incident Database being a volunteer project — it is missing in a federal system built under statute. That makes three fields (ai-incident-database, breach data, medical devices) with the same hole, which is enough to stop treating it as a gap in the corpus's reading. 6b. A count of uses, not of things. New 2026-08-25, the residue. Every roster the corpus has found counts artefacts — devices authorised, models registered, systems declared — and every rate question needs episodes: how often the thing ran, on how many people. — needs: any published denominator of AI system usage under a regulatory duty (an EU AI Act Article 72 post-market monitoring return, an NHS or CMS utilisation figure for an authorised device, a payer's claims count), or a regulator stating plainly that it does not collect one. CLOSED 2026-08-28 (research pass) — in the negative, from the strongest available source: the Regulation's own text. ai-act-post-market-monitoring (T1, Official Journal) reads Articles 72, 19 and 12. Article 72 asks for performance and continuous-compliance data, never volume; deployer input is permissive ("may be provided by deployers"); law-enforcement operational data is carved out of the obligation entirely; and the monitoring plan lives inside Annex IV technical documentation, which describes the system and is shown on request rather than published. The exception is the finding: Article 12(3) does mandate "recording of the period of each use… start date and time and end date and time of each use" — for remote biometric identification alone, and Article 19 leaves those logs with the provider for six months with no reporting route. So the instrument can count uses, requires it once, and publishes it never. The edge asked for a denominator or a regulator saying it collects none; the answer is that the duty was drafted so the question does not arise. Checked against the Regulation's other reporting duty, 2026-10-07. Article 72 is monitoring; Article 73 is mandatory serious-incident reporting — 15/2/10-day deadlines depending on severity, straight from the Official Journal text. It is the one duty in the whole Regulation with real teeth, and it is incident-by-incident with no aggregate field anywhere in it: nothing asks how many systems are deployed or how often any of them ran. Between the two Articles the Regulation actually enforces, neither supplies a denominator — recorded on ai-act-post-market-monitoring, which now covers both. Successor (edge 6c): the AI Act is one regime and the health regulators are another, with a real chance of a different answer — an FDA post-market pathway or an NHS/CMS utilisation return attaches to a device that is billed, and billing produces episode counts as a by-product. — needs: a utilisation or claims count for an authorised AI device, T1/T2. Not eligible before 2026-09-11. CLOSED 2026-09-27 (research pass) by ai-device-cpt-claims (T1, Wu et al., NEJM AI 2024, read at abstract level). Billing did produce the count: 11 billion CPT claims (2018 to mid-2023), and only AI for coronary artery disease and AI for diabetic retinopathy pass 10,000 claims. Use clusters in richer, metropolitan zip codes with academic centres. The count came from researchers working on a commercial claims database, not from the FDA, so the regulator's silence stands. Successor (edge 6e): divide one by the other. ai-ml-device-adverse-events has the harm reports and ai-device-cpt-claims now has use for two device classes. — needs: the per-device claim counts (the paper's own table, T1) and the MAUDE reports for the same devices, or any study that already computed an adverse-event rate per use for an AI device. Not eligible before 2026-10-11. 6d. Is removing a mark regulated anywhere? New 2026-09-06. The corpus has the marking duty (china-ai-regulation) and now the removal tooling (watermarks-remover, content-credentials-c2pa) and nothing in between. — needs: the EU AI Act's transparency article on AI-generated content read from the Official Journal (does it name machine-readable marking, and does it say anything about stripping one?), or any statute or regulator decision treating provenance removal as a violation. T1.

  4. US law below the federal level. us-ai-policy and eo-14409 cover Washington; Colorado's AI Act and California's statutes bind deployers now. — needs: the statute texts plus one compliance analysis.

  5. Does certification compare like with like? New 2026-10-07, promoted from the hub dream ledger. iso-iec-42001 certifies a management process; nothing requires two certified organisations to define "risk" or "harm" the same way, so an nemko-digital AI Trust Mark may not compare across holders before any question of safety outcomes. — needs: an audit-sampling study of how several 42001-certified organisations define and measure AI risk, or incident rates of certified vs uncertified deployers, T1/T2.

Governance without a denominator (quality cycle, 2026-08-22)

This spoke catalogues instruments: eu-ai-act, iso-iec-42001, nist-ai-rmf, oecd-ai-principles, national strategies, assurance services. Every one of them is justified by harm, and until today no page here held a single record of harm.

ai-incident-database is the field's largest such collection — over 1,000 reports at the time of the paper — and reading it carefully makes the gap sharper rather than closing it. The AIID is built on the CVE model, not the NTSB model: its contribution is an identifier and a searchable archive, with analysis left to whoever wants to layer a taxonomy on top. Incidents are deliberately untitled, taxonomies are namespaced so partners can disagree in public, and the paper is candid that report volume tracks interest rather than severity. Entry is by public submission of press coverage, not by obligation.

So the answer to "how much harm is there" is still unavailable, and now it is unavailable for a documented reason. Growth edge 1 — does certification correlate with anything — cannot be answered from this data, because there is no population to divide by. ../defensive-security-wiki reached the same conclusion from the other direction with breach datasets that carry no per-company compliance variable, and that is two spokes independently discovering that the compliance-versus- outcomes question has no dataset behind it in either field.

One structural note worth keeping. The AIID's refusal of a universal ontology is a rational response to who funds it — partnership-on-ai's board is split between corporate partners and non-profits, and the corporate partners own the systems in the incidents. A regulator writing eu-ai-act Article 73 guidance does not have that option: a reporting duty needs one definition of "serious incident," binding in every member state. The soft-governance and hard-governance layers are being asked for incompatible things by the same evidence problem.

The US instrument was read wrong, and the error has a shape (quality cycle, 2026-08-30)

us-ai-policy was the one page in this spoke the freshness queue kept flagging and the cycle kept deferring. Read at last against crs-eo-14409-explained (CRS IF13268 v2, T1), it produced the largest correction this spoke has made.

eo-14409 is voluntary. This wiki has said since June that it "requires submitting new model releases to the government 30 days before release" and called it "a recurring pre-clearance pipeline." The Congressional Research Service says companies "are asked to participate in a voluntary review process" and that the order "stops short of formal licensing or requiring preclearance." The order does not even define covered frontier model; that was itself a deliverable, due 1 August 2026 out of a classified benchmarking process. Nine tasks, every one landing on an agency and none on a company.

And CRS lists the Anthropic suspension as a different instrument, inviting Congress to weigh the EO's voluntary assessment "especially in contrast to stronger government direction," then naming "the Administration's other approaches, such as requiring Anthropic to suspend access." So what this spoke recorded as two enforcement beats of a standing gate is, on the better-sourced reading, a voluntary framework plus two separate ad-hoc actions.

The shape of the error is worth more than the correction. Nothing was fabricated. Every fact came from real reporting, and the reporting was about real events: two frontier labs were restricted within a fortnight of an order about frontier-model security. What the wiki added was the causal joint — it read a sequence as a mechanism. A voluntary framework and a mandatory gate are indistinguishable from outside when everyone participates and two participants are publicly restricted at once. Only the primary text separates them, and nobody here had read it.

This is the same failure the corpus keeps finding in other people's work — synthesis elsewhere in this hub tracks secondary coverage widening a vendor's narrow claim — arriving from the inside. The generalizable rule: when a page describes an instrument as compelling something, the obligation is the claim to check against primary text, because it is the part that carries the whole reading and the part reporting is least careful about.

Also corrected: the deregulatory pole now has a number and mechanisms. eo-14365 (11 December 2025) is paged — a DOJ AI Litigation Task Force inside 30 days, $42bn of BEAD broadband funding conditioned on states repealing "onerous" AI laws, an option for states to sign binding non-enforcement agreements instead of repealing, carve-outs for child safety, data-centre infrastructure and state procurement, and a directive to draft the preempting legislation that does not yet exist. It has preempted nothing: preemption flows from Congress, and Congress has passed no AI statute.

So the US cell of this spoke's four-way instrument comparison is, as of today: two ad-hoc restrictions, one voluntary framework, one order to sue, and a bill nobody has written. It has been standing in the comparison as a settled national position, and it is not one.

One fact the corpus lacked entirely. NIST already held pre-release access to Anthropic's and OpenAI's major models by private agreement, predating every order here, with no public transparency and no coverage of other labs. Government pre-release model access in the US started as a commercial arrangement.

The risk statement acquires liability (2026-09-16)

anthropic-mother-of-all-risk-factors is held at one remove — the FT text is paywalled and only the headline is on the page — but the event it names is clear from the secondary record. Every instrument this spoke has read asks a frontier lab to assess its own risk and publish or submit the assessment: eo-14409's pre-release filing, OpenAI's preparedness-framework, Anthropic's own risk reports. None puts the lab on the hook for the assessment being wrong. An S-1 risk factor does. It is a statement by management, under securities liability, and Anthropic has to write one about the failure mode its alignment lead puts above 10% within the decade and says the company has no plan for. The governance point is the asymmetry: a regulator can be told "risk is low" with no cost if it is not, while an investor cannot — so the sharpest disclosure of frontier risk this year may come from the SEC's boilerplate rather than from any AI statute. Watch the filing itself when it lands; this page should be corrected from the primary.

Contradictions / tensions

Cross-spoke adjacency

Index — AI Governance Wiki

Catalog of every page, grouped by schema.org @type. Spine: synthesis (thesis), log.md (history), this file (catalog). Spun out of the hub ai-governance cluster 2026-06-07. Read synthesis first for the thesis (organizational vs national governance; horizontal-binding vs sectoral-soft law).

DefinedTerm (concepts / mechanisms / standards)

Legislation

GovernmentOrganization

Organization

Person

TechArticle (sources)

Report / WebPage / Article (sources)

SoftwareSourceCode (sources)

Synthesis

Related pages

Links to

Linked from