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:

• EU (eu-ai-act) — rights-driven, horizontal-binding: one cross-cutting risk-tiered law, enforced. nemko-digital‘s services largely exist to help firms comply with this kind of regime.
• China — state-driven, content-binding: early, targeted, enforceable rules (2022 algorithm-registration with the CAC — 1,400+ filed; 2023 generative-AI Interim Measures; 2025 mandatory AI-content labeling/watermarking), organized around content control + data localization rather than EU-style rights tiers — the hard-law end of the AI-provenance spectrum the sibling speech-audio-wiki tracks as voluntary SynthID.
• US — market-driven, fragmented/deregulatory: no federal horizontal law; the Biden 2023 safety-testing EO was rescinded, and the Dec-2025 Trump EO moves to preempt state laws (California/Colorado). The US alone adds a federal-vs-subnational preemption conflict inside the jurisdiction, not just variation across them.
• UK — pro-innovation, cross-sector top-down: five central principles (safety/security/robustness; transparency/explainability; fairness; accountability/governance; contestability/redress) delegated to existing sector regulators, with no single AI law.
• Israel (responsible-innovation) — innovation-driven, sectoral-soft: AI rules kept light, leaning bottom-up on existing statutes reinterpreted per sector, with sandboxes and an “export-ready” motive (israel-ai-regulation-overview).

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:

• Regulation by existing law. lofrayer-bar-association shows incumbent professional-licensing law doing the work: Israel’s Bar Association invokes a 1961 statute (unauthorized practice of law) to threaten an AI traffic-fine startup (professional-licensing-and-ai) — no AI-specific law required, the sharp end of “existing statutes reinterpreted for AI.” The incumbent’s case (algorithms lack oversight, ethics, insurance, accountability) is itself the assurance logic nemko-digital sells, here weaponized by a profession to defend its monopoly.
• Binding international treaty. Above the voluntary tier sits a treaty binding states — the CoE Framework Convention (signed by EU+US+UK) — a different shape from a regulation of products (EU AI Act). So even the binding tier is not one thing.
• Security-driven release control — now a named, repeated regime. What looked like a one-off export designation is a standing instrument: Executive Order 14409 (2 June 2026) requires submitting frontier models 30 days before release (export-controls-on-ai). It governs AI by gatekeeping distribution on national-security grounds rather than setting conduct rules, pointing the chip-era export-control instrument inward at domestic frontier labs. Two enforcement beats in two weeks prove it repeats: Anthropic’s Fable 5 / Mythos 5 (worldwide suspension, misuse/jailbreak rationale) and OpenAI’s GPT-5.6 (release limited to government-approved partners, approved customer-by-customer) — suspension and graduated gating as two settings of the same dial. This complicates the US “deregulatory” cell: the same administration clearing domestic regulation runs a blunt, discretionary security gate on release. “Market-driven / light-touch” describes US conduct regulation, not its security posture. Three further threads sharpen here. (a) Efficacy — practitioners argue the gate accelerates the shift to self-hostable open-weight / Chinese models (GLM-5.2 already beating GPT-5.5 on coding benchmarks), so the access-denial lever may feed the proliferation it targets — “the security goal and the market reaction now run against each other.” (b) Evidence weight — the EO text is verifiable (Federal Register), but per-model application is still secondhand (“reportedly,” via trade press / staff Q&As); weight by who benefits (Amazon = Anthropic investor and competitor; Sacks = political principal; TNS owner Insight Partners invests in both OpenAI and Anthropic). (c) The capability premise is now provider-documented. OpenAI’s GPT-5.6 system card rates the model High in cyber and High in bio/chem under its Preparedness Framework (none Critical) and states “government coordination was requested before broader deployment” — so the too-dangerous-to-ship- freely premise behind the GPT-5.6 gate is no longer only “reportedly,” and the provider’s internal capability gate now sits visibly alongside the state’s external release gate (the same staged release described from both ends). The standing caution: a system card is a self-assessment — the lab grading its own model against self-defined “High/Critical” thresholds with no independent auditor, the assurance gap public standards (iso-iec-42001) exist to close. (This does not retro-verify the separate Anthropic jailbreak claim.) (d) The regime has now run one full cycle — and it revealed a missing half. Anthropic’s models came back online on 1 July 2026 after a 20-day ban→review→re-release loop, the first completed enforcement run. The exit was an ad hoc, multi-agency crash review: the Center for AI Standards and Innovation and the NSA rejected the first fixes, agency heads approved gradually, Commerce coordinated — with no published criteria, threshold, or timeline. So EO 14409 mandates a defined entry (submit 30 days out) but leaves the exit (approval) to discretion — the reporting’s own bottom line is that a “framework for approving future models with transparency standards and timelines” still doesn’t exist. This also names who adjudicates: CAISI+NSA are the technical gate — the US standards-body-as-binding-regulator, the coercive counterpart to the voluntary iso-iec-42001/nist-ai-rmf assurance role (and the same “safety→security/standards” renaming the UK made). The contested trigger meanwhile got more contested: cybersecurity experts wrote an open letter saying rival models share the flaw, deepening the “already exists in other public models” dispute rather than settling it.
• Governance by technical standard (the input side). EFF surfaces a venue the rest of the map misses: the IETF, where AI rules get written in protocol language before any legislature acts. Two working-group efforts — ai-preferences (machine-readable “don’t train on this” signals, an evolution of robots.txt) and web-bot-auth (cryptographic bot identification) — target AI’s inputs (its access to web content) rather than its conduct or distribution. This adds two firsts to the map: a non-state, standards-body venue (private ordering that law may later make binding), and an input-side lever (the training-data supply chain) distinct from EU-style conduct rules and China’s output-side content labeling. EFF’s warning is that a gate built for AI crawlers is indifferent to who crawls — the same standard throttles archiving, research, security scanning, and accessibility tools. Caveat: EFF is an advocacy source (T2, POV) and the drafts are unratified; the mechanism is solid, the threat scope is contested.
• Governance by collective industry defense (a non-state, operational venue). akrites (akrites-launch-tns) adds a lever none of the above had: ~20 competitors (Anthropic, Google, OpenAI, Microsoft, AWS…) pooling coordinated open-source vulnerability response under the neutral Linux Foundation (a shared SIRT consolidating findings; CVE/CVSS; OpenSSF/ Alpha-Omega funding). It governs AI-related security risk not by rules on the model but by hardening the targets — fix OSS vulnerabilities before AI-armed attackers exploit them. Two firsts for the map: (1) a second non-state venue alongside the IETF, but an operational defense-coordination body rather than a rule-writing one — so “non-state governance” now spans input-side rule-writing (IETF) and defense-side incident coordination (Akrites); (2) it is the defense-side counterpart to export control on the same trigger — the Fable 5/Mythos 5 incident produced both a state access-denial instrument and an industry collective-defense one. And it is the constructive corollary of the efficacy critique: precisely because the offensive capability replicates (Anthropic’s own rebuttal — “already exists in other public models”), the industry bet shifts from gating the tool to coordinating the defense. (Day-one reporting on intent; T3, one stat unverified — outcomes unproven.)

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

• Who governs the open web — and on whose behalf? EFF frames the ietf AI-crawling work as publisher/platform interests capturing a neutral standards process to gate public content. Open: do ai-preferences/web-bot-auth settle as narrow AI-training opt-outs or as a general access-licensing toll booth? Does any jurisdiction actually make a preference signal legally binding (the step that turns convention into right)? And does civil society (EFF) hold the openness line, or do standards venues default toward the parties who fund participation?
• Does soft law hold under pressure? Israel’s bet is that sectoral soft law + sandboxes beats a horizontal act. Untested against a major AI harm; a “Framework Law” on algorithmic discrimination is already mooted (israel-ai-regulation-overview) — the first crack toward horizontal rules?
• Do certifications mean anything? nemko-digital‘s “AI Trust Mark” and iso-iec-42001 certification are assurance signals — but is there evidence they correlate with actual safety, or are they (like early security certs) compliance theater? No outcome data yet.
• Whose standards win? EU AI Act vs. NIST RMF vs. ISO vs. OECD — convergent or competing? Israel hedges by aligning to OECD and “watching” the EU. Track whether a dominant standard emerges. Mapped further (2026-06-09): the oecd-ai-principles turn out to be the shared substrate — the eu-ai-act, the CoE treaty, US policy and UN guidance all reuse the OECD AI-system definition, so convergence is real at the definitional/principles layer even where instruments diverge. The binding tier now has two distinct shapes: a regulation of products (EU AI Act) and a treaty binding states (framework-convention-on-ai, signed by EU+US+UK), both above the voluntary oecd-ai-principles/iso-iec-42001/nist-ai-rmf tier. So “horizontal vs sectoral” isn’t binary — there’s a binding-international-treaty option too.
• What does the approval half of the release gate look like? EO 14409‘s first completed cycle (Anthropic, 20 days, re-released 1 July) ran the approval as ad hoc discretion — CAISI+NSA judging fixes with no published criteria or clock. Open: does the “framework with transparency standards and timelines” the reporting says is still missing actually get built — turning a discretionary security reach into a legible, repeatable process — or does approval stay case-by-case leverage? And does CAISI harden into a standing frontier-model regulator (the US analogue to an EU notified body) or remain an incident-driven reviewer? GPT-5.6 still on hold is the next test.
• Vendor & state incentives. One source is a governance vendor (sells compliance); two are a government (sells its own competitiveness narrative). Both have skin in the framing — weight claims accordingly.

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.

4. The regime already in force. The spoke tracks AI-specific instruments and has no page on data protection — GDPR, impact assessments, the lawful basis a training set needs — which binds today. — needs: the regulation text plus one enforcement decision touching AI.
5. 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.
6. Incident reporting. eu-ai-act creates reporting duties and the corpus holds no page on any incident taxonomy or database, so governance instruments are catalogued with no evidence of harm to govern. — needs: the AI Incident Database’s methodology or the Act’s Article 73 guidance.
7. 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.

Contradictions / tensions

• Horizontal-binding vs. sectoral-soft (above) — not a fact conflict but a genuine policy fork between the eu-ai-act model and Israel’s responsible-innovation model. Worth tracking which delivers better safety and innovation outcomes.
• Deregulatory yet interventionist (US). anthropic-export-ban-2026 sits in tension with us-ai-policy‘s “clearing regulation away” read: a hard export-control strike on a domestic lab. Not a fact conflict — different layers (domestic conduct deregulation vs security intervention) — but a caution against reading the US cell as uniformly light-touch. Also a disputed-facts case: Anthropic says the flagged capability “already exists in other public models,” contra the government/Amazon misuse framing — recorded, unresolved.
• Public-protection vs. access / anti-monopoly. lofrayer-bar-association pits the assurance rationale (only licensed, insured, accountable humans should do this work) against access-to-justice / anti-guild (licensing prices ordinary people out, so the AI tool’s real alternative is nothing). The same “AI lacks oversight/accountability” argument that justifies governance can also be regulatory capture — a caution that governance rhetoric and incumbent protection are hard to tell apart.

Cross-spoke adjacency

• ../agentic-tooling-wiki — owns the tools for building/running agents; governance of those agents (risk, compliance, assurance) lives here.
• ../platform-ops-wiki — operating AI in production (SRE/observability); the parked ai-productionization cluster (engineering-delivery reality) is adjacent-but-distinct — this spoke is the regulatory/assurance layer, not the delivery-engineering one.
• ../llm-providers-wiki — the model/provider market that these rules regulate.

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)

• ai-governance — the discipline: setting & enforcing the rules AI operates under (accountability, risk, trust) · domain
• risk-based-regulation — scaling oversight intensity to an AI system’s potential impact; the shared principle · practice
• professional-licensing-and-ai — incumbent professional/licensing law (unauthorized practice of law) as de facto AI regulation · practice
• responsible-innovation — Israel’s governance framing: enable trustworthy AI without a horizontal law · practice
• china-ai-regulation — China’s state-driven, content-binding regime (CAC algorithm registration; 2023 generative-AI measures; 2025 content labeling) · source · domain
• us-ai-policy — the US market-driven, fragmented/deregulatory approach (rescinded Biden EO; Trump 2025 preemption; state patchwork) · source · domain
• uk-ai-regulation — the UK “pro-innovation” approach: 5 cross-sector principles delegated to existing regulators; no single AI law (statutory framework mooted 2026) · source · domain
• iso-iec-42001 — ISO/IEC 42001, the AI management-systems standard (certifiable) · standard
• nist-ai-rmf — NIST AI Risk Management Framework; voluntary US risk-management guidance · standard
• oecd-ai-principles — the first intergovernmental AI standard (2019/2024); the soft-law convergence layer · source · standard
• export-controls-on-ai — export control as an AI-governance instrument (security/distribution-gatekeeping); the new lever beyond conduct rules · standard
• preparedness-framework — OpenAI’s internal dangerous-capability risk-tiering (High/Critical across cyber/bio/self-improvement); frontier-lab self-governance (vs public standards) · standard
• ai-preferences — machine-readable “don’t train on this” signals (robots.txt evolution); IETF input-side lever on AI’s training data · mechanism
• web-bot-auth — cryptographic bot identification; an IETF anti-abuse standard with selective-access (licensing) double-use · mechanism

Legislation

• eu-ai-act — the EU’s horizontal, binding, risk-tiered AI law; the reference regime
• framework-convention-on-ai — Council of Europe; the first legally binding international treaty on AI (2024) · source
• eo-14409 — US Executive Order 14409 (2 Jun 2026); 30-day pre-release submission of frontier models; the standing instrument behind the Anthropic + GPT-5.6 release controls

GovernmentOrganization

• caisi — Center for AI Standards and Innovation; the US federal frontier-model evaluator (with the NSA) that gates release under eo-14409 — standards-body-as-binding-regulator · url · axios.com

Organization

• nemko-digital — AI governance/assurance & certification provider (“AI Trust Mark”)
• israel-innovation-authority — Israel’s gov innovation agency; publisher of the national AI strategy
• ietf — Internet Engineering Task Force; standards body as an AI-governance venue (protocol-as-regulation)
• electronic-frontier-foundation — digital-rights NGO; civil-society voice for the open web in standards/policy venues
• akrites — Linux Foundation body coordinating open-source vulnerability defense vs AI-enabled threats; industry collective self-governance (a new modality) · source · akrites.org
• linux-foundation — non-profit host of Akrites + OpenSSF/Alpha-Omega; a non-state, defense-side governance venue (the operational counterpart to the IETF)
• the-new-stack — tech-news publisher of the Akrites + GPT-5.6 sources; owner Insight Partners invests in OpenAI & Anthropic (disclosed conflict)

Person

• tori-noble — EFF author of the IETF open-web critique
• sam-altman — OpenAI CEO; named voice on the receiving end of EO 14409 (the GPT-5.6 customer-by-customer gate)
• howard-lutnick — US Commerce Secretary; the regulator at the center of the EO 14409 release controls (made the 12 June call taking Anthropic’s models offline) · url

TechArticle (sources)

• nist-ai-rmf-overview — NIST’s official AI RMF landing page; AI RMF 1.0 (2023-01-26), Playbook, 7 trustworthiness characteristics, GenAI Profile NIST-AI-600-1 (2024-07-26) · source · T1 · nist.gov
• iso-42001-align-explainer — A-LIGN explainer on ISO/IEC 42001 structure (Annex A–D, six PDCA requirement areas) and its non-harmonized status vs the EU AI Act · source · T2 · a-lign.com
• cyber-export-control-history — TechCrunch: PGP→spyware→Mythos history arguing cyber/AI export controls don’t stop proliferation (the efficacy critique) · source · T2 · techcrunch.com

Report / WebPage / Article (sources)

• israel-ai-strategy-2026 — Israel Innovation Authority national AI strategy; “Excellence Outweighs Scale,” 4 pillars · source · raw PDF
• israel-ai-regulation-overview — Israel’s sectoral soft-law regulatory framework (Regulations.ai) · source · regulations.ai
• nemko-ai-governance-services — Nemko’s AI-GRC services & “AI Trust Mark” certification · source · digital.nemko.com
• lofrayer-bar-association — Israel Bar Association moves to shut an AI traffic-fine startup (unauthorized practice of law) · source · calcalistech.com
• anthropic-export-ban-2026 — reported US export-control ban on Anthropic’s Fable 5 / Mythos 5 over a contested jailbreak/cyber-misuse claim · source · T4 · techcrunch.com
• eff-web-under-attack-ietf — EFF: publishers/Big Tech bending IETF AI-crawling standards into a gate on the open web · source · T2 · eff.org
• akrites-launch-tns — The New Stack: ~20 firms launch Akrites (Linux Foundation) for coordinated OSS vuln defense after the Fable 5 ban; Anthropic a founding member · source · T3 · thenewstack.io
• openai-gpt56-access-restriction — The New Stack: US limits GPT-5.6 to government-approved partners (customer-by-customer) under EO 14409; the 2nd lab hit, naming the standing regime · source · T3 · thenewstack.io
• gpt56-system-card — OpenAI’s official GPT-5.6 system card: High cyber + bio/chem under the Preparedness Framework, layered mitigations, “government coordination requested before broader deployment” — the primary-source anchor for the restriction (self-assessment) · source · T1 · deploymentsafety.openai.com
• anthropic-models-revived-2026 — Axios behind-the-scenes: Anthropic’s Fable 5 / Mythos 5 back online 1 July after a 20-day multi-agency review (CAISI+NSA gated it); first completed EO 14409 cycle, and the approval framework still doesn’t exist · source · T2 · axios.com

Synthesis

• synthesis — the evolving thesis: organizational vs national governance; the horizontal-vs-sectoral fork