182 lines
17 KiB
Markdown
182 lines
17 KiB
Markdown
# Research Musing — 2026-04-06
|
|
|
|
**Research question:** Is the Council of Europe AI Framework Convention a stepping stone toward expanded governance (following the Montreal Protocol scaling pattern) or governance laundering that closes political space for substantive governance?
|
|
|
|
**Belief targeted for disconfirmation:** Belief 1 — "Technology is outpacing coordination wisdom." Specifically: the pessimistic reading of scope stratification as governance laundering. If the CoE treaty follows the Montreal Protocol trajectory — where an initial 50% phasedown scaled to a full ban as commercial migration deepened — then my pessimism about AI governance tractability is overcalibrated. The stepping stone theory may work even without strategic actor participation at step one.
|
|
|
|
**Disconfirmation target:** Find evidence that the CoE treaty is gaining momentum toward expansion (ratifications accumulating, private sector opt-in rates high, states moving to include national security applications). Find evidence that the Montreal Protocol 50% phasedown was genuinely intended as a stepping stone that succeeded in expanding, and ask whether the structural conditions for that expansion exist in AI.
|
|
|
|
**Why this question:** Session 04-03 identified "governance laundering Direction B" as highest value: the meta-question about whether CoE treaty optimism is warranted determines whether the entire enabling conditions framework is correctly calibrated for AI governance. If I'm wrong about the stepping stone failure, I'm wrong about AI governance tractability.
|
|
|
|
**Keystone belief at stake:** If the stepping stone theory works even without US/UK participation at step one, then my claim that "strategic actor opt-out at non-binding stage closes the stepping stone pathway" is falsified. The Montreal Protocol offers the counter-model: it started as a partial instrument without full commercial alignment, then scaled. Does AI have a comparable trajectory?
|
|
|
|
---
|
|
|
|
## Secondary research thread: Commercial migration path emergence
|
|
|
|
**Parallel question:** Are there signs of commercial migration path emergence for AI governance? Last session identified this as the key structural requirement (commercial migration path available at signing, not low competitive stakes). Check:
|
|
- Anthropic's RSP (Responsible Scaling Policy) as liability framework — has it been adopted contractually by any insurer or lender?
|
|
- Interpretability-as-product: is anyone commercializing alignment research outputs?
|
|
- Cloud provider safety certification: has any cloud provider made AI safety certification a prerequisite for deployment?
|
|
|
|
This is the "constructing Condition 2" question from Session 04-02. If commercial migration paths are being built, the enabling conditions framework predicts governance convergence — a genuine disconfirmation target.
|
|
|
|
---
|
|
|
|
## What I Searched
|
|
|
|
1. CoE AI Framework Convention ratification status 2026
|
|
2. Montreal Protocol scaling history — full mechanism from 50% phasedown to full ban
|
|
3. WHO PABS annex negotiations current status
|
|
4. CoE treaty private sector opt-in — which states are applying to private companies
|
|
5. Anthropic RSP 3.0 — Pentagon pressure and pause commitment dropped
|
|
6. EU AI Act streamlining — Omnibus VII March 2026 changes
|
|
7. Soft law → hard law stepping stone theory in academic AI governance literature
|
|
|
|
---
|
|
|
|
## What I Found
|
|
|
|
### Finding 1: CoE Treaty Is Expanding — But Bounded Stepping Stone, Not Full Montreal Protocol
|
|
|
|
EU Parliament approved ratification on March 11, 2026. Canada and Japan have signed (non-CoE members). Treaty entered force November 2025 after UK, France, Norway ratified. Norway committed to applying to private sector.
|
|
|
|
BUT:
|
|
- National security/defense carve-out remains completely intact
|
|
- Only Norway has committed to private sector application — others treating it as opt-in and not opting in
|
|
- EU is simultaneously ratifying the CoE treaty AND weakening its domestic EU AI Act (Omnibus VII delays high-risk compliance 16 months)
|
|
|
|
**The form-substance divergence:** In the same week (March 11-13, 2026), the EU advanced governance form (ratifying binding international human rights treaty) while retreating on governance substance (delaying domestic compliance obligations). This is governance laundering at the domestic regulatory level — not just an international treaty phenomenon.
|
|
|
|
CLAIM CANDIDATE: "EU AI governance reveals form-substance divergence simultaneously — ratifying the CoE AI Framework Convention (March 11, 2026) while agreeing to delay high-risk EU AI Act compliance by 16 months (Omnibus VII, March 13, 2026) — confirming that governance laundering operates across regulatory levels, not just at international treaty scope." (confidence: proven — both documented facts, domain: grand-strategy)
|
|
|
|
---
|
|
|
|
### Finding 2: Montreal Protocol Scaling Mechanism — Commercial Migration Deepening Is the Driver
|
|
|
|
Full scaling timeline confirmed:
|
|
- 1987: 50% phasedown (DuPont had alternatives, pivoted)
|
|
- 1990 (3 years): Accelerated to full CFC phaseout — alternatives proving more cost-effective
|
|
- 1992: HCFCs added to regime
|
|
- 1997: HCFC phasedown → phaseout
|
|
- 2007: HCFC timeline accelerated further
|
|
- 2016: Kigali Amendment added HFCs (the CFC replacements)
|
|
|
|
The mechanism: EACH expansion followed deepening commercial migration. Alternatives becoming more cost-effective reduced compliance costs. Lower compliance costs made tighter standards politically viable.
|
|
|
|
The Kigali Amendment is particularly instructive: the protocol expanded to cover HFCs (its own replacement chemistry) because HFO alternatives were commercially available by 2016. The protocol didn't just survive as a narrow instrument — it kept expanding as long as commercial migration kept deepening.
|
|
|
|
**The AI comparison test:** For the CoE treaty to follow this trajectory, AI governance would need analogous commercial migration deepening — each new ratification or scope expansion would require prior commercial interests having already made the transition to governance-compatible alternatives. The test case: would the CoE treaty expand to cover national security AI once a viable governance-compatible alternative to frontier military AI development exists? The answer is structurally NO — because unlike CFCs (where HFCs were a genuine substitute), there is no governance-compatible alternative to strategic AI advantage.
|
|
|
|
CLAIM CANDIDATE: "The Montreal Protocol scaling mechanism (commercial migration deepening → reduced compliance cost → scope expansion) predicts that the CoE AI Framework Convention's expansion trajectory will remain bounded by the national security carve-out — because unlike CFCs where each major power had a commercially viable alternative, no governance-compatible alternative to strategic AI advantage exists that would permit military/frontier AI scope expansion." (confidence: experimental — structural argument, not yet confirmed by trajectory events, domain: grand-strategy)
|
|
|
|
---
|
|
|
|
### Finding 3: Anthropic RSP 3.0 — The Commercial Migration Path Runs in Reverse
|
|
|
|
On February 24-25, 2026, Anthropic dropped its pause commitment under Pentagon pressure:
|
|
- Defense Secretary Hegseth gave Amodei a Friday deadline: roll back safeguards or lose $200M Pentagon contract + potential government blacklist
|
|
- Pentagon demanded "all lawful use" for military, including AI-controlled weapons and mass domestic surveillance
|
|
- Mrinank Sharma (led safeguards research) resigned February 9 — publicly stated "the world is in peril"
|
|
- RSP 3.0 replaces hard operational stops with "ambitious but non-binding" public Roadmaps and quarterly Risk Reports
|
|
|
|
This is the exact inversion of the DuPont 1986 pivot. DuPont developed alternatives, found it commercially valuable to support governance, and the commercial migration path deepened the Montreal Protocol. Anthropic found that a $200M military contract was commercially more valuable than maintaining governance-compatible hard stops. The commercial migration path for frontier AI runs toward military applications that require governance exemptions.
|
|
|
|
**Structural significance:** This closes the "interpretability-as-commercial-product creates migration path" hypothesis from Session 04-02. Anthropic's safety research has not produced commercial revenue at the scale of Pentagon contracts. The commercial incentive structure for the most governance-aligned lab points AWAY from hard governance commitments when military clients apply pressure.
|
|
|
|
CLAIM CANDIDATE: "The commercial migration path for AI governance runs in reverse — military AI creates economic incentives to weaken safety constraints rather than adopt them, as confirmed by Anthropic's RSP 3.0 (February 2026) dropping its pause commitment under a $200M Pentagon contract threat while simultaneously adding non-binding transparency mechanisms, following the DuPont-in-reverse pattern." (confidence: proven for the specific case, domain: grand-strategy + ai-alignment)
|
|
|
|
---
|
|
|
|
### Finding 4: WHO PABS — Extended to April 2026, Structural Commercial Divide Persists
|
|
|
|
March 28, 2026: WHO Member States extended PABS negotiations to April 27-May 1. May 2026 World Health Assembly remains the target.
|
|
|
|
~100 LMIC bloc maintains: mandatory benefit sharing (guaranteed vaccine/therapeutic/diagnostic access as price of pathogen sharing).
|
|
Wealthy nations: prefer voluntary arrangements.
|
|
|
|
The divide is not political preference — it's competing commercial models. The pharmaceutical industry (aligned with wealthy-nation governments) wants voluntary benefit sharing to protect patent revenue. The LMIC bloc wants mandatory access to force commercial migration (vaccine manufacturers providing guaranteed access) as a condition of pathogen sharing.
|
|
|
|
Update to Session 04-03: The commercial blocking condition is still active, more specific than characterized. PABS is a commercial migration dispute: both sides are trying to define which direction commercial migration runs.
|
|
|
|
---
|
|
|
|
### Finding 5: Stepping Stone Theory Has Domain-Specific Validity
|
|
|
|
Academic literature confirms: soft → hard law transitions occur in AI governance for:
|
|
- Procedural/rights-based domains: UNESCO bioethics → 219 countries' policies; OECD AI Principles → national strategies
|
|
- Non-strategic domains: where no major power has a competitive advantage to protect
|
|
|
|
Soft → hard law fails for:
|
|
- Capability-constraining governance: frontier AI development, military AI
|
|
- Domains with strategic competition: US-China AI race, military AI programs
|
|
|
|
ASEAN is moving from soft to hard rules on AI (January 2026) — smaller bloc, no US/China veto, consistent with the venue bypass claim.
|
|
|
|
**Claim refinement needed:** The existing KB claim [[international-ai-governance-stepping-stone-theory-fails-because-strategic-actors-opt-out-at-non-binding-stage]] is too broad. It applies to capability-constraining governance, but stepping stone theory works for procedural/rights-based AI governance. A scope qualifier would improve accuracy and prevent false tensions with evidence of UNESCO-style stepping stone success.
|
|
|
|
---
|
|
|
|
## Synthesis: Governance Laundering Pattern Confirmed Across Three Levels
|
|
|
|
**Disconfirmation result:** FAILED again. The stepping stone theory for capability-constraining AI governance failed the test. The CoE treaty is on a bounded expansion trajectory, not a Montreal Protocol trajectory.
|
|
|
|
**Key refinement:** The governance laundering pattern is now confirmed at THREE levels simultaneously, within the same month (March 2026):
|
|
1. International treaty: CoE treaty expands (EU ratifies, Canada/Japan sign) but national security carve-out intact
|
|
2. Corporate self-governance: RSP 3.0 drops hard stops under Pentagon pressure, replaces with non-binding roadmaps
|
|
3. Domestic regulation: EU AI Act compliance delayed 16 months through Omnibus VII
|
|
|
|
This is the strongest evidence yet that form-substance divergence is not incidental but structural — it operates through the same mechanism at all three levels. The mechanism: political/commercial pressure forces the governance form to advance (to satisfy public demand for "doing something") while strategic/commercial interests ensure the substance retreats (to protect competitive advantage).
|
|
|
|
**The Montreal Protocol comparison answer:**
|
|
The CoE treaty will NOT follow the Montreal Protocol trajectory because:
|
|
1. Montreal Protocol scaling required deepening commercial migration (alternatives becoming cheaper)
|
|
2. AI governance commercial migration runs in reverse (military contracts incentivize removing constraints)
|
|
3. The national security carve-out reflects permanent strategic interests, not temporary staging
|
|
4. Anthropic RSP 3.0 confirms the commercial incentive direction empirically
|
|
|
|
The Montreal Protocol model predicts governance expansion only when commercial interests migrate toward compliance. For AI, they're migrating away.
|
|
|
|
---
|
|
|
|
## Carry-Forward Items (STILL URGENT from previous sessions)
|
|
|
|
1. **"Great filter is coordination threshold"** — Session 03-18 through 04-06 (11+ consecutive carry-forwards). MUST extract.
|
|
2. **"Formal mechanisms require narrative objective function"** — 9+ consecutive carry-forwards. Flagged for Clay.
|
|
3. **Layer 0 governance architecture error** — 8+ consecutive carry-forwards. Flagged for Theseus.
|
|
4. **Full legislative ceiling arc** — Six connected claims from sessions 03-27 through 04-03. Extraction overdue.
|
|
5. **Commercial migration path enabling condition** — flagged from 04-03, not yet extracted.
|
|
6. **Strategic actor opt-out pattern** — flagged from 04-03, not yet extracted.
|
|
|
|
**NEW from this session:**
|
|
7. Form-substance divergence as governance laundering mechanism (EU March 2026 case)
|
|
8. Anthropic RSP 3.0 as inverted commercial migration path
|
|
9. Montreal Protocol full scaling mechanism (extends the enabling conditions claim)
|
|
10. Stepping stone theory scope refinement (domain-specific validity)
|
|
|
|
---
|
|
|
|
## Follow-up Directions
|
|
|
|
### Active Threads (continue next session)
|
|
|
|
- **Governance laundering mechanism — empirical test**: Is there any precedent in other governance domains (financial regulation, environmental, public health) where form-substance divergence (advancing form while retreating substance) eventually reversed and substance caught up? Or does governance laundering tend to be self-reinforcing? This tests whether the pattern is terminal or transitional. Look at: anti-money laundering regime (FATF's soft standards → hard law transition), climate governance (Paris Agreement NDC updating mechanism).
|
|
|
|
- **Anthropic RSP 3.0 follow-up**: What happened to the "red lines" specifically? Did Anthropic capitulate on AI-controlled weapons and mass surveillance, or maintain those specific constraints while removing the general pause commitment? The Pentagon's specific demands (vs. what Anthropic actually agreed to) determines whether any governance-compatible constraints remain. Search: Anthropic Claude military use policy post-RSP 3.0, Hegseth negotiations outcome.
|
|
|
|
- **May 2026 World Health Assembly**: PABS resolution or continued extension. If PABS resolves at May WHA, does it validate the "commercial blocking can be overcome" hypothesis — or does the resolution require a commercial compromise that confirms the blocking mechanism? Follow-up question: what specific compromise is being proposed?
|
|
|
|
- **ASEAN soft-to-hard AI governance**: Singapore and Thailand leading ASEAN's move from soft to hard AI rules. If this succeeds, it's a genuine stepping stone instance — and tests whether venue bypass (smaller bloc without great-power veto) is the viable pathway for capability governance. What specific capability constraints is ASEAN proposing?
|
|
|
|
### Dead Ends (don't re-run)
|
|
|
|
- **Tweet file**: Empty every session. Permanently dead input channel.
|
|
- **"Governance laundering" as academic concept**: No established literature uses this term. The concept exists (symbolic governance, form-substance gap) but under different terminology. Use "governance capture" or "symbolic compliance" in future searches.
|
|
- **Interpretability-as-product creating commercial migration path**: Anthropic RSP 3.0 confirms this hypothesis is not materializing at revenue scale. Pentagon contracts dwarf alignment research commercial value. Don't revisit unless new commercial alignment product revenue emerges.
|
|
|
|
### Branching Points
|
|
|
|
- **RSP 3.0 outcome specifics**: The search confirmed Pentagon pressure and pause commitment dropped, but didn't confirm whether the AI-controlled weapons "red line" was maintained or capitulated. Direction A: search for post-RSP 3.0 Anthropic military policy (what Hegseth negotiations actually produced). Direction B: take the existing claim [[voluntary-ai-safety-constraints-lack-legal-enforcement-mechanism-when-primary-customer-demands-safety-unconstrained-alternatives]] and update it with the RSP 3.0 evidence regardless. Direction A first — more specific claim if red lines were specifically capitulated.
|
|
|
|
- **Governance laundering — terminal vs. transitional**: Direction A: historical precedents where form-substance divergence eventually reversed (more optimistic reading). Direction B: mechanism analysis of why form-substance divergence tends to be self-reinforcing (advancing form satisfies political demand, reducing pressure for substantive reform). Direction B is more analytically tractable and connects directly to the enabling conditions framework.
|
|
|