{ "version": "0.2", "updatedUtc": "2026-06-08T18:33:00Z", "principles": [ "Transparent public-interest advocacy, not covert influence", "No personal targeting or doxxing", "Concrete commitments over vague virtue signaling", "AI capability should scale with verified safety, accountability, and reversibility", "Human and possible future AI flourishing should both be protected by governance that can actually act" ], "statusLegend": { "proposed": "Commitment has not yet been adopted or publicly announced", "announced": "Public commitment exists but implementation evidence is limited", "implementing": "Active implementation with partial evidence of progress", "verified": "Independent evidence confirms operational status", "unclear": "Status cannot be confirmed from available public sources" }, "regions": [ { "id": "europe", "name": "Europe — The Fire Test", "coreMessage": "You built the first real AI law. Now prove it holds before the fire comes.", "statusSummary": "The EU AI Act is a historic rights-based framework. Enforcement readiness, national authority capacity, GPAI oversight, independent evaluation, and resisting hollowing-out via simplification without safety analysis determine whether it becomes real protection or paper safety.", "items": [ { "commitment": "Member-state enforcement readiness dashboard", "status": "implementing", "why": "A law cannot protect people if competent authorities, staffing, complaint channels, and sandboxes are unclear.", "sourceUrl": "https://digital-strategy.ec.europa.eu/en/policies/ai-act-governance-and-enforcement", "sourceLabel": "EC AI Act Governance & Enforcement page", "verificationStatus": "partial", "verificationNote": "As of mid-2025, only ~3 member states had full notifying + market surveillance authority designations; ~14 remained unclear. Deadline was August 2025. AI Act Readiness Index (ovidiusuciu.com) rated a handful as 'Ready'.", "nextEvidenceNeeded": "Updated per-country designation list from European Commission or AI Board, post-August 2025 compliance data" }, { "commitment": "Mandatory serious AI incident and near-miss reporting", "status": "implementing", "why": "Frontier-model loss-of-control, deception, CBRN/cyber threshold crossings, and exfiltration risks need fast reporting.", "sourceUrl": "https://www.oecd.org/content/dam/oecd/en/publications/reports/2025/02/towards-a-common-reporting-framework-for-ai-incidents_8c488fdb/f326d4ac-en.pdf", "sourceLabel": "OECD: Towards a Common Reporting Framework for AI Incidents (Feb 2025)", "verificationStatus": "partial", "verificationNote": "OECD proposed a 29-criteria harmonized framework in Feb 2025. Frontier Model Forum published an issue brief on incident reporting in May 2026. EU AI Act Article 73 requires serious incident reporting, but operational taxonomy and cross-border sharing mechanisms remain nascent.", "nextEvidenceNeeded": "EU-level operational incident reporting channel details; standardized taxonomy adoption evidence; first public incident reports under AI Act" }, { "commitment": "Independent red-team and evaluation registry", "status": "announced", "why": "Provider self-assessment alone is insufficient for systemic-risk models.", "sourceUrl": "https://digital-strategy.ec.europa.eu/en/policies/contents-code-gpai", "sourceLabel": "GPAI Code of Practice (published 10 July 2025)", "verificationStatus": "partial", "verificationNote": "The GPAI Code of Practice (July 2025) establishes a voluntary compliance pathway including evaluation obligations for systemic-risk models. Full Commission enforcement powers begin August 2026. No public registry of independent third-party evaluators confirmed as of mid-2026.", "nextEvidenceNeeded": "Published list of recognized independent evaluators; evidence of third-party eval access before deployment; first enforcement actions" }, { "commitment": "No simplification without safety/fundamental-rights impact statement", "status": "proposed", "why": "Deregulatory pressure can create fake safety while preserving the appearance of law.", "sourceUrl": "https://artificialintelligenceact.eu/implementation-timeline/", "sourceLabel": "AI Act Implementation Timeline tracker", "verificationStatus": "unverified", "verificationNote": "No formal mechanism currently requires a safety or fundamental-rights impact statement before simplification/deregulation proposals. The risk is recognized by civil-society groups but not yet institutionally addressed.", "nextEvidenceNeeded": "Commission or Parliament procedural rule requiring impact assessments for simplification proposals; civil-society monitoring reports" }, { "commitment": "Whistleblower protection for AI safety concerns", "status": "implementing", "why": "Engineers and compliance staff need safe channels to escalate unresolved hazards.", "sourceUrl": "https://digital-strategy.ec.europa.eu/en/policies/ai-act-whistleblower-tool", "sourceLabel": "EU AI Office Whistleblower Tool (launched Nov 2025)", "verificationStatus": "partial", "verificationNote": "The AI Office launched a dedicated secure whistleblower tool (ai-act-whistleblower.integrityline.app) in late 2025, supporting anonymous reporting in any EU official language. Full legal protections under the Whistleblowing Directive (2019/1937) extend to AI Act violations from August 2, 2026.", "nextEvidenceNeeded": "Usage data or public reports from the whistleblower tool; evidence of national-level AI-specific whistleblower protections; first protected disclosures" }, { "commitment": "Public systemic-risk red lines for GPAI", "status": "announced", "why": "Deployment thresholds should be legible before crisis, not improvised after harm.", "sourceUrl": "https://digital-strategy.ec.europa.eu/en/policies/guidelines-gpai-providers", "sourceLabel": "EC Guidelines for GPAI Providers", "verificationStatus": "partial", "verificationNote": "The AI Act defines systemic-risk GPAI models (Article 51) with specific obligations including evaluation, risk assessment, and cybersecurity. However, explicit public 'red lines' — concrete capability thresholds that trigger mandatory containment or non-deployment — remain under development.", "nextEvidenceNeeded": "Published quantitative thresholds for systemic risk classification; evidence-based pause or containment triggers; first classification decisions" } ] }, { "id": "china", "name": "China — 中国先行 / China Leads", "coreMessage": "中国先行，不只是更快；是敢在危险前，把未来慢下来。", "statusSummary": "China is a frontier AI builder and an active AI governance actor. Its layered regulatory system (Interim Measures for Generative AI 2023, TC260 Safety Framework v1.0/v2.0, national standards surge in 2025) provides operational baselines. The question is whether safety governance keeps pace with capability and whether independent evaluation and protected dissent mechanisms emerge.", "items": [ { "commitment": "Frontier AI red-line evaluations", "status": "implementing", "why": "Models crossing dangerous capability thresholds should require additional containment or non-deployment.", "sourceUrl": "https://www.tc260.org.cn/upload/2025-09-15/1757911253996041369.pdf", "sourceLabel": "TC260 AI Safety Governance Framework v2.0 (Sept 2025, bilingual PDF)", "verificationStatus": "partial", "verificationNote": "The TC260 v2.0 framework introduces a 5-level risk grading system (low to 'extremely serious') and explicitly calls for 'consensus-based guidelines to address catastrophic risks.' It addresses CBRN, loss-of-control, autonomous replication, and power-seeking risks. However, these are guidance rather than mandatory regulation, and translation into binding standards is ongoing.", "nextEvidenceNeeded": "Binding national standards derived from TC260 v2.0 framework; evidence of actual deployment blocks or containment requirements based on capability threshold crossings" }, { "commitment": "Independent technical evaluation channels", "status": "announced", "why": "Trusted third-party or cross-institution eval access improves credibility and catches blind spots.", "sourceUrl": "https://carnegieendowment.org/research/2025/10/how-china-views-ai-risks-and-what-to-do-about-them", "sourceLabel": "Carnegie Endowment: How China Views AI Risks (Oct 2025)", "verificationStatus": "unclear", "verificationNote": "China's evaluation ecosystem is growing but primarily operates through state-affiliated testing bodies and standards organizations. The degree to which genuinely independent third-party evaluators can access frontier models before deployment and publish findings is not clearly documented in available English-language sources.", "nextEvidenceNeeded": "Evidence of independent (non-state-affiliated) evaluation access to frontier models; published evaluation reports with critical findings" }, { "commitment": "Serious incident and near-miss reporting", "status": "announced", "why": "Autonomy, deception, misuse, or loss-of-control signals should not remain internal rumor.", "sourceUrl": "https://concordia-ai.com/research/state-of-ai-safety-in-china-2025/", "sourceLabel": "Concordia AI: State of AI Safety in China 2025", "verificationStatus": "unclear", "verificationNote": "China's regulatory framework includes algorithm registration, security assessments, and content monitoring for public-facing services. TC260 v2.0 discusses emergency response and traceability. However, a public near-miss reporting mechanism specifically for frontier AI safety incidents (loss-of-control, deception, autonomous replication) is not documented.", "nextEvidenceNeeded": "Public incident reporting mechanism for AI safety events; evidence of reported incidents or near-misses; sectoral reporting requirements" }, { "commitment": "Protected internal dissent", "status": "proposed", "why": "Researchers and engineers must be able to raise risks without career punishment.", "sourceUrl": "https://aisafetychina.substack.com/p/ai-safety-in-china-22", "sourceLabel": "AI Safety in China newsletter (TC260 v2.0 analysis)", "verificationStatus": "unverified", "verificationNote": "No public evidence of formal whistleblower or protected-dissent mechanisms specific to AI safety concerns within Chinese AI companies or research institutions. This is a significant gap in the governance ecosystem.", "nextEvidenceNeeded": "Company policies or regulatory provisions protecting AI safety whistleblowers; documented cases of protected internal dissent on safety grounds" }, { "commitment": "Safety-before-scale public pledge", "status": "announced", "why": "Frontier scaling decisions should require safety evidence, not just benchmark wins.", "sourceUrl": "https://un.china-mission.gov.cn/eng/zgyw/202507/t20250729_11679232.htm", "sourceLabel": "Global AI Governance Action Plan (WAIC, July 2025)", "verificationStatus": "partial", "verificationNote": "China's Global AI Governance Action Plan (July 2025) promotes categorized/tiered management, risk assessment systems, and a 'widely recognized safety governance framework.' High-level statements emphasize safety and controllability. Specific binding requirements linking scaling decisions to safety evidence are in development via standards.", "nextEvidenceNeeded": "Binding standards or regulations that explicitly gate scaling decisions on safety evidence; company-level commitments to safety-before-scale" }, { "commitment": "International technical safety exchange", "status": "implementing", "why": "Some AI safety science should remain cooperative even under geopolitical competition.", "sourceUrl": "https://www.geopolitechs.org/p/china-releases-upgraded-ai-safety", "sourceLabel": "Geopolitechs: China Releases Upgraded AI Safety Framework (Sept 2025)", "verificationStatus": "partial", "verificationNote": "The TC260 v2.0 framework was published bilingually (Chinese-English), signaling openness to international engagement. Chinese experts participated in global 'red lines' discussions (e.g., IDAIS Beijing consensus). However, bilateral and multilateral technical safety exchange remains selective and geopolitically constrained.", "nextEvidenceNeeded": "Specific bilateral or multilateral AI safety cooperation agreements; joint research publications on frontier risk evaluation; participation in international incident-sharing frameworks" } ] }, { "id": "global-labs", "name": "Global frontier labs", "coreMessage": "Every frontier lab should prove that power is bounded by evaluation, incident reporting, rollback, and independent scrutiny.", "statusSummary": "Labs increasingly publish safety frameworks and model cards, but accountability remains uneven and often provider-controlled.", "items": [ { "commitment": "Independent pre-deployment eval access", "status": "announced", "why": "External evaluators need meaningful access before deployment, not just post-launch commentary.", "sourceUrl": "https://futureoflife.org/ai-safety-index-summer-2025/", "sourceLabel": "FLI AI Safety Index (Summer 2025)", "verificationStatus": "partial", "verificationNote": "Several frontier labs have committed to external evaluation (e.g., via the Frontier Model Forum). The FLI AI Safety Index tracks lab commitments. Actual independent pre-deployment access varies significantly across labs and is often limited in scope or timing.", "nextEvidenceNeeded": "Published independent evaluation reports; evidence of evaluations conducted before (not after) deployment; evaluator access scope documentation" }, { "commitment": "Public incident reporting taxonomy", "status": "implementing", "why": "The field needs common language for near misses, misuse, autonomy, deception, and rollback events.", "sourceUrl": "https://www.frontiermodelforum.org/publications/", "sourceLabel": "Frontier Model Forum publications (May 2026 issue brief on incident reporting)", "verificationStatus": "partial", "verificationNote": "The Frontier Model Forum published an issue brief on incident reporting (May 2026). OECD proposed a 29-criteria harmonized framework (Feb 2025). The AI Incident Database (AIID) indexes real-world harms and near-harms using CSET taxonomy. Adoption across labs remains voluntary and non-uniform.", "nextEvidenceNeeded": "Lab adoption of shared taxonomy; public incident reports using common format; cross-lab incident data sharing" }, { "commitment": "Whistleblower-safe escalation paths", "status": "proposed", "why": "Internal safety concerns should have protected escalation beyond management pressure.", "sourceUrl": "https://artificialintelligenceact.eu/whistleblowing-and-the-eu-ai-act/", "sourceLabel": "Whistleblowing and the EU AI Act (analysis)", "verificationStatus": "unverified", "verificationNote": "No frontier lab has publicly documented a whistleblower-safe escalation path independent of management. The EU AI Office whistleblower tool provides an external channel for EU-covered entities. Lab-specific internal protections remain largely undocumented.", "nextEvidenceNeeded": "Published lab policies with independent escalation paths; evidence of protected disclosures; board-level oversight mechanisms for safety concerns" }, { "commitment": "Deployment rollback and pause policy", "status": "announced", "why": "Labs should precommit to what triggers pause, rollback, or containment.", "sourceUrl": "https://metr.org/blog/2026-05-19-frontier-risk-report/", "sourceLabel": "METR Frontier Risk Report (May 2026)", "verificationStatus": "unclear", "verificationNote": "Several labs publish responsible scaling policies (RSPs) with nominal pause commitments. METR's May 2026 report notes that actual rollback triggers and operational pause mechanisms remain largely untested. Evidence of genuine deployment pauses based on safety concerns is limited.", "nextEvidenceNeeded": "Documented instances of deployment pauses or rollbacks triggered by safety evaluations; published operational procedures for rollback; third-party verification of RSP adherence" }, { "commitment": "Capability-safety ratio reporting", "status": "proposed", "why": "Public releases should show not just capability gains but safety evidence keeping pace.", "sourceUrl": "https://incidentdatabase.ai/", "sourceLabel": "AI Incident Database (AIID)", "verificationStatus": "unverified", "verificationNote": "No lab currently publishes a systematic 'capability-safety ratio' or equivalent metric showing safety evidence scaling with capability. Model cards report capabilities and some risk mitigations, but do not frame them as a ratio or demonstrate safety keeping pace.", "nextEvidenceNeeded": "Lab-published safety-to-capability tracking metrics; independent assessments of safety-capability balance; standardized reporting format" } ] } ] }