AI Governance Interventions Timeline

Data table page — This page feeds the AI Governance and PolicyCruxAI Governance and PolicyComprehensive analysis of AI governance mechanisms estimating 30-50% probability of meaningful regulation by 2027 and 5-25% x-risk reduction potential through coordinated international approaches. ...Quality: 66/100 analysis. For conceptual background on how governance strategy interacts with AGI timelines, see Short AI Timeline Policy ImplicationsAnalysisShort AI Timeline Policy ImplicationsAnalyzes how AI policy priorities shift under 1-5 year timelines to transformative AI, arguing that interventions requiring less than 2 years (lab safety practices, compute monitoring, emergency co...Quality: 62/100.

Key Links

Overview

This page maintains a chronological log of major AI governance interventions from 2022 through mid-2026, intended as a data resource for the AI Governance Effectiveness Analysis. Entries cover a wide range of intervention types: executive orders and presidential directives, domestic legislation (enacted and failed), international agreements and summit declarations, voluntary industry commitments, export controls, AI safety institute formations, and technical standards bodies. Each row includes an assessment field coded as effective, partial, ineffective, or too early to tell, based on available evidence of implementation and observable outcomes.

The period covered is unusually dense. The EU AI Act moved from political agreement (December 2023) through adoption (July 2024) into phased implementation across 2025–2030. The US oscillated between the safety-oriented Executive Order 14110 (October 2023) and its revocation under the deregulatory AI Action Plan (July 2025). Internationally, a sequence of frontier AI safety summits at Bletchley (November 2023), Seoul (May 2024), and Paris (February 2025) produced escalating but largely non-binding commitments. Export controls on advanced AI chips, first imposed in October 2022, were tightened in successive rounds through 2023 and 2025, making them among the more consequential and enforceable interventions in the dataset.

Assessments are provisional. Many interventions are either too recent to evaluate or lack independent monitoring mechanisms. Where disagreement exists in the literature about whether an intervention has had its intended effect, the assessment defaults to partial or too early to tell. The methodology section below describes coding decisions in more detail.

Methodology

Unit of analysis. Each row represents a discrete policy instrument: a signed executive order, an enacted statute, a formally adopted international agreement, a published framework or standard, or a documented voluntary commitment. Draft proposals and white papers without formal adoption are excluded unless they generated observable downstream effects.

Assessment coding. The four assessment codes are defined as follows:

• Effective — the intervention was implemented substantially as intended and evidence suggests it achieved at least its proximate stated goal.
• Partial — the intervention was implemented but with significant gaps, delays, or scope limitations that reduced its effectiveness relative to stated goals.
• Ineffective — the intervention was not implemented, was rescinded before taking effect, or evidence suggests it failed to achieve its proximate goals.
• Too early to tell — the intervention has been adopted but insufficient time or evidence exists to assess outcomes.

Scope. Interventions are included if they meet one or more of the following criteria: they were adopted by a jurisdiction with significant AI development activity (US, EU member states, China, UK, Singapore, India); they operated at an international or multilateral level; or they were assessed in the AI safety literature as potentially significant for catastrophic risk reduction.

Source limitations. The research base for this table is strongest for US and EU interventions and weakest for China, ASEAN, and Africa. Post-2025 entries carry higher uncertainty. Assessments for interventions still in phased rollout (notably the EU AI Act) may be updated as implementation progresses.

Intervention Log

The table below contains 44 rows covering the period October 2022 through mid-2026. Rows are ordered chronologically by the date the intervention entered into force or was formally adopted.

Notable Patterns and Cross-Cutting Observations

Several patterns emerge across the intervention log that are relevant to the broader AI governance effectiveness analysis.

The enforcement gap. The majority of interventions coded as partial share a common structural feature: they establish substantive requirements but lack independent monitoring or verification mechanisms. The White House Voluntary Commitments (row 4), the Seoul Frontier AI Safety Commitments (row 20), and the Singapore Model AI Governance Framework (row 19) all fall into this category. Voluntary frameworks without third-party audit produce compliance that is difficult to distinguish from performative adoption. This is noted in the broader literature as a central limitation of the governance landscape prior to the EU AI Act's binding phase.

Export controls as outlier. The October 2022 semiconductor export controls (row 1) and their 2023–2024 updates (row 23) represent the most enforcement-capable interventions in the dataset. They operate through the existing US export licensing system, carry criminal penalties, and have demonstrated observable effects on chip acquisition patterns. The assessments for these rows nonetheless remain partial because documented circumvention through third-country re-export, cloud-access workarounds, and redesigned products limits their effectiveness relative to stated goals. Compute GovernanceConceptCompute GovernanceThis is a comprehensive overview of U.S. AI chip export controls policy, documenting the evolution from blanket restrictions to case-by-case licensing while highlighting significant enforcement cha...Quality: 58/100 and Hardware-Enabled GovernanceApproachHardware-Enabled GovernanceRAND analysis identifies attestation-based licensing as most feasible hardware-enabled governance mechanism with 5-10 year timeline, while 100,000+ export-controlled GPUs were smuggled to China in ...Quality: 70/100 literature treats these controls as a significant proof-of-concept for technical AI governance, while noting the ongoing cat-and-mouse dynamic.

The US oscillation. Rows 6, 15, and 29 trace a sharp reversal in US federal AI governance posture. Executive Order 14110 (October 2023) established the most comprehensive federal AI safety requirements to date, including mandatory reporting of large-model safety tests to the government. Its revocation via Executive Order 14179 (January 2025) before most implementing rules had been finalized means the intended outcome was never realized, producing an ineffective assessment on safety grounds — though EO 14179 was effective as a statement of the Trump administration's deregulatory intent. The AI Action Plan (row 36) that followed delegated safety responsibility to the private sector and conditioned federal funding on states not enacting AI restrictions, a posture assessed by critics at Harvard's ethics program as prioritizing economic competition over harm prevention.

The Bletchley–Seoul–Paris sequence. The three frontier AI safety summits produced escalating formality: Bletchley (2023) generated a political declaration; Seoul (2024) produced company-level safety commitments from 16 frontier labs; Paris (2025) built on these with further governmental statements. All three are assessed as partial rather than effective because they generated no binding obligations and relied on companies self-reporting implementation. The summits were nonetheless significant in establishing a shared vocabulary around catastrophic and systemic AI risk and in providing political cover for the creation of national AI Safety Institutes in the UK and US.

EU AI Act as structural anchor. The EU AI Act (rows 9, 11, 17, 22, 28, 38, 44) is the most structurally significant governance intervention in the dataset. Its phased implementation creates binding legal obligations for market access in the EU, backed by penalties and a formal regulatory architecture. Its assessments remain too early to tell throughout because the binding enforcement phases have only recently begun. The proposed delay to the August 2026 high-risk deadline (row 44) illustrates the political pressure on implementation timelines and introduces uncertainty that the dataset tracks but cannot yet resolve.

State-level governance in the US. The failure of California SB 1047 (row 24) and the Senate rejection of the federal preemption moratorium (row 37) together illustrate the contested terrain of US AI governance. SB 1047 was the most ambitious attempt at compute-threshold–based safety requirements for frontier models in any US jurisdiction; its veto removed a potential model for state-level frontier AI governance. The moratorium's 99–1 Senate rejection — opposed by both conservative governors and civil society — preserved state authority but left a fragmented regulatory landscape with over 480 enacted state AI bills as of 2025. Texas TRAIGAPolicyTexas Responsible AI Governance Act (TRAIGA)TRAIGA represents a state-level AI regulation focused on intent-based liability for harmful AI practices rather than comprehensive safety requirements, creating enforcement mechanisms but avoiding ...Quality: 55/100 represents a continuing state-level effort in this space.

Interventions by Type: Summary Count

Criticisms and Concerns

The governance landscape documented in this table draws significant criticism from multiple directions, which the effectiveness assessments attempt to reflect rather than adjudicate.

Voluntary frameworks are insufficient for catastrophic risks. A recurring critique in both the academic and EA/rationalist literatures is that the majority of governance interventions through 2025 remain voluntary, lack independent verification, and are therefore unlikely to constrain frontier AI development in ways that matter for catastrophic risk. The predominance of partial assessments for voluntary commitments in the table reflects this structural limitation. Critics have argued that relying on the private sector to self-report compliance — as both the White House Voluntary Commitments and Seoul Frontier Commitments do — repeats the pattern of social media self-regulation that produced documented harms without accountability.

Deregulatory US shift creates global governance gap. The Trump administration's January 2025 revocation of EO 14110 and the subsequent AI Action Plan have been characterized by governance researchers as creating a significant gap in the world's most AI-capable jurisdiction. Harvard ethics commentators described the plan as prioritizing economic competition over ethical safeguards and delegating risk management to the private sector without a statutory backstop. The US federal moratorium's failure (99–1 Senate vote) simultaneously prevented the kind of national coordination that could have replaced EO 14110 with a legislative equivalent.

AGI timeline dependency limits governance strategy. Research cited in the EA forum and LessWrong communities argues that the efficacy of governance interventions is deeply sensitive to AI TimelinesConceptAI TimelinesForecasts and debates about when transformative AI capabilities will be developedQuality: 95/100. Under pre-2030 timelines, broad regulatory and advocacy approaches may be too slow — awareness campaigns and legislative cycles typically take years to produce results. This suggests a mismatch between the timescale of most interventions in this dataset and the timescale on which transformative AI risks may materialize. Some researchers have argued for prioritizing corporate governance and security-focused interventions within existing frontier labs as faster-acting alternatives under short timelines.

Methodological concerns about timeline forecasting itself. A separate strand of criticism questions the value of timeline-dependent governance analysis, arguing that AGI timeline forecasting suffers from deference cycles (where a small number of influential forecasters tacitly inform a wider consensus, creating a false appearance of independent agreement) and poor operationalization of what AGI means. This limits confidence in the conditional assessments.

Implementation gaps even for binding frameworks. The EU AI Act's phased implementation, the proposed delay to the August 2026 high-risk deadline, and the significant variation in national authority capacity across EU member states illustrate that formal adoption does not guarantee effective implementation. The table captures this with too early to tell assessments for most EU AI Act milestones, reflecting genuine uncertainty rather than optimism.

Key Uncertainties

• Whether EU AI Act enforcement capacity will be sufficient to produce meaningful compliance by frontier model providers, particularly non-EU companies
• Whether the US federal AI governance vacuum following EO 14110's revocation will be filled by state legislation, industry self-governance, or eventual congressional action
• Whether export controls on advanced AI chips will remain effective as chip design, manufacturing, and cloud delivery continue to evolve
• Whether the frontier AI safety summit process (Bletchley → Seoul → Paris) will produce binding mechanisms or remain in the voluntary commitment paradigm
• Whether the Council of Europe AI Convention will attract ratification from major AI-developing states or remain limited in scope
• How governance effectiveness assessments for 2025–2026 interventions should be updated as implementation evidence accumulates

Changelog

Sources