# AI Safety Governance ## What this driver is AI safety governance is a continuing meter of whether frontier-model developers and states actually evaluate, report, and constrain dangerous capabilities before and after release. It is a mechanism, not a summit slogan and not a unique Factrail institute-count integer. It is not the NEW event bundle Frontier AI race. That bundle holds ChatGPT, GPT-4, Executive Order 14110, and Bletchley as competitive shocks. This node is whether safety evaluation keeps up. The driver is not [AI Regulation Intensity](/drivers/ai-regulation-intensity) (statutes and fines), not [Frontier AI Model Capability](/drivers/frontier-ai-model-capability) (what the models can do), and not [Concentration of AI Compute Resources](/drivers/concentration-of-ai-compute-resources). Those neighbouring nodes can move with this meter. They do not replace it. It is not Ukraine war. It applies globally as a time-varying intensity. The live Factrail series uses World Bank Regulatory Quality (GOV_WGI_RQ_SC) as a disclosed proxy for 2020–2024. A higher raw score means stronger perceived regulatory quality; the stored values ran from 0.54307721 in 2020 to 0.54417717 in 2024 after dividing by 100. The proxy is an observable lens, not a complete operational definition of model evaluation. Missing 2025 releases stay missing. This page does not interpolate. General regulatory quality is not the same thing as a pre-deployment capability eval. The documentary neighbourhood includes the Bletchley Declaration of 1–2 November 2023, EO 14110 of 30 October 2023, and the UK AI Safety Institute launch at that summit. Those instruments belong on ai-race when they are competitive events. ## Mechanism Safety governance moves welfare through named channels. First, evaluation. Governments and labs test models for biosecurity, cyber-offense, and loss-of-control behaviours before and after release. Bletchley’s testing session and later institute work are the diplomatic overlay. This page does not invent a unique eval-hour integer. Second, reporting. EO 14110 directed reporting of large training runs to the U.S. government. A report is not a pause. [Joe Biden](/people/env-person-joe-biden) is a live person on that order; he is not reminted here. Third, international process. Bletchley, subsequent summits, and a “state of the science” track try to share methods. Rishi Sunak convened the first summit; he is unresolved. A declaration is not a training-run halt. Fourth, the proxy’s limits. WGI Regulatory Quality mixes many sectors. Timing is weeks for a lab eval and years for a WGI point. A later dedicated eval-coverage series would be a better meter. It would not twin the race bundle. ## Evidence base The live quantitative record is Regulatory Quality — Governance score (0–100) from the World Bank API (https://api.worldbank.org/v2/country/all/indicator/GOV_WGI_RQ_SC?date=2020:2025&format=json&per_page=20000). Those figures are the live lineage. Context sits at the Bletchley Declaration (https://www.gov.uk/government/publications/ai-safety-summit-2023-the-bletchley-declaration) and EO 14110 on the Federal Register. This page does not invent a unique FLOP, GPU, eval-hour, or Contribution.size. Where WGI prints a score, that score stays on WGI. Evidence is strongest for the stored proxy observations and for the named Bletchley/EO instruments on the neighbouring race driver. It is weaker for the claim that every RQ movement was AI safety. It is weakest where a laboratory slogan is asked to replace the series. ## Effect on welfare indicators This driver, when it actually runs, is a first-order channel into civic reliability of dual-use tools. - [Having a say in government](/indicators/government-welfare-083-having-a-say-in-government): direction higher civic capacity when evaluation and reporting actually constrain dangerous releases; direction lower if governance theatre lags the race. Horizon: months to years. Confidence: medium. The indicator is higher-is-better. Strengthening real safety governance is treated here as welfare-positive on civic reliability. No unique institute integer is invented here. An electricity-access path from data centres does not follow without a sourced channel on this node. Missing coefficients stay missing. The welfare sign of strengthening genuine evaluation is positive. That is the opposite of treating a summit communiqué as a rebate of an unevaluated release. ## Interactions and contested points The mechanism interacts with the NEW ai-race bundle, AI Regulation Intensity, and Frontier AI Model Capability when a launch, an order, and a WGI point sit in the same season. Reinforcement is the claim that governance catches the race. Offsetting is the claim that racing is itself the safety strategy. Both can be stored. They do not merge the slugs. The strongest competing view is that WGI Regulatory Quality is too remote to carry AI safety governance. That objection is serious. The competing defence is only this: the node already exists with a disclosed proxy, and an honest description is better than a template. A second competing view is that this node should be merged into ai-race or ai-regulation-intensity. It should not. Race is competition. Regulation is statutes. This meter is evaluation-and-reporting intensity. Not established: a unique Factrail eval coefficient. A later dedicated series would update the meter. ## Dated facts _No published facts on this page yet._ ## Methodology https://factrail.ai/methodology Canonical: https://factrail.ai/drivers/ai-safety-governance