> ## Content Index
> Fetch the complete content index at: https://www.aigl.blog/llms.txt
> Use this file to discover other available public pages before exploring further.

# AI Governance Scorecard and Safety Standards Policy
- URL: https://www.aigl.blog/ai-governance-scorecard-and-safety-standards-policy/
- Published: 2026-09-18T18:59:12.000Z
- Updated: 2026-09-18T18:59:12.000Z
- Description: A Future of Life Institute policy document that scores AI governance proposals against safety and regulatory criteria. It proposes a four-level safety framework with registration, audits, approvals, risk bounds and provable-safety requirements.
- Author: Jakub Szarmach
- Tags: Framework, AI Governance, AI Safety, Risk Management, Compliance & Audit, Technical Controls, #aigl-library

## ⚡ Quick Summary

Published by Future of Life Institute (FLI), this policy document compares AI governance proposals and presents FLI’s Safety Standards Policy (SSP), a proposed regulatory framework for AI safety standards, measures and oversight. It argues that voluntary guidelines are insufficient under competitive pressures and distinguishes proposals with binding safety standards from those relying on industry self-regulation or voluntary commitments. Its scorecard assesses proposals against concrete requirements including registration, third-party audits, developer burden of proof, quantitative risk bounds, liability, compute limits, coverage of open-source models, large language models and military AI, and international coordination.

The SSP sets four AI Safety Levels (ASL-1 to ASL-4), each triggered by hardware or capabilities. ASL-2 requires registration and national-authority audits; ASL-3, triggered at 100 yottaFLOP (10²⁶) or specified capabilities, requires pre-approval and quantitative safety bounds; and ASL-4, triggered at ronnaFLOP (10²⁷) or AGI potential, requires IAIA approval and formal verification of provable safety. FLI also advocates an immediate pause on systems that do not meet the proposed standards while they are enacted.

## 🧩 What’s Covered

The document moves from a comparative scorecard to its proposed safety regime and its rationale.

- **Introduction and policy premise**: Presents risks associated with increasingly capable AI, including disinformation, deepfakes, bias, destabilisation of labour markets and political institutions, concentration of power, and assistance with chemical, biological and cyber weapons. It calls for lawmakers to require demonstrable compliance with safety standards, drawing comparisons with the FDA, FAA and NRC.
- **AI Governance Scorecard**: Table 1 compares named AI governance proposals and non-AI regulatory examples. It marks whether each proposal contains safety requirements, registration, independent audits, a developer burden of proof, quantitative risk bounds, liability, compute limits, non-exemptions for open source, LLMs and military AI, an international body, and no call for human replacement.
- **Binding standards distinction**: Explains that the scorecard concentrates on concrete, enforceable requirements. It contrasts FLI’s SSP with Anthropic’s Responsible Scaling Policy, noting that both use four safety levels but that the SSP calls for an immediate pause where standards are not met.
- **Four ASL classifications**: Table 2 defines ASL-1 as negligible catastrophic risk and ASL-2 as potential minor harm; it gives examples including a 2018 LLM, a chess-playing system, and unreliable bioweapon instructions. ASL-3 addresses potential major harm and low-level autonomous capabilities, while ASL-4 covers systems with AGI potential.
- **Training and deployment requirements**: Specifies registration with a national authority for ASL-2; pre-approval of safety plans for ASL-3; and IAIA pre-approval for ASL-4\. Deployment requirements range from whitebox and blackbox red-teaming to approved quantitative risk assessments and formal verification covering cybersecurity, controllability, kill-switches, alignment and robustness to malicious use.
- **Clarifications and rationale**: Defines what registration, audits, liability, risk quantification and compute triggers entail. It proposes international coordination through an IAIA and UN Security Council approval for ASL-4 in the interim, then argues that shortening projected AGI timelines make regulation urgent.
- **Commercial AI and AGI pursuit**: Separates currently valuable commercial uses from the pursuit of AGI and superintelligence. It argues that the benefits associated with commercial AI do not require AGI pursuit, which it places under ASL-4.

## 💡 Why it matters?

The framework gives policymakers, AI labs and assurance functions a concrete way to connect model capability thresholds to pre-training and deployment controls. Rather than stopping at general commitments, it specifies registrations, safety plans, audits, liability, quantitative risk assessments and approval responsibilities. This can support assessment of whether a proposal places the onus on developers to show safety before training or release.

It also connects model safety and cybersecurity: audits include preventing unauthorised model access as well as whitebox and blackbox model-safety evaluations. For higher-risk systems, the document combines national oversight with a proposed international coordination body and describes an interim role for the UN Security Council.

## ❓ What’s Missing

The document acknowledges that its framework is “by no means exhaustive” and will require further specification. Although it calls for quantitative safety bounds, it does not set the authorised numerical levels for major harms. It requires formal verification at ASL-4 but does not provide a procedure for performing or independently evaluating that verification. The scorecard supplies criteria, definitions and checkmarks for proposals, but does not set out a detailed scoring methodology or evidence base for every classification. Operational detail is also limited on how national authorities would be established, how their approvals would be coordinated, or how the proposed IAIA would be constituted. The compute thresholds may be updated by the IAIA, but no process for making those updates is specified.

## 👥 Best For

Policy teams and regulators designing binding AI safety rules, as well as AI-lab governance and compliance leads assessing pre-training approval, audit and incident-reporting expectations. It is particularly relevant to teams mapping capability- or compute-based thresholds to safeguards for systems that may create major harm or enable AGI.

## 📄 Source Details

*AI Governance Scorecard and Safety Standards Policy* is an English-language Future of Life Institute document dated 30th October 2023 and marked “Last update: 6th November 2023.” No individual authors are credited; Dr. Peter Park is thanked for research underlying the scorecard. The supplied complete PDF contains six pages. No URL to the PDF is printed in the document.