Selected reading.
Curated reading on pre-execution AI governance, AI safety frameworks, and the operational reality of compliance — combining original essays with selected analysis from regulators, research institutes and frontier labs.
01 · Articles & analysis
19 items · May 2025 – April 2026 · external links open in a new tab
Responsible Scaling Policy Version 3.0
Anthropic ·
Anthropic replaces the categorical training halt with a dual trigger combining race-leadership and material catastrophic risk, and embeds external review authority for the Long-Term Benefit Trust.
Assessing the state of AI adoption across the federal government
Brookings ·
More than 85% of US federal AI use cases declared 'high-impact' in 2025 have incomplete disclosures on pre-deployment testing, impact assessments or monitoring — despite explicit OMB requirements.
12 Graphs That Explain the State of AI in 2026
IEEE Spectrum ·
Visualisation of the Stanford AI Index 2026: rising incidents, falling transparency scores, and a widening governance gap between model capabilities and institutional oversight.
The 2026 AI Index Report
Stanford HAI ·
Annual report documents AI regulatory activity across 47 countries, a 55% rise in documented AI incidents, and an 18-point drop in the Foundation Model Transparency Index.
The HAIP Reporting Framework: What works, what's next
Brookings ·
Analysis of the first reporting cycle of the G7 Hiroshima AI Process Reporting Framework identifies gaps in comparability and verification of voluntary risk disclosures by leading AI developers.
Next-generation Constitutional Classifiers: More efficient protection against universal jailbreaks
Anthropic ·
Updated Constitutional Classifiers architecture reduces inference overhead through representation re-use while maintaining the block rate against universal jailbreaks across thousands of red-team hours.
The Part of an AI System We Never See — But That Determines Everything
Medium ·
The structural layer beneath every AI action — and why governance starts there, not at the API surface.
Beyond Compliance: Why AI Needs Clarity Before It Acts
Medium ·
Compliance reports describe past behaviour. Clarity decides future behaviour. The distinction matters operationally.
The Layer Above All Layers: Why Autonomy Requires a Governance Substrate Before It Requires Capability
Medium ·
Capability without a governance substrate compounds risk faster than it compounds usefulness.
Why Autonomous Systems Need a Preconditions Logic — And Why We've Ignored It Until Now
Medium ·
Autonomy without an explicit preconditions logic is a missing safety primitive. The case for making it first-class.
Before Retrieval Comes Permission: The Missing Layer in Modern AI Systems
Medium ·
Retrieval-Augmented Generation skips a step. Permission-checking belongs above retrieval, not after it.
Before AI Acts: A Technical Prerequisite for Trustworthy Autonomy
Medium ·
A technical prerequisite for trustworthy autonomy: every request passes a deterministic gate before the system acts on it.
AISI Frontier AI Trends Report (2025)
UK AI Security Institute ·
First public AISI evaluation of more than 30 frontier models shows cyber-task length doubling roughly every eight months and first expert-level successes on autonomous tasks.
The 2025 Foundation Model Transparency Index
Stanford CRFM ·
Third edition of the index shows average transparency scores falling from 58 to 41 points, with the steepest drops in training-data and environmental-impact disclosures.
International AI Safety Report 2025 — First Key Update: Capabilities and Risk Implications
International AI Safety Report (Bengio et al.) ·
Update to the international Bengio-led safety report assesses how reasoning training and inference-time enhancements shift biothreat, cyber, and controllability risks since the January 2025 report.
Strengthening our Frontier Safety Framework
Google DeepMind ·
Version 3.0 of the Frontier Safety Framework introduces a Critical Capability Level for 'harmful manipulation' and extends evaluation to scenarios where models might undermine human shutdown or control.
Guidelines on the scope of obligations for providers of general-purpose AI models under the AI Act
European Commission · AI Office ·
Official Commission interpretive guidelines on the scope of GPAI obligations taking effect 2 August 2025, including the 10^25-FLOPs threshold for systemic-risk classification.
The General-Purpose AI Code of Practice
European Commission · AI Office ·
Final Code of Practice with three chapters (Transparency, Copyright, Safety & Security) as the voluntary compliance path for GPAI providers under the AI Act.
The EU's AI Power Play: Between Deregulation and Innovation
Carnegie Endowment ·
Argues that Europe's real AI-competitiveness barriers are fragmented markets, under-funded VC structures and cloud dependence — not the AI Act itself.
02 · About this list
Articles are listed strictly by date, newest first — no separation between own writing and external sources. Selection criteria for external pieces: substantive content (not press releases), credible sources (regulators, standards bodies, peer-reviewed research, leading think tanks, frontier-lab safety publications), and relevance to operational AI governance.
Original essays by the Noetik founder are also available on Medium: medium.com/@mikefarrell478
For company enquiries: Contact · for technical depth: Capabilities.