Andrew Clearwater is an AI governance practitioner who writes analyses on his Substack andrewclearwater.substack.com, aimed at what he describes as "people actually running AI governance programs." His writing addresses the operational reality of compliance practitioners at enterprises, connecting AI policy debates to day-to-day governance work. A recurring argument across his work is that voluntary standards are becoming the de facto US AI legal infrastructure.
Positions and frameworks
Clearwater's writing returns to several frameworks, most of which are tracked on dedicated concept pages.
Standards as the de facto governance layer. Clearwater argues that US states are pursuing three approaches to standards-based AI governance at once: an incentive approach (Texas TRAIGA: follow NIST to obtain a safe harbor), a mandate approach (Illinois SB 3312, Utah HB 286, Washington HB 2157: standards as obligation), and a transparency approach (California TFAIA, NY: standards as accountability). He describes this as a three-lane standards-based governance arrangement (see Three-Lane Standards-Based AI Governance).
He further argues that courts are already using the NIST AI RMF to define the standard of care in negligence and strict-liability cases, regardless of statutory adoption. On his account, decades of product-liability precedent mean that compliance with widely recognized standards becomes evidence of good faith, while non-adoption becomes evidence of negligence (see Standards as Litigation Evidence).
In a related line of argument, Clearwater contends that NIST, ISO, IEEE, and CEN-CENELEC were never designed as quasi-legislative bodies but that standards-based governance is making them so, a dynamic he labels the standards-industrial complex. He points to the EU's CEN-CENELEC fast-track process, which he says was needed because only 15 of the EU AI Act's harmonized standards were published by late 2025, with half projected to miss the August 2026 deadline; in his view, the fast-track approach undercut the consensus legitimacy that makes standards worth following (see Standards-Industrial Complex).
Procurement-driven governance. Clearwater argues that California's EO N-5-26 identified a legal seam in the Trump administration's December 2025 federal preemption order, which carves out state procurement from the scope of preemption. He predicts that procurement-driven governance will become the operating model for US AI policy for the foreseeable future (see California Executive Order N-5-26 (Trusted AI Procurement) and Procurement-Driven AI Governance).
System cards and model-level due diligence. Clearwater argues that model system cards are among the most important AI governance documents, on the grounds that they contain admissions that would not appear in marketing materials; he cites Anthropic's 200+ page Claude Opus 4.6 card and 244-page Mythos Preview card as examples. He provides a structured prompt template for LLM-assisted system-card analysis (see System Card Due Diligence).
He also describes what he calls a defensive AI paradox: the argument that the only way to protect against advanced AI is with equally advanced AI, which on his account leaves organizations and governments without frontier-model access at a structural disadvantage (see Defensive AI Paradox). Relatedly, he argues that Anthropic's two-question alignment risk update framework — intent (will the model attempt a harmful action?) and monitoring/security (will the attempt succeed despite mitigations?) — should become standard practice for every company deploying AI agents (see Alignment Risk Update).
Works
Clearwater authored a cluster of Substack pieces ingested in April 2026:
- NIST Just Told Us What's Actually Broken in AI Governance (Clearwater, March 2026) (Mar 10, 2026) — analysis of NIST AI 800-4
- Standards Are the New Legislation (Clearwater, March 2026) (Mar 24, 2026) — three-lane framework
- California Just Played the One AI Card the Feds Can't Block (Clearwater, March 2026) (Mar 31, 2026) — procurement-as-legal-seam
- System Cards Are the Most Important AI Governance Document You're Not Reading Carefully Enough (Clearwater, February 2026) (Feb 19, 2026) — Opus 4.6 system-card walkthrough
- You Need the Model to Fight the Model: What 300 Pages of Anthropic's Brutally Honest Safety Docs Reveal (Clearwater, April 2026) (Apr 8, 2026) — Mythos / alignment risk update
- Colorado AI Act repeal/replace analysis (Mar 17, 2026) — folded into Colorado AI Act (SB 24-205)
Relationships
- supports: NIST AI Risk Management Framework 1.0, ISO/IEC 42001 — AI Management System (advocates these as the operational substrate of US AI governance).
- depends-on: Claude Opus 4.6, Claude Mythos Preview (his system-card analyses are based on these).
- related: Government AI Procurement (he extends this concept to a full operating model).
- contradicts: the federal-preemption framing in EO — Trump Federal Preemption of State AI Laws (Dec 11, 2025) — argues the procurement carve-out makes preemption an incomplete ceiling.