The twelve leverage points from Meadows’ Thinking in Systems, applied to AI governance. Ranked from weakest (#12) to most powerful (#1). Full treatment in Part 2: What Can I Do About AI?
Low leverage — parameters, buffers, structure (#12–10)
#12 Constants & Parameters — Content filters, rate limits, safety thresholds, tax rates. Real engineering work; the system absorbs changes and the underlying loops continue. Most “AI safety announcements” live here. Useful signal: what a lab announces at #12 tells you what tier they’re actually operating at.
#11 Buffer Sizes — Mandatory review windows, safety team headcount ratios, oversight budget relative to industry scale. The US AISI’s pre-cut budget vs. the industry it was meant to oversee is a buffer asymmetry in real time. When review periods get shortened under competitive pressure, a buffer is being drained.
#10 Physical Structure of Stocks & Flows — Chip fabrication networks, data center geography, power grid architecture. Very hard to change; constrains everything that flows through it. The electricity grid is already acting as a structural brake — not by intent, but because interconnection queues average 7+ years in key markets.
Medium leverage — delays, feedback strength, loop gain (#9–7)
#9 Length of Delays — The lag between cause and effect. Capability deploys on 2–4 month cycles; harms are documented months to years later; regulatory response takes years. Aviation’s near-miss reporting system (ASRS) is the model for shortening this: mandatory, anonymous, fast. An AI equivalent doesn’t exist.
#8 Strength of Balancing Feedback Loops — How vigorously the oversight loop counteracts unchecked growth. Currently: underfunded regulators, voluntary self-reporting, no mandatory audits. One appellate liability ruling would strengthen the legal balancing loop more than most regulatory rulemaking.
#7 Gain Around Reinforcing Feedback Loops — How fast the capability-capital-compute loop amplifies. Taxing AI compute at the point of training runs, or requiring profit-sharing contributions from AI-related gains, would reduce loop gain. Resisted by everyone at the center of the loop.
High leverage — information, rules, who writes them (#6–4)
#6 Structure of Information Flows — Who knows what, when. Making something visible that wasn’t visible changes behavior without changing any rules. When electricity meters moved from basements to front porches, consumption fell ~30% with no other policy change. Mandatory AI incident reporting, public capability evaluations, required training data disclosures — these are the AI equivalents. Labs currently self-select what gets reported.
#5 Rules of the System — Incentives, punishments, deployment gates. Rules with enforcement behind them are categorically different from voluntary commitments. The EU AI Act penalty structure (€35M or 7% of global turnover) activates August 2026 — the first #5 lever with structural enforcement at scale. Product liability case law is the other active front.
#4 Power to Add, Change, or Self-Organize Structure — Who gets to write the rules. When the regulated write their own oversight framework, the framework protects the regulated. The US AI Action Plan (July 2025) was developed with significant lab involvement. Academic independence in AI safety research is a #4 asset.
Transformative — goals, paradigm, coordination (#3–1)
#3 Goals of the System — What the system actually produces, not what it says it’s for. POSIWID: the purpose of a system is what it does. The revealed goal of current AI deployment — in deployment pattern after deployment pattern — is replacement and market capture, not broad benefit distribution. Financial liability, profit-sharing law, and labor protection change actual goals faster than mission language.
#2 Mindset / Paradigm — The deep assumptions nobody questions because nobody sees them. “Efficiency gains are inherently good” and “the race to AGI is inevitable” are paradigm-level claims that function as facts in the current system. They’re choices. Paradigms shift through accumulated evidence and visible counterexample — not through argument alone.
#1 The Coordination Problem — At civilizational scale, leverage means building the conditions for coordination between actors with misaligned incentives. The Montreal Protocol worked because scientists made the threat legible before it was politically convenient, and gave actors a face-saving exit. Those three conditions — legible shared threat, manageable actor set, face-saving exit — don’t yet exist for AI. They can be built.
Framework: Donella H. Meadows, Thinking in Systems (Chelsea Green, 2008), pp. 145–165.