Section 3 of 4
Strong reinforcing loops have been contained before. The structural sequence is consistent across cases: harm was made legible, financial cost was attached to specific behavior, and rules were written by actors other than the industry being regulated. None of it was fast.
Part 1 documented what happens when balancing loops fail to keep pace with reinforcing loops. This section grounds the other side: what happened when they succeeded. AI is not the same as tobacco, aviation, or nuclear power — the capability development loop is faster, the harm signals are harder to attribute, and the actor set is more fragmented. But the loop dynamics are similar enough that the structural lessons transfer directly to the leverage tier choices that follow.
Aviation: length of delays in practice
Commercial aviation in the 1950s–60s was averaging roughly one fatal accident per week. The industry’s loop (more flights → more revenue → more routes → more flights) had no built-in brake for safety. What changed it was not a ban or a boycott — it was the National Transportation Safety Board’s Aviation Safety Reporting System (ASRS, launched 1975), which shortened the delay between when a dangerous situation occurred and when the industry could learn from it. Anonymous voluntary reporting — protecting reporters from punishment — created the information flow that mandatory top-down reporting couldn’t. The loop was not stopped. Its feedback structure was changed. Within two decades, commercial aviation became statistically the safest form of mass transportation in history.
The AI equivalent — a mandatory, standardized incident reporting system covering model failures, near-misses, and alignment gaps — does not exist. Aviation didn’t build its system until after several catastrophic failures created the political will. The structural question for AI is whether that will arrives before or after an equivalent threshold event.
Tobacco: enforceable rules closing the financial loop
The tobacco industry operated its reinforcing loop for decades after harm was scientifically documented. Awareness campaigns ran from the 1960s. Consumer behavior shifted. The loop continued. What eventually closed it was enforceable rules operating through a specific mechanism: state attorneys general filed coordinated suits treating tobacco harm as a public health cost the state was absorbing and had standing to recover. The 1998 Master Settlement Agreement — $206 billion over 25 years plus advertising restrictions — permanently changed the financial structure of what recklessness cost the industry.
The mechanism that worked was not consumer pressure (tried; insufficient). Not public information campaigns (tried; insufficient). It was making the externalized harm financially recoverable by actors with the standing and resources to pursue it. For AI, the equivalent mechanism is product liability — and it is currently in early formation. Garcia v. Character.AI reached a product liability framework in federal court before settling in January 2026 without establishing precedent. The mechanism is real. The precedent is not yet set.
Nuclear nonproliferation: coordination with rivals at the table
The Nuclear Non-Proliferation Treaty (1968) was signed by states in active geopolitical competition with every incentive to defect. What made coordination possible was a specific combination of conditions: the threat was legible (nuclear winter scenarios were scientifically documented and publicly communicable), the actor set was small enough to negotiate with, and signatories had a face-saving path to restraint. The NPT is imperfect — India, Pakistan, and Israel never signed; North Korea withdrew. But it demonstrably slowed proliferation below what uncoordinated development would have produced.
The AI coordination problem has harder structure: more actors, less legible risk (alignment failure doesn’t have a visible signature the way a nuclear test does), and no established face-saving framework. The Bletchley–Seoul–Paris–New Delhi summit chain shows states can sit at the same table. The conditions for binding coordination haven’t been created yet. Whether they can be depends significantly on whether interpretability research matures enough to make AI risk as legible as nuclear risk became — the same role nuclear winter modeling played in making the NPT politically viable.
The pattern
Across cases, the structural sequence is consistent:
- Harm was documented and made publicly legible — not predicted, documented
- Financial or reputational cost attached to specific behavior by actors with legal standing to impose it
- Rules were written by actors outside the industry being regulated
- The goal of the system shifted from “maximize without constraint” to “maximize within constraint”
None of these transitions were fast. Tobacco took forty years from documented harm to the MSA. Aviation took thirty years from systematic accidents to a functioning reporting system. Nuclear coordination required US–Soviet agreement during the Cold War and still has significant gaps.
The pace problem
AI capability is compressing the timeline available for the same structural transitions. The question isn’t whether the playbook works — it does, across sufficiently different domains that the structural logic appears robust. The question is whether the sequenced steps can happen fast enough. Some of the windows are narrowing specifically because the fast loop (R1) is running much faster than the fast loops in any of these predecessor cases. The timing dimension is mapped in section 4.