A Boring Blueprint for AI Governance

Context ~ In his Substack post "On Effective Altruism" (October 2026), written in response to The Economist's cover story calling effective altruism "this century's biggest idea," Adam Thierer argues that much effective-altruist thinking about AI treats morality as an optimization problem and AI safety as a matter for top-down control by enlightened elites. Drawing on Hayek's "pretence of knowledge," Thomas Sowell's "vision of the anointed," and Aaron Wildavsky's Searching for Safety, he contends that safety emerges from decentralized experimentation, feedback, and learning from failure rather than from anticipatory controls, and that speculative extinction probabilities ("p(doom)") have no validated method behind them. He proposes instead what he calls a "boring blueprint" of pragmatic, incremental measures, most of which are already in place and being continually refined. Acknowledged Risks ~ Thierer does not dismiss AI risks. He concedes that effective altruists are right that some AI-enabled capabilities pose enormous risks, some already materializing, and that certain applications will require more safeguards, testing, liability rules, security practices, and targeted policies. Stakeholder Authority ~ A question the blueprint raises is who imposes its safeguards, and by what authority, if not a central regulator. One answer is the stakeholders of AI services: both those who use them and those who may be affected by them. Users can impose conditions through purchasing, contracting, insurance, and investment decisions. Affected non-users, who have no purchasing power to withhold, rely mainly on liability for actual harms and on enforcement of existing consumer-protection and product-safety laws, which the blueprint explicitly endorses. Published commitments and performance reports strengthen both forms of leverage by supplying the evidence needed to show whether duties have been met. Limitation ~ Stakeholder leverage binds most tightly where stakeholders are concentrated, such as large purchasers, insurers, and investors, and least where they are diffuse. The developers of the most consequential systems may be the least dependent on any single stakeholder's goodwill. How decentralized arrangements can enforce cooperation, as distinct from coordinating those already willing to cooperate, remains an open question. Symmetry ~ The evidentiary discipline Thierer calls for applies in both directions. If no validated method supports high estimates of catastrophic risk, none supports confidence that such risk is negligible. Likewise, the commitments of developers who treat safety as an engineering problem, such as Nvidia CEO Jensen Huang's statement that a product should not be released unless its maker is confident in its safety, can and should be made verifiable in the same way as the commitments of those who emphasize catastrophic risk. Verifiable Virtue ~ This plan is complementary to Verifiable Virtue (https://stratml.us/docs/VV4AI.xml), which seeks accountability for AI developers that does not depend on trusting anyone's stated good intentions. That plan's goals for independent verification, epistemic diversity, and public performance reporting would put into practice several elements of this blueprint, including independent testing and certification, incident reporting, and continuous refinement through trial and error, while subjecting risk estimates to stated assumptions, adversarial review, and tracking over time. This plan has been inferred and rendered in StratML format by Claude.ai in dialog with Owen Ambur focusing on Adam Thierer's Substack post: https://adamthierer.substack.com/p/on-effective-altruism It can be edited in the form at https://stratml.us/forms/Claude/Part1.html