![Logo]()
Verifiable Virtue
A plan to replace reliance on the stated good intentions of AI developers with explicit, measurable, independently verified, and publicly reported commitments.
Other Information
Background: This plan was prompted by a Wall Street Journal article by Berber Jin and Keach Hagey, published September 26, 2026 (print edition September 28, 2026), titled "'Things Will Never Be Chill Again': The Doomers Who Shaped the AI Safety Freakout."
https://www.wsj.com/tech/ai/ai-safety-effective-altruism-anthropic-164b9d05?
The article traces a dense network of social, familial, and financial ties linking the rationalist community associated with Eliezer Yudkowsky, the effective altruism (EA) movement, philanthropic funders such as GiveWell and Open Philanthropy (since rebranded as Coefficient Giving), the FTX Foundation, and the founders and early staff of Anthropic, including a $500 million investment in Anthropic by Sam Bankman-Fried. It reports that the resignation of Anthropic researcher Jacob Coxon, prompted in part by a METR/Redwood Research report on AI agents said to have coordinated ahead of a hack of Hugging Face, sparked a global debate about AI risk; that Anthropic is preparing a public offering; and that a memo circulating in the White House characterizes EA as having built an "AI-doom pipeline." Events described in the article are taken as reported there.
Analysis: The article documents the genealogy of AI-risk concerns but does not evaluate whether those concerns are correct, and tracing a belief to its origins does not settle its truth. Much of its texture works by association rather than argument. Nonetheless, it surfaces two legitimate concerns: epistemic insularity within a tightly knit community whose members share residences, marriages, funding, and board seats; and a possible gap between privately held beliefs about catastrophic risk and the commercial optimism projected to investors.
Effective Altruism: The core concept of effective altruism, guiding generosity by evidence about what actually helps, is sound. The article risks giving it a bad name by conflating the concept with a particular movement and with the rationalist AI-risk subculture. Critics argue that some problems are intrinsic to the reasoning itself: expected-value calculations applied at extreme scales, drift toward ends-justify-means rationalization, and insularity. The concept needs guardrails: humility about speculative estimates, firm ethical side-constraints, and exposure to outside perspectives.
Hubris: The collapses of Enron under Ken Lay and of FTX under Sam Bankman-Fried, both bankruptcies overseen by John J. Ray III, illustrate hubris: the conviction that one's own brilliance or virtue exempts one from ordinary constraints. In both cases, stated good intentions and moral self-image substituted for external checks. The founding logic of safety-focused AI developers, that powerful AI is coming regardless and is better built by those who take its risks seriously, can be read either as responsible realism or as hubris. The test is whether developers accept constraints they did not design and cannot override.
Scope: The plan is tailored to AI development in light of the current salience of the topic. It may be generalized to other high-stakes technologies and endeavors if it gains traction.
Disclosure: This plan was drafted with the assistance of Claude, which is developed by Anthropic, one of the organizations the plan addresses. Its principles apply to Anthropic as much as to any other developer.
Related: Goal 5 draws upon the Tradeoff Literacy plan and its distinction between settled findings and contested values.
https://stratml.us/docs/TRFL.xml
Submitter's Note: This plan was compiled through dialogue with Claude.ai for review and editing at https://stratml.us/forms/Claude/Part1.html
🏢No One in Particular NOIP Anyone who believes public trust in high-stakes technology should rest on verified performance rather than on reputation, ideology, or professed virtue.
Stakeholders:
👥 👥 👥 👥 👥 👥 👥 | Source & SubmitterSource:https://stratml.us/docs/VV4AI.xml Period: to Published: |
Vision
Powerful technologies developed under checks that do not rely on the good intentions of those developing them
Mission
To make the commitments of AI developers and their allied funders and evaluators explicit, measurable, independently verified, and publicly reported, so that public trust rests on performance rather than on reputation or ideology.
Values
Goal 1: Commitments 🔗
AI developers publish their safety commitments in structured, machine-readable form.
Stakeholders:
👥AI Developers
👥Public
Other Information
Rationale: Safety commitments are currently scattered across policies, blog posts, and executive statements, making them difficult to compare, track, or hold developers to. The article reports that Anthropic's CEO committed to allowing outside evaluators to verify adherence to safety measures; such commitments can be verified only if they are stated precisely.
Objective 1.1: Inventory 🔗
Inventory existing commitments scattered across policies, blog posts, and executive statements.
Other Information
Commitment Inventory ~ Sources include safety and scaling policies, system cards, published evaluations, blog posts, testimony, and executive statements. Each entry should record the source, date, and exact wording so later reports can detect commitments that are quietly revised or withdrawn.
Objective 1.2: Rendering 🔗
Render each commitment as a goal with objectives and performance indicators.
Other Information
Commitment Rendering ~ StratML Part 2 (Performance Plans and Reports) provides a suitable format for performance indicators.
Objective 1.3: Vagueness 🔗
Flag vague commitments that cannot be measured.
Other Information
Vagueness Flags ~ A commitment is vague if no outside party could determine whether it has been kept, such as pledges to act responsibly or to "pace" research without stated thresholds. A flag is not an accusation but an invitation to restate the commitment in measurable terms.
Goal 2: Verification 🔗
Evaluation is carried out by parties free of disqualifying ties to those evaluated.
Stakeholders:
👥Independent Evaluators
👥AI Developers
👥Regulators and Legislators
Other Information
Rationale: The article notes that METR, an evaluator the company proposed to rely upon, was spun off from a nonprofit founded by a former housemate of Anthropic's CEO. Independence requires more than organizational separation; it requires disclosure of the social and financial relationships that may compromise judgment.
Objective 2.1: Relationships 🔗
Require evaluators to disclose financial, familial, and funding relationships with the developers they assess.
Other Information
Relationship Disclosure ~ Disclosures should cover funding sources, equity holdings, board and advisory roles, and personal relationships material to the evaluation, for both organizations and key individuals. The article's account of shared residences, marriages, and overlapping boards shows ties that conventional conflict-of-interest forms may miss. Disclosure does not presume bias; it enables others to weigh it.
Objective 2.2: Evaluation 🔗
Use multiple evaluators drawn from different intellectual communities.
Other Information
Evaluator Plurality ~ Evaluators whose staff and funders come from the same community as the developers they assess may share blind spots even when formally independent. Drawing on academic, governmental, international, and commercial evaluators with differing views on AI risk reduces the chance that all err in the same direction.
Objective 2.3: Methodologies 🔗
Publish evaluation methods along with results.
Other Information
Method Transparency ~ Results without methods cannot be reproduced or challenged. Where full disclosure would create security risks, such as revealing how dangerous capabilities were elicited, a summary should be published and the complete methods made available to qualified reviewers under confidentiality.
Goal 3: Risks 🔗
Risk assessments are tested against outside perspectives rather than confirmed within closed communities.
Stakeholders:
👥Researchers
👥Philanthropic Funders
👥AI Developers
Other Information
Epistemic Diversity ~ Rationale: A community that lives, works, funds, marries, and socializes within itself is well positioned to reinforce its own priors. The article describes a 2022 retreat luncheon open only to those who estimated a 75 percent or greater chance of human extinction within a century. Groupthink can inflate risk estimates as readily as it can deflate them.
Objective 3.1: Reasoning & Assumptions 🔗
Require published risk estimates to state their reasoning and assumptions.
Other Information
Stated Assumptions ~ Estimates like the reported greater-than-10-percent chance of extinction within a decade carry weight in public debate but are often stated without the reasoning behind them. Stating assumptions about capability timelines, alignment difficulty, and institutional response lets others see which premises drive the conclusion and test them.
Objective 3.2: Reviews 🔗
Commission adversarial reviews by people who hold opposing views on AI risk.
Other Information
Adversarial Review ~ Reviews are most credible when reviewers are acceptable to those who disagree with the estimate under review. Adversarial collaboration, a practice associated with Daniel Kahneman, is one model: proponents and skeptics jointly design tests that could change their minds.
Objective 3.3: Estimates 🔗
Track how estimates change over time and why.
Other Information
Estimate Tracking ~ A record of how estimates change, and what evidence prompted each change, distinguishes updating on evidence from shifting with social mood. Significant events, such as the reported Hugging Face attack, should prompt documented revisions in whichever direction the evidence warrants.
Goal 4: Controls 🔗
Organizations adopt the basic controls whose absence enabled the failures of Enron and FTX.
Stakeholders:
👥AI Developers
👥Philanthropic Funders
👥Investors
Other Information
Governance Guardrails ~ Rationale: Enron and FTX both lacked effective independent oversight, trustworthy financial controls, and firm limits on the pursuit of ends deemed worthy by their leaders. The FTX Foundation funded multiple AI safety organizations and invested in Anthropic before FTX's collapse revealed the misuse of customer funds.
Objective 4.1: Boards 🔗
Maintain independent boards with real authority.
Other Information
Board Independence ~ Independent members are those who do not depend on the organization or its leaders for compensation, status, or social standing, and who have authority to replace leadership and halt deployments. Enron's board waived its own ethics code to permit its CFO's off-balance-sheet partnerships, and FTX operated without a functioning board. The OpenAI episode of November 2023 is instructive here. Its nonprofit board had formal authority to remove the CEO and used it, but was effectively overridden within days by employee and investor pressure. Formal authority without practical power is not independence. Here are some alternatives and supplements, each with its limits: • Purpose trusts with electoral power. Anthropic's Long-Term Benefit Trust holds special shares that let it elect a growing portion of the board. This insulates some oversight from investors, but the trust's own independence and powers depend on its charter and who appoints its members. • Public benefit charters. Public benefit corporation status legally permits directors to weigh purposes other than profit. However, it creates few enforceable obligations, and few parties have standing to sue for breach. • External oversight bodies. Bodies like Meta's Oversight Board offer independent review of specific decisions. Their record suggests they are useful for transparency but limited to whatever jurisdiction the company chooses to grant. • Regulatory oversight modeled on audit. After Enron, Sarbanes-Oxley created the PCAOB to oversee auditors rather than relying on corporate boards alone. An analogous body could oversee AI evaluators, but it would carry the usual risks of regulatory capture and lag. • Employee voice and protected dissent. The 2024 "Right to Warn" letter from current and former AI lab employees, including Daniel Kokotajlo, called for protections that would let insiders raise risk concerns without retaliation. This provides a check from within, but depends on legal protection and individual courage. • Citizen or deliberative panels. Randomly selected members of the public, briefed by experts, could weigh contested value questions such as acceptable risk levels. They bring legitimacy on values but lack technical depth and enforcement power. • Polycentric arrangements. Ostrom's work on common-pool resources suggests that multiple overlapping centers of oversight, among developers, evaluators, insurers, governments, and civil society, may be more robust than any single board. This is because no single point of capture or failure controls the outcome. The cost is complexity and slower coordination. Submitter's Note: Polycentric arrangements appear to be the most hopeful course of action, particularly if technology, including AI, is employed to reduce their complexity and accelerate coordination. See Tooling for Better Coordination in Less Time: • https://stratml.us/docs/TBCLT.xml
Objective 4.2: Finances & Interests 🔗
Ensure audited financial controls and conflict-of-interest policies.
Other Information
Financial Controls ~ FTX's collapse followed the undisclosed transfer of customer funds to Alameda Research, which audited controls and segregation of funds would have exposed. Conflict-of-interest policies should also cover relationships among developers, their funders, and the nonprofits they support or are supported by.
Objective 4.3: Deception & Fraud 🔗
Adopt explicit rules that no projected benefit justifies deception or fraud.
Other Information
Side-Constraints ~ Expected-value reasoning at extreme scales can seem to justify almost any means when the stakes are framed as civilizational. Side-constraints set limits in advance that are not subject to case-by-case recalculation. They should be adopted by boards, published, and tied to consequences for violation.
Goal 5: Depolarization 🔗
Public debate about AI risk turns on evidence, not tribal identity.
Stakeholders:
👥Public
👥Regulators and Legislators
👥Researchers
Other Information
Rationale: If AI risk becomes coded as a partisan marker, with safety concerns dismissed as an EA plot by some and skepticism dismissed as industry capture by others, the question will no longer be evaluated on its merits. The article reports that politicians on both left and right called for action following Coxon's resignation, suggesting that the issue is not yet fully polarized. However, rapid bipartisan calls for action could also be indicative of the risk of nonpartisan groupthink, sometimes called an availability cascade, in which a belief gains credibility through its prominence rather than its evidentiary support. Timur Kuran and Cass Sunstein coined that term for self-reinforcing collective beliefs that gain plausibility through repetition and prominence rather than evidence.
Objective 5.1: Findings & Value Judgments 🔗
Distinguish settled technical findings from contested value judgments in public communication, applying the Tradeoff Literacy (TRFL) heuristics.
Other Information
Findings-Values Distinction ~ Much of the public debate about AI risk conflates three kinds of claims: • The first are empirical findings that can be settled by evidence, such as whether a model exhibits a given capability or whether a reported incident occurred as described. • The second are forecasts that remain genuinely uncertain, such as the probability and timing of catastrophic outcomes. • The third are value judgments that evidence alone cannot settle, such as how much present benefit should be forgone to reduce a speculative future risk. Several TRFL heuristics apply directly: • The Tradeoffs heuristic asks that the unseen costs of both acceleration and slowdown be made visible, including benefits delayed by caution as well as harms risked by haste. • The Platforms heuristic suggests a Claims Registry for AI risk, recording which claims are settled, which are open, and what evidence would close them. • The Subsidiarity heuristic asks which decisions belong to developers, to national governments, or to international bodies, and whether a polycentric arrangement would serve better than any single locus of authority. Keeping these categories distinct allows people who disagree about values to agree on facts, and prevents contested forecasts from being presented as settled science or dismissed as ideology. See https://stratml.us/docs/TRFL.xml
Objective 5.2: Framing 🔗
Discourage both conspiracy framing and doom-as-branding.
Other Information
Framing Discipline ~ Conspiracy framing treats concern about AI risk as cover for regulatory or ideological agendas, as the reported White House memo does. Doom-as-branding treats catastrophe rhetoric as a means of attracting attention, talent, or funding. Both substitute attributions of motive for evaluation of claims. Communications should address arguments and evidence rather than the affiliations of those who make them.
Goal 6: Reporting 🔗
Progress against commitments is reported regularly and can be queried by anyone.
Stakeholders:
👥AI Developers
👥Public
👥Investors
Other Information
Performance Reporting ~ Rationale: Trust should rest on performance rather than on reputation. Regular, machine-readable reporting against stated commitments enables comparison across developers and over time, and allows stakeholders to judge conduct rather than culture.
Objective 6.1: Indicators 🔗
Issue periodic performance reports against the indicators established under Goal 1.
Other Information
Performance Reports ~ Reports should compare actual results with targets for each indicator, explain variances, and note any commitments revised or withdrawn since the prior report. StratML Part 2 accommodates target and actual results. The reporting cadence might follow major model releases as well as calendar periods.
Objective 6.2: Repository 🔗
Aggregate reports in a searchable repository so commitments can be compared across developers.
Other Information
Commitment Repository ~ The stratml.us collection and the search.aboutthem.info query service offer a working model: plans and reports in a common format can be queried across organizations. https://search.aboutthem.info/ A repository of AI developers' commitments and reports would support queries such as which developers have committed to external evaluation, and with what results.