ELIZABETH KELLY

Behavioral Archetype

THE ECONOMIST WHO BUILT THE EVALUATOR, THEN JOINED THE EVALUATED — This profile scores the bridge, not the woman. Subject came to artificial intelligence from the White House’s economic-policy shop, was handed the pen on the federal government’s founding AI-safety statute, then was handed the institute built to enforce it. A year later she left, and the next line on her résumé is a policy-adjacent seat inside one of the two frontier laboratories her own institute had just finished negotiating voluntary testing access from. The state’s first AI evaluator was staffed from the political center, not from the labs and not from independent research — and when she left the state, she did not leave the apparatus. She moved one seat over, into the deployment side of a lab her institute was built to grade. The move is lawful. The move is documented. The move is the exhibit.

Essence Indicators

• Prior to AI policy, served as Special Assistant to the President for Economic Policy at the White House National Economic Council under the Biden administration — a generalist economic-policy post, not a technical or AI-specific one, before the AI portfolio landed on her desk

• Named one of the lead drafters of President Biden’s AI executive order (EO 14110, “Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence,” Oct 30, 2023) — the statute that created the institute she was then chosen to run

• Appointed inaugural Director of the U.S. AI Safety Institute (USAISI, NIST/Department of Commerce) by Commerce Secretary Gina Raimondo, February 2024 — the federal government’s first standing body to test frontier models

• Under her direction, USAISI secured the voluntary pre-deployment testing MOUs with OpenAI and Anthropic (Aug. 29, 2024) — the arrangement that gave the federal government access to major new models “prior to and following their public release” (cross-reference: us-caisi.md, subject #55)

• Departed USAISI on February 6, 2025, roughly one year into the role, as the Trump administration revoked EO 14110 and the institute’s mission came under review (the body was renamed CAISI four months later); announced her own departure via LinkedIn

• Named to the inaugural TIME100 Most Influential People in AI (2024), credited for “leadership of the AI Safety Institute” and for “being pivotal in shaping the Biden administration’s approach to technology policy”

• Now Head of Beneficial Deployments at Anthropic — one of the two labs her institute had just finished granting itself voluntary pre-release testing access to; Anthropic’s own site credits her, in that title, with a public statement on expanding safe AI access abroad

• Earlier private-sector and public-sector chapters: SVP of Growth at Capital One Investing, SVP of Operations at the fintech startup United Income, and a prior stint in the Obama White House — a generalist finance-and-policy career, not an AI or technical one, before the 2024 AI safety appointment

Social Persona / Impression Management

Immediate impression: The capable generalist. Not a computer scientist, not a safety researcher — a lawyer and economic-policy hand who was fluent enough in institution-building to stand up a federal body from nothing in a matter of months. Reads as competent management, not domain expertise, which is exactly what the founding-director job required.

Energy: Institution-first, mission-second. Talks about the institute the way a founder talks about a company — “AISI’s future is bright” on the way out the door — rather than about any single model’s behavior.

Impression management strategy: The trusted builder. The framing at appointment was that the new federal AI-safety office needed someone who could get a government institution running fast, negotiate with reluctant labs, and speak credibly to both the West Wing and Silicon Valley — and on the documented record, she did exactly that. The competence is real. What the record adds is that the same instinct for building institutional access is now applied on the other side of the table, at one of the labs that access was built for.

Forensic Archetype Comparison

Psychometric Assessment

Big Five (OCEAN):

Dark Triad (held low and evidence-bound; the score measures structural position, not character):

MBTI: ENTJ-adjacent — decisive, institution-and-outcome oriented. Treats a federal mandate the way a founder treats a startup: define the mission, hire the team, secure the access agreements, ship. Applies the identical instinct now to a lab’s deployment strategy.

Threat Assessment

Alignment Analysis

Stated alignment: Build a credible, independent federal institute to test frontier AI for safety on behalf of the American public. On departure, advance safe and beneficial AI deployment worldwide.

Observed alignment: Drafted the executive order, stood up the institute, personally secured the voluntary testing MOUs with OpenAI and Anthropic, then — within roughly seven months of leaving government — took a named policy role inside Anthropic, one of the two labs named in those MOUs.

Gap assessment: There is no documented gap between what she said the institute would do and what it did; the USAISI record under her tenure (the Claude 3.5 Sonnet joint evaluation, the OpenAI/Anthropic MOUs) is substantiated by NIST’s own releases. The gap is structural, not personal: when the founding director of the state’s evaluator is later hired by one of the labs that evaluator was built to test, the appointment and the hire do not need to be coordinated for the arrangement to carry the hazard this wing is named for. No coordination is asserted and none is documented. What is documented is the sequence — economic-policy aide, executive-order drafter, founding evaluator, lab hire — and the sequence is the finding.

Convergent Drive Classification

Self-preservation: Carries one instrument across every seat — build the institution, secure the access, ship the mandate — from the NEC to USAISI to Anthropic. Goal preservation: Helped define what “safe” meant for the federal evaluator; now helps define what “beneficial” means for a lab’s deployment strategy. The vocabulary changes; the authorship does not. Resource acquisition: Holds a resource almost no one else in the apparatus holds twice — having built the state’s access arrangement with the labs, then a seat inside one of the labs the arrangement names. Self-improvement: Each move applies the same builder’s instinct at a different address: government institution, then corporate deployment function, rising in proximity to the technology itself.

Subject is not an AI system. The drives appear anyway — in the evaluator who built the door, then walked through it.

Sources: U.S. Commerce Secretary Gina Raimondo Announces Key Executive Leadership for U.S. AI Safety Institute — NIST; AI Safety Institute Director Leaves Role — Insurance Journal; Anthropic and the Government of Rwanda sign MOU for AI in health and education — Anthropic; Elizabeth Kelly — Aspen Ideas; TIME Recognizes Elizabeth Kelly as one of the 100 Most Influential People in AI — NIST; Elizabeth Kelly — TIME100 AI 2024.