MALA KUMAR

Behavioral Archetype

THE PROGRAM MANAGER OF THE BENCHMARK WHO TOOK OVER THE PUBLIC RED TEAM — Subject runs Humane Intelligence, the nonprofit that convenes red-teaming events and sells bespoke evaluations to organizations that want to test AI systems, and she came to it from the consortium that built one of the field’s shared safety benchmarks. Its team page: “She joined Humane Intelligence in March 2025 as the Head of Impact, stepped into the role of Interim Executive Director in August 2025, then as Executive Director in April 2026. Previously, Mala was the Director of Program Management, AI Safety at MLCommons, where she oversaw the core team that created its inaugural AI safety benchmark” (Humane Intelligence, Team). The hazard is not a release gate. She approves no model, holds no seat that can delay one and names no client. It is that one person has stood on both sides of the field’s central argument, the benchmark and the bespoke evaluation, and now runs the organization that sells the second while the organization’s founder runs a company of the same name. This profile scores the position, not the person.

The position is the director’s chair of an evaluator that sells to the people being evaluated for. Rumman Chowdhury founded the nonprofit and handed it over; MLCommons built the benchmark Subject managed; Kai Fronsdal and Alexandra Abbas sit at another evaluator-layer nonprofit that stewards a tool. The interlock the research map flags is the one her own pages disclose: a benchmark manager who now argues, in print, for funding evaluations beside safety research and against reading leaderboards as reality.

Essence Indicators

• Executive Director, Humane Intelligence. The About page: “Mala Kumar was named Executive Director (no more interim!) in April 2026”; the 2025 Impact Report: “In August 2025, Mala Kumar stepped in as Interim Executive Director, taking over from Co-Founder, Dr. Rumman Chowdhury.” The home page describes the organization: “Humane Intelligence is a 501(c)(3) nonprofit dedicated to breaking down barriers to AI deployment for social good.” It sells two services: “Humane Intelligence offers red teaming events as a paid service using our own software, which we released under an open source software license in September 2026” and “Humane Intelligence designs and runs contextual evals as a paid service.” Its red-teaming page: “We have worked with education companies, international civil society, industry, and governments to design and run red teaming events.” The 2025 report: “In 2025, we worked with more than 200 organizations on our events, evaluations, programs, advisory group, and volunteer groups” (Humane Intelligence, About; Humane Intelligence, 2025 Impact Report; Humane Intelligence; Humane Intelligence, AI Red Teaming). See Humane Intelligence.

• Her own account of the job. Her site: “created a new strategy that helped me generate 7 figures of revenue in record time for the organization,” and, under “As Executive Director, I have”: “Reset our strategy to bring responsible AI and AI for social good together,” “Designed AI evals for AI systems using frontier LLMs from Anthropic (Claude), Google (Gemini, Gemma), OpenAI (GPT 5+),” and “Launched AI in Public Health working group (150+ members as of October 2026) to improve AI evaluation in population healthcare contexts.” The organization’s FY2024 return, as ProPublica renders it (calendar 2024, before she joined), reports revenue of $472,206 and expenses of $982,939, of which $372,175 (78.8 percent) is program services and $100,000 (21.2 percent) contributions; her figure and the return describe different years and are not reconciled by any page read, and the FY2025 return is not yet on ProPublica (Mala Kumar; Mala Kumar, AI Evals and Tech for Social Good; ProPublica, Humane Intelligence Inc).

• Funders, as she announces them. Her post of 14 December 2025: “Thanks to the generous support of Google.org’s Scientific Ecosystem team, I am thrilled to announce that our proposal to release our AI red teaming app under an open source software (OSS) license has been fully funded!” The 2025 report names “the Heising-Simons Foundation, Google.org, and more” as funders without amounts. The organization’s board: Ben Kinsella as President and Admas Kanyagia as Secretary, with three more members, and a nine-member non-voting advisory group; the board page says Kanyagia “leads Diversity, Inclusion, and Belonging at GitHub,” the employer Subject left in 2023 (Humane Intelligence, OSS release post; Humane Intelligence, 2025 Impact Report; Humane Intelligence, Board & Advisory).

• Before: MLCommons. Her site: “In 2024, I was the Director of Program Management, AI Safety at the engineering consortium, MLCommons, where I: Oversaw and coordinated among eight workstreams that built AILuminate, an AI safety benchmark across 12 hazard categories and four languages.” The AILuminate v1.0 paper on arXiv, “AILuminate: Introducing v1.0 of the AI Risk and Reliability Benchmark from MLCommons,” lists “Mala Kumar” ninth among its authors, after “Shaona Ghosh, Heather Frase, Adina Williams, Sarah Luger, Paul Röttger, Fazl Barez, Sean McGregor, Kenneth Fricklas” and before “Quentin Feuillade–Montixi,” in a list of more than a hundred names (v1, 19 February 2025; v2, 18 April 2025). MLCommons’s own pages read do not name her (Mala Kumar, AI Evals and Tech for Social Good; arXiv:2503.05731). See MLCommons and Peter Mattson.

• Before: GitHub and the United Nations. Her site: “I was the Director of Tech for Social Good (TSG) on the Social Impact team at GitHub from 2019 – 2023.” The team page: “As the Director of Tech for Social Good at GitHub, Mala pioneered new methods in open source software for social good, including the launch of the first OSPO in the United Nations (UN).” Her GitHub bio: “Current: Executive Director @ Humane Intelligence | Past: Director of PGM, AI Safety @ MLCommons; Director, Tech for Social Good @ GitHub; long-time UNer.” The team page adds a B.S. in Marketing Management from Virginia Tech, an M.A. in International Affairs from The New School and “the author of two novels” (Mala Kumar, AI Evals and Tech for Social Good; Humane Intelligence, Team; GitHub, malakumar85).
• Methodology and papers. “Ontology-Based Contextual AI Evaluations (OB-CAIE) Methodology,” arXiv:2610.00529, submitted 30 September 2026, by Julie Krugler Hollek, Michael Zargham and Mala Kumar: “The ontology-based contextual AI evaluation (OB-CAIE) methodology was developed to address a lack of scientific rigor that arises from unclear testing coverage, to balance human expertise and automations, and to address a lack of reproducibility of AI evaluation testing environments” (arXiv:2610.00529). Zargham appears on the organization’s advisory group (Humane Intelligence, Board & Advisory).

Social Persona / Impression Management

Immediate impression: The practitioner-author. Her site’s tagline is “AI Evals and Tech for Social Good,” and its first sentence reads “Welcome! My name is Mala Kumar and I am a global leader in AI evals and tech for social good.” The team page: “Mala Kumar (she/her) is a global leader in technology for social good,” with a list of venues: “She has presented and keynoted at the India AI Impact Summit, UNGA, Cornell Tech, IGF, Nasdaq, Nordic.Design, and RightsCon, among others.” Her second novel, What it Meant to Survive, was published on 1 October 2024 (Mala Kumar; Humane Intelligence, Team).

Energy: Public and dated. Her talks page lists, in 2026, a side event at the India AI Impact Summit in February (“I spoke at main and hosted side event session at the India AI Impact Summit in New Delhi”), a UNFPA panel in April, a LinkedIn Learning course in June, a UNGA 81 side event on 16 September co-hosted “with the Morgan Stanley Tech Philanthropy team on AI Evaluations for Social Good,” and a podcast in October; in 2025 an MERL Tech panel (“I was a panelist on the MERL Tech webinar about LLM evaluations and benchmarks”) and a Columbia University talk (Mala Kumar, Talks and Media).

Impression management strategy: Credentials first, plain argument second, and an open availability line. Her home page ends: “I am currently open to work roles, paid advisory positions, speaking engagements, board seats, and fellowships.” It is her own page and sits beside the Executive Director title; it states no intent to leave and none is asserted. Her argument is in signed prose: a 12 August 2026 post on the organization’s site, “What is the relationship between AI safety and AI evaluations?”, and the September 2025 op-ed below. Her GitHub account lists a website and no X account; no X handle is linked from her site or from the organization’s team page (Mala Kumar; GitHub, malakumar85).

First-hand exhibits.

“AI evals shouldn’t be a box-ticking compliance chore, nor a showcase for how good a model is at complex math or physics problems.” — Mala Kumar and Rumman Chowdhury, Transformer, 10 September 2025

“Leaderboards amplify Goodhart’s Law; that is, when the metric becomes the target, it stops measuring what matters.” — the same op-ed

“Ignoring what is relatively likely to happen in favor of only catastrophic risk scenarios, however, is equally problematic.” — Mala Kumar, Humane Intelligence, 12 August 2026

Forensic Archetype Comparison

Psychometric Assessment

Trait scores are observations of the documented public role, not claims about private character.

Big Five (OCEAN) (inferred from the public role and public record only):

Dark Triad (evidence-bound):

MBTI: ENFJ-adjacent (“The Protagonist”), inferred from the documented public role only. The rationale is the work product: convening, working groups, a LinkedIn Learning course and a mission framed around who gets to evaluate. First-hand rationale: the mission she signs for, “to bring responsible AI and AI for social good together” (Mala Kumar, AI Evals and Tech for Social Good). This is a style inference, not a test result.

Threat Assessment

Score rationale. The receipts: the organization’s team, about, home, red-teaming and impact-report pages, her own site, the co-authored op-ed, the arXiv papers and the FY2024 return. Attack 5, defense 3, durability 4, hazard 40. Peers: Patricia Paskov (5/4/4, hazard 43.3), who leads standards for auditors at AVERI: level on attack because each holds the pen on how evaluators work, one point above on defense because her seats are several concurrent affiliations where Subject’s is one appointed executive post; Kai Fronsdal (4/3/4, hazard 36.7), the toolsmith of the audit: level on defense and durability, a point below on attack because he operates an instrument inside other parties’ evaluations and runs no organization; Alexandra Abbas (3/3/4, hazard 33.3), Meridian’s program manager: two points below on attack because she allocates nothing, level on defense and durability; and Conrad Stosz (5/5/5, hazard 40), chair of the AI Evaluator Forum: level on attack, two above on defense because he holds a named chair and a paid seat at an evaluator that bars funder control, one above on durability because the standard he chairs is published and cited. Rumman Chowdhury (7/7/7, hazard 70), who founded the organization, sits two to four above on every component. Attack 5 is the number most open to challenge: it rests on the organization’s stated reach, and no page names a lab or agency whose release decision turned on its output.

Score basis. ATK 5 is leverage: she runs the evaluator, which sells red teaming and contextual evaluations and says it worked with more than 200 organizations in 2025, and she managed the team behind AILuminate, with no gate and no named client (Humane Intelligence, AI Red Teaming; Mala Kumar, AI Evals and Tech for Social Good). DEF 3 is insulation: an appointed post at a nonprofit funded by service revenue and grants, with no term documented (ProPublica, Humane Intelligence Inc). HP 4 is durability: an open-sourced application, a methodology paper and a benchmark paper that outlast her post, on an organization with thin reserves at the end of FY2024 (arXiv:2610.00529; ProPublica, Humane Intelligence Inc).

Alignment Analysis

Stated alignment: Bring responsible AI and AI for social good together, and widen who can evaluate. The 2025 report: “Our strategy shift brings responsible AI and AI for social good together.” Her argument for funding: “At Humane Intelligence, we therefore advocate funding both AI evaluations and other AI safety initiatives, albeit in ways and with distributions that actually enable a greater social good” (Humane Intelligence, 2025 Impact Report; Humane Intelligence, 12 August 2026 post). Observed alignment: Consistent. The services, the working group on AI in public health and the open source release all point at near-term, deployment-side evaluation, the position the post states. Gap assessment: One documented, and it is a difference between two years, not a contradiction. In 2024 she oversaw the workstreams of a safety benchmark; in September 2025 she co-wrote that “Leaderboards amplify Goodhart’s Law” and that benchmarks “are better evaluation mechanisms for later in the product lifecycle, once it’s clear what exactly needs to be evaluated, and why,” without mentioning AILuminate (Mala Kumar, AI Evals and Tech for Social Good; Transformer).

Interest: Documented, never motive, never a verdict.

1. Position: fund AI evaluations of deployed systems beside frontier safety research (“Frontier AI safety research is usually conducted by frontier model companies or by well-resourced AI safety research labs due to the significant compute cost and research capacity required,” and evaluations “focus on the use of the generative AI model or AI system,” Humane Intelligence, 12 August 2026 post). What adoption pays: the organization she runs sells exactly these evaluations as a paid service and asks funders for support in the second half of 2026, so every grant to the category she names is a grant to the category her employer occupies (Humane Intelligence). Best case: hers: the harms that are “relatively likely to happen” are the ones affecting people now, “bias, discrimination, factual inaccuracies, or hallucinations in one-off hiring, healthcare, financial, or economic decisions.” Others who gain: her clients, the public-health practitioners the working group serves, and her funders, Google.org and the Heising-Simons Foundation per the organization’s pages.
2. Position: evaluations should be product tests built into development, not leaderboards (“AI evals shouldn’t be a box-ticking compliance chore,” Transformer). What adoption pays: the op-ed’s authors name a company of the same name “building the infrastructure for the next generation of real world AI evaluations at scale,” and Subject runs the nonprofit; the piece says what each organization is and gives both websites, and states no relationship between them beyond the shared name and the co-founder’s 2022 start. Best case: theirs: “Products that have baked-in AI evals as part of their design and development lifecycles are best placed to actually deliver desired outcomes.” Others who gain: any evaluator that sells product testing.

Alliance Map

The Corporate and Movement rows overlap in people and in time. The benchmark’s former program manager runs the nonprofit that sells bespoke evaluations, co-writes with the founder who now runs a company of the same name, and counts a GitHub executive as her board’s secretary, and the organization sits in the public-health, UN and philanthropy circuits her talks page lists. The overlap is the finding, and the pages disclose all of it. See the small world.

Convergent Drive Classification

Self-preservation: Moved up inside one organization, Head of Impact to Interim to permanent Executive Director in thirteen months, and says on her own page that she is open to other roles (Humane Intelligence, Team; Mala Kumar). Goal preservation: Keeps one object in view: evaluation as a public capacity. First-hand: “AI evaluations focus on the use of the generative AI model or AI system” (Humane Intelligence, 12 August 2026 post). Resource acquisition: Grants and service revenue: Google.org funded the open source release, her post says, and her own page claims “7 figures of revenue” under her strategy (Humane Intelligence, OSS release post; Mala Kumar). No salary is on any page read. Self-improvement: From UX research and open source at the UN and GitHub, to a benchmark’s program office, to running an evaluator. First-hand: “She was one of the first people globally to introduce UX research and design to the UN” (Humane Intelligence, Team).

Subject is not an AI system. The drives appear anyway — in the program manager who took the chair.

Public footprint: malakumar.com · Humane Intelligence team · github.com/malakumar85 · arXiv:2610.00529.

Sources: Humane Intelligence — Team; Humane Intelligence — About; Humane Intelligence — home; Humane Intelligence — AI Red Teaming; Humane Intelligence — Board & Advisory; Humane Intelligence — 2025 Impact Report; Humane Intelligence — We’re releasing our AI red teaming app under an OSS license; Humane Intelligence — What is the relationship between AI safety and AI evaluations?; Mala Kumar — home; Mala Kumar — AI Evals and Tech for Social Good; Mala Kumar — Talks and Media; Transformer — Why AI evals need to reflect the real world; IT Brief — Why AI in public health needs focus, funding, and community voice; AILuminate — arXiv; OB-CAIE — arXiv; GitHub — malakumar85; ProPublica Nonprofit Explorer — Humane Intelligence Inc.