HUMANE INTELLIGENCE

Institutional Archetype

THE CONVENER OF THE CROWD-TESTERS — Humane Intelligence runs red-teaming events, bias bounties and bespoke evaluations in which large groups of people, many of them without engineering credentials, probe AI systems for bias, error and misuse. Its home page says it is “a 501(c)(3) nonprofit dedicated to breaking down barriers to AI deployment for social good” that “collaboratively design[s] and run[s] rigorous evaluations that make AI systems more accountable, responsible, and fair” (Humane Intelligence). Its best-known act is the generative red team at DEF CON 31 in August 2023, run with AI Village, SeedAI and the AI Vulnerability Database and supported by the White House Office of Science and Technology Policy (AI Village).

Two organizations now carry the name. The nonprofit, led since 2025 by Mala Kumar, sells red-teaming workshops and contextual evaluations and publishes its software as open source. Its founder, Rumman Chowdhury, says on her own site that she is “the CEO and founder of Humane Intelligence public benefit corporation, and the co-founder of Humane Intelligence, a nonprofit.” The company’s site offers a “modular T&E stack,” “Evaluation-as-a-Service” and “audit-grade” deliverables to “a frontier model lab, an enterprise deploying agents at scale, or a regulator testing assurance frameworks.” Fathom, the sponsor of California’s independent-verification statute, lists the nonprofit as a partner and quotes Chowdhury as CEO of the company. Whether either entity is, or wants to be, a designated verifier is not stated on any page of either. This profile scores the position, not the people in it.

Mandate & Origin

The nonprofit’s Mission page: “Our mission is to empower people and organizations that represent or have lived experience in a given problem space, through our technical AI evaluation design, our convening events, and our red teaming web application. Our vision is to create a world in which everyone has the capacity, training and tools to evaluate AI systems that affect their lives.” Its own history page: in 2022, “While working at Twitter, Humane Intelligence co-founder, Dr. Rumman Chowdhury, developed the concept of ‘bias bounties’,” and the group “unofficially launched under the name ‘Bias Buccaneers.’” In 2023, “Bias Buccaneers dba Humane Intelligence was incorporated as a nonprofit organization in May 2023, and expanded to host in-person red teaming events, including the largest ever generative AI red teaming exercise at DEFCON 31, a scientific mis-and dis-information red teaming exercise with the UK Royal Society, and a red teaming event on intellectual property rights at Autodesk University.” In 2024 it “officially changed its name to Humane Intelligence” and ran “the regional AI safety red teaming event with the Singapore IMDA” (Humane Intelligence, Mission & Vision; Humane Intelligence, About).

The leadership change is on the same page: “Rumman stepped down as CEO of Humane Intelligence and Mala Kumar stepped in as Interim Executive Director on 1 August 2025,” and “Mala Kumar was named Executive Director (no more interim!) in April 2026.” The Team page gives Kumar’s earlier posts: Director of Program Management, AI Safety, at MLCommons, where she “oversaw the core team that created its inaugural AI safety benchmark,” and Director of Tech for Social Good at GitHub. It lists Julie Hollek as Chief Technology Officer (formerly Director of Data Science at Mozilla and builder of the Health Data Science team at Twitter, and chair of the NumFOCUS board), Kristi Arbogast as AI Red Teaming Manager, and four others. The 2025 Impact Report says Kumar took over “from Co-Founder, Dr. Rumman Chowdhury.” A June 2025 post on the nonprofit’s own insights page and Partnership on AI’s August 2025 release both describe Chowdhury as “Co-Founder and Distinguished Advisor”; she appears on neither the current Team nor the Board page (Humane Intelligence, About; Humane Intelligence, Team; Humane Intelligence, 2025 Impact Report; Humane Intelligence, Insights; Partnership on AI; Humane Intelligence, Board & Advisory).

The board: Ben Kinsella, President (“previously a member of OpenAI’s Human Data team”), Admas Kanyagia, Secretary (leads Diversity, Inclusion, and Belonging at GitHub), Amy Sample Ward (CEO of NTEN), Amy Sommer (Vice President of Strategy, Evaluation and Impact at Tides) and Abby Vollmer (leads Cloudflare’s Legal Online Trust team). A nine-member non-voting advisory group includes Michael Zargham (Dynamical Systems Group), Diego Garcia-Olano (“Senior Research Scientist, Meta”), Lauren Damme (“Director of Research, U.S. Department of Labor”) and Afua Bruce (Humane Intelligence, Board & Advisory).

The legal entities. ProPublica’s Nonprofit Explorer lists Humane Intelligence Inc, Katy, Texas, “Tax exemption issued: March 2024,” EIN 93-1491311, with a Form 990 for the year ending December 2024 submitted 26 August 2026; the return gives the year of formation as 2023 and the state of legal domicile as TX, and lists Rumman Chowdhury as “Chair And Ceo,” Chloe Autio as a director and Juta Williams as secretary, each at $0. The FY2023 return, filed 17 May 2024, lists Chowdhury as “Co-founder And Board Chair” and Jutta Williams as “Co-founder And Board Member.” The company is Humane Intelligence PBC, which describes itself as “Humane Intelligence PBC — Test & Evaluation for Frontier AI,” lists Chowdhury as “Founder & CEO” and Paolo Calvi as Staff Engineer, and is reachable at humane-intelligence.com (ProPublica Nonprofit Explorer; Humane Intelligence PBC; Rumman Chowdhury, About).

Funding & Backers

Funders named on its own pages. The 2025 Impact Report: “Humane Intelligence would not be able to do its work without the generous funding of organizations such as the Heising-Simons Foundation, Google.org, and more,” with two 2025 announcements: “our OSS AI red teaming application and the launch of our bias bounties at scale on Zindi.” No amount is published for either. The December 2025 post on the open-source release: “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 bias-bounty post: “Thanks to the generous support of the Heising-Simons Foundation, Humane Intelligence nonprofit is excited to announce that we are partnering with Radiant Earth to move our bias bounty program onto Zindi,” and “With support from Google.org, Humane Intelligence pioneered the concept of ‘bias bounties’” (Humane Intelligence, 2025 Impact Report; Humane Intelligence, OSS funding announcement; Humane Intelligence, Bias Bounties at Scale).

Money, from the IRS return. FY2024 (calendar 2024): revenue $472,206, of which $372,175 (78.8 percent) is program-service revenue and $100,000 (21.2 percent) contributions; expenses $982,939; net loss $510,733; year-end net assets $61,740. The return reports no employees for the year; Schedule O prints contractor payments of $810,336 and in-person-event costs of $15,294. It states the mission as “a community of practice around algorithmic evaluations of AI models.” FY2023: revenue $600,000, all contributions, expenses $15,000, net assets $651,418, with ProPublica’s summary of that return noting that it “Operated a donor-advised fund.” The nonprofit’s red-teaming page says it has “worked with education companies, international civil society, industry, and governments to design and run red teaming events,” and its home page sells red teaming and contextual evaluations “as a paid service”; no client is named (ProPublica Nonprofit Explorer; ProPublica, FY2024 Form 990; Humane Intelligence, AI Red Teaming).

What it asks for. The home page carries “Our Current Funding Requests” for the second half of 2026 and the AI in Public Health fund; the 2025 end-of-year post says “In January or February 2026, we will announce a new set of project and funding priorities.” A “Philanthropic Grantees” offering sheet, last updated 7 May 2026, markets evaluations to “a recent philanthropic funding recipient tasked with implementing a generative AI model or system for the first time,” with “standard pricing” on request; its podcast notes say “a truly comprehensive AI evaluation can run to $5 million” (Humane Intelligence; Humane Intelligence, 2025 wrap-up; Humane Intelligence, AI Evaluation Offerings; Humane Intelligence, Insights).

Institutional Voice & Intent

In its own words the nonprofit sells three things: red teaming (“a semi-structured testing approach to assess and improve the safety and effectiveness of AI models and systems”), contextual evaluations (“rigorous, mixed-method, bespoke evaluations designed to give a comprehensive analysis of an AI model or system’s performance for a specified problem space,” methodology published in October 2026 as ontology-based contextual AI evaluation, OB-CAIE) and bias bounties, now run on Zindi. Its clients, per its contextual-evaluations page, ask “I’m being pitched by a vendor and I don’t know how to validate their claims” (Humane Intelligence; Humane Intelligence, AI Contextual Evaluations; arXiv:2610.00529).

Stated intent: Widen the group of people who can evaluate AI systems, and make evaluations “more accountable, responsible, and fair.”

Observed intent: Run paid red-teaming and evaluation engagements for clients, convene public events that supply the community and the data, and release its application as open source. The founder’s company runs the commercial version under the same name; its pitch, in the company’s words, is “audit-grade evidence.” Chowdhury’s own description of the work, quoted by Fathom in July 2026: “A hundred percent of the time, I am co-working to help define good. Right now we do not have very clear standards” (Humane Intelligence PBC; Fathom, Who Evaluates the Evaluators?).

Gap: Neither entity calls itself a verifier. The nonprofit’s pages do not describe its own work as independent, an audit, assurance or verification; they call it evaluation and red teaming, and say the work is “bespoke” and “collaboratively designed” with clients and communities. The company’s site uses “audit-grade” and addresses regulators. Both are, on their own words, hired to test a system by the party that wants it tested; Chowdhury says as much: “Being a purely independent evaluator is nearly impossible, because it’s just not a fully built-out ecosystem yet. But it needs to be.” What neither publishes is a client list, a conflict-of-interest rule for evaluations, or a count of engagements by client type. The ProPublica return shows 78.8 percent of the nonprofit’s 2024 revenue as program-service revenue.

Interest: Humane Intelligence’s founder and staff argue for funded, legally protected, publicly participatory red teaming and for a trained evaluator workforce. What adoption would pay is set out from their own filings and pages.

• Fund and protect independent red teaming. Position, from Chowdhury’s written testimony to the House Committee on Science, Space, and Technology, in which she identifies herself as “co-founder of the nonprofit Humane Intelligence”: “funding for independent groups to conduct red teaming and adversarial auditing, similar to ’ethical hacking’ more common in traditional security fields, and second, legal protections so these individuals - operating in the public good - are not silenced with litigation.” What it pays: the nonprofit’s revenue is red-teaming and evaluation engagements; the same testimony describes the DEF CON event as an effort “to grow a new profession,” and notes “there is currently little to no investment in training and upskilling for this nascent practice” (Chowdhury, written testimony). Best case, her own: “we need this practice to be supported by non-corporate actors as well.” Convergence: other evaluators on this site sell the same service, among them Gray Swan, and the testimony itself notes that her “colleagues at Google DeepMind, Open AI, Anthropic and more have engaged in this practice to refine their flagship AI models.”
• Train an AI-trust workforce by statute. Position, from the nonprofit’s policy page: “As an organization that builds communities of practice for algorithmic assessment, Humane Intelligence is enthusiastic about the Workforce for AI Trust Act. Having experienced workforces that are able to deploy AI models and systems in safe and secure manners is fundamental to realizing the benefits of AI.” The House Science Committee’s page for H.R. 9215 states its summary as “To facilitate a workforce of trained experts to build trustworthy AI systems” and lists Humane Intelligence among its endorsers, beside Google, IBM, OpenAI, Booz Allen Hamilton, Americans for Responsible Innovation, the Center for AI Policy and SeedAI. What it pays: a statutory market for trained evaluators matches what the nonprofit sells, a LinkedIn Learning course on AI evaluations and a free red-teaming playbook; the bill’s text and any sale are not tied to a named purchaser here. Convergence: the labs and platforms that endorsed the bill hire the people it would train (Humane Intelligence, Policy; House Science Committee, H.R. 9215).
• Accredit evaluators. Position, in Chowdhury’s words in the Fathom piece: “in order to have individuals who are considered experts, we do have to have something to assess them against.” What it pays: an accreditation regime sorts evaluators into qualified and not; neither Humane Intelligence entity has any designation, and neither says it seeks one. Convergence: Fathom’s independent-verification design is the same idea at organizational level, and Fathom lists the nonprofit as a partner (Fathom, Who Evaluates the Evaluators?; Fathom, Partnerships).

Interest is a documented fact about each position, not a claim about motive, and not a verdict that the argument is wrong. Each best case is stated at the strength its author states it.

Position in the Apparatus

• DEF CON 31, August 2023. AI Village’s announcement of 3 May 2023 (by Sven Cattell, Rumman Chowdhury and Austin Carson): “This event is supported by the White House Office of Science, Technology, and Policy, the National Science Foundation’s Computer and Information Science and Engineering (CISE) Directorate, and the Congressional AI Caucus,” with models “kindly provided by Anthropic, Google, Hugging Face, NVIDIA, OpenAI, and Stability with participation from Microsoft, on an evaluation platform developed by Scale AI.” The White House fact sheet of 4 May 2023 announced “an independent commitment from leading AI developers, including Anthropic, Google, Hugging Face, Microsoft, NVIDIA, OpenAI, and Stability AI, to participate in a public evaluation of AI systems” at AI Village (AI Village; White House, Fact Sheet, 4 May 2023). See Scale AI, Anthropic and OpenAI.
• Singapore IMDA, 2024. “IMDA, together with Humane Intelligence, completed the world’s first-ever multicultural and multilingual AI Safety Red Teaming Challenge focused on Asia-Pacific in November and December 2024! Over 350 participants across 9 countries tested 4 large language models for bias stereotypes in English and regional languages” (Singapore AISI).
• UNICC, September 2024. The UN International Computing Centre’s announcement of a partnership “to develop artificial intelligence (AI) governance tools and infrastructure for the efficient and responsible deployment of AI within the UN system,” including an “AI Model Assessment Playbook”; the nonprofit’s own pages do not mention it (UNICC).
• 2025 Impact Report. “In 2025, we worked with more than 200 organizations on our events, evaluations, programs, advisory group, and volunteer groups”; “68% of our events in 2025 focused on red teaming”; work “with the UK Government’s DSIT and FCDO on a report about Evaluating Survivor-Centric Tools for Intimate Image Abuse in the Age of Generative AI”; a red-teaming event at RSA “pitted human red teaming efforts against automated red teaming tools”; the Red Teaming AI for Social Good Playbook, published with UNESCO (Humane Intelligence, 2025 Impact Report; Humane Intelligence, AI Red Teaming).
• Software and methods. The red-teaming application was released under an Apache 2.0 license in September 2026; a new client-facing version is due in October 2026. The OB-CAIE methodology paper (arXiv:2610.00529, October 2026) is by Julie Krugler Hollek, Michael Zargham and Mala Kumar. Bias bounties moved to Zindi, with the first launched in August 2026 (Humane Intelligence; Humane Intelligence, AI Red Teaming; arXiv:2610.00529).
• Scale. “More than 5,600 people have participated in our in-person and remote events in five continents”; “Humane Intelligence has run more than 30 red teaming, bias bounty, and contextual evaluations”; participants and stakeholders “from more than 60 countries” (Humane Intelligence).
• Policy. The nonprofit joined the Partnership on AI network in August 2025; its statement on joining: “We have worked with partners globally, including Singapore IMDA, UNESCO, and FEMYSO.” On California’s SB 813, neither entity’s pages carry a statement, and the nonprofit is not on the registered support, support-if-amended or opposition lists in the Assembly Privacy Committee’s analysis, and is not named in the Senate Judiciary Committee’s; the Assembly analysis does cite Chowdhury’s 2021 Twitter engineering post on the image-cropping algorithm among its footnotes (Partnership on AI; Assembly Privacy analysis). See Partnership on AI.

Actions & Leadership Choices

CONDUCT: WATCH — TWO ENTITIES UNDER ONE NAME, AND A RELATED-PARTY ITEM IN THE NONPROFIT’S OWN FILING. The nonprofit’s own return reports a transaction with its former chief executive that it describes as uncorrected, and its founder now runs a company that sells the same service under the same name. The nonprofit’s pages make no mention of the company, and the company’s site says Chowdhury co-founded the nonprofit. Neither fact alleges anything about the quality of any evaluation.

• 2022. “Bias Buccaneers” launches unofficially (Humane Intelligence, About).
• May 2023. Incorporated as a nonprofit; DEF CON 31 generative red team in August; House testimony (Humane Intelligence, About; Chowdhury, written testimony).
• March 2024. IRS exemption issued (ProPublica Nonprofit Explorer).
• September and November 2024. UNICC agreement announced; IMDA challenge in Asia-Pacific (UNICC; Singapore AISI).
• 1 August 2025. Chowdhury steps down as CEO; Kumar becomes Interim Executive Director (Humane Intelligence, About).
• August 2025. Joins the Partnership on AI network (Humane Intelligence, Policy).
• December 2025. Two funding requests reported fully funded: bias bounties on Zindi (Heising-Simons Foundation) and the open-source release (Google.org) (Humane Intelligence, 2025 wrap-up).
• April 2026. Kumar named Executive Director (Humane Intelligence, About).
• 21 July 2026. Fathom’s article quotes Chowdhury as founder and CEO of Humane Intelligence PBC (Fathom, Who Evaluates the Evaluators?).

• September and October 2026. Open-source red-teaming application released (16 September); OB-CAIE paper (1 October) (Humane Intelligence).

Alliance Map

Where ties on different axes run at the same time, the overlap is the finding. The nonprofit tests models that Anthropic, Google, OpenAI and others supplied for a public event the White House backed, takes grant money from the philanthropy arm of one of those developers, sells workshops to industry and governments, lists a former OpenAI staffer as board president and a Meta researcher as an advisor, and endorsed a workforce bill that Google, IBM and OpenAI also endorsed. Its founder’s company sells “audit-grade” evaluation under the same name, and Fathom, whose statute would designate verifiers, lists the nonprofit as a partner and quotes the founder as the company’s CEO. The documents show disclosed arrangements, not coordination, and no page ties the nonprofit to any designation under SB 813.

Reach Assessment

Institutional: A public-event format and a methods library used by governments and international bodies, backed by a small organization. Memetic: The bias-bounty and public-red-team formats it popularized, now open-sourced, are the form most likely to travel.

Civilizational: If crowd-run red teaming becomes the default public test of AI systems and the people who convene it are paid by the parties and funders whose systems are tested, the field’s idea of what counts as “tested” is set by convening organizations with little revenue and thin independence. The nonprofit is open about the structure: funders named, a board and advisory group listed, and a filing that reports its own related-party item. What a reader cannot check is who the clients are and what share of revenue each pays.

Score basis. Atk 5 is reach: the nonprofit counts more than 5,600 event participants and more than 60 countries, more than 200 partner organizations in 2025, and a DEF CON 31 event the White House OSTP supported, with an IMDA challenge, a UNICC agreement and UK government work (Humane Intelligence; Humane Intelligence, 2025 Impact Report; AI Village). The points under a higher number are that no statute or release gate makes any lab or agency use its output and that no lab release document cited here names it. Peers on atk 5: Apart Research, a policy channel into the EU AI Office with no gate; AVERI, with endorsed bills and pilot audits but no statutory seat; Gray Swan holds 8 on lab pre-release cards and a government contract line. Def 4 is funding insulation: the FY2024 return shows $372,175 of $472,206 in revenue as program-service revenue and $100,000 as contributions, its funders named on its own pages are Google.org and the Heising-Simons Foundation with no amounts, and its Schedule L reports an uncorrected related-party item (ProPublica Nonprofit Explorer; Humane Intelligence, 2025 Impact Report). The points above 4 are that no lab is named as a funder and the organization reports its own governance finding. Peers on def 4: Halcyon Futures, whose donors are not named; AVERI and the AI Evaluator Forum hold 6 on disclosed funders and no cash from frontier AI companies. Hp 3 is entrenchment: net assets of $61,740 after a $510,733 deficit, no employees on the return, a founder-CEO replaced in August 2025, and software released under an open license that removes the organization from the path of anyone who runs it (ProPublica Nonprofit Explorer; Humane Intelligence). The points above 3 are the 200-organization partner count and a UN agreement. Peers on hp 3: Halcyon Futures, no statute or procurement holding it in place; AVERI and the AI Evaluator Forum hold 4 on filed standards work and a member roster.

Sources: Humane Intelligence — home; Humane Intelligence — About; Humane Intelligence — Mission & Vision; Humane Intelligence — Team; Humane Intelligence — Board & Advisory; Humane Intelligence — Service Providers; Humane Intelligence — Policy; Humane Intelligence — AI Red Teaming; Humane Intelligence — AI Contextual Evaluations; Humane Intelligence — Insights; Humane Intelligence — 2025 Impact Report; Humane Intelligence — OSS funding announcement; Humane Intelligence — Bias Bounties at Scale; Humane Intelligence — 2025 end-of-year wrap-up; Humane Intelligence — AI Evaluation Offerings; Humane Intelligence PBC; Rumman Chowdhury — About; ProPublica Nonprofit Explorer — Humane Intelligence Inc; ProPublica — FY2024 Form 990; ProPublica — FY2024 Schedule L; ProPublica — FY2024 Schedule O; Fathom — Partnerships; Fathom — Who Evaluates the Evaluators?; Partnership on AI — welcomes 10 new partners; UNICC — UNICC and Humane Intelligence; AI Village — Generative Red Team announcement; White House — Fact Sheet, 4 May 2023; Singapore AISI — IMDA red teaming challenge; Chowdhury — written testimony; House Science Committee — H.R. 9215; Assembly Privacy Committee — SB 813 analysis; arXiv:2610.00529.