Making 2IA the definitive directory of AI organizations

Full report

On this page

Executive summary

As of May 17, 2026, the public site at https://2ia.org is not yet an AI organization directory. It currently presents itself as “2IA – Two Identities Of Anonymous,” with top-level navigation for Home, Start Here, Research Archive, Methodology, Support, and About, and it describes itself as an independent public-intelligence and civil-liberties research project. The archive is organized around issue hubs, dossiers, investigations, guides, toolkits, and corrections, not organization profiles. One public archive page still exposes the default WordPress “Hello world!” post in May 2026, and searches on the homepage for terms such as “directory,” “LinkedIn,” and “GitHub” did not surface organization-listing functionality.

That means the opportunity is not to refine an existing directory, but to launch one on top of a useful editorial foundation. The strongest parts of the current site are exactly the pieces a trustworthy “who’s who” directory needs: a stated source hierarchy, visible confidence labels, a right-of-reply and corrections pathway, and a privacy-first posture that rejects unnecessary tracking. Those policies already distinguish 2IA from low-trust link farms and generic SEO directories.

The highest-leverage move is to make 2IA entity-centric. Each organization should get a stable canonical profile page with normalized fields, exact official links, a verification ledger, tags, and change history. Around those profiles, 2IA should publish faceted browse pages by focus area, country, organization type, and update status, along with a public dataset export. The first public launch should target roughly 100–150 anchor organizations across frontier labs, safety and evaluation groups, governance and civil-society institutions, standards bodies, academic institutes, and open-source communities.

The first-priority organizations to add are the institutions that define the field’s structure: major frontier labs such as OpenAI, Anthropic, Google DeepMind, and Ai2; measurement and benchmark institutions such as MLCommons and METR; safety and governance nonprofits such as CAIS, Partnership on AI, GovAI, Ada Lovelace Institute, and IAPS; public-sector nodes such as the UK AI Security Institute and NIST’s CAISI; and standard-setting bodies such as GPAI and ISO/IEC JTC 1/SC 42. These will give 2IA immediate credibility across research, deployment, governance, and standards.

A crucial implementation detail is to treat the directory as a dataset as well as a website. Schema.org supports exactly the structures needed here: Organization for each profile, sameAs for canonical social identities, ContactPoint for member/contact paths, foundingDate for provenance, CollectionPage and ItemList for browse pages, BreadcrumbList for navigation, and Dataset for downloadable exports. Google’s documentation also recommends using structured data, clear canonical URLs, crawlable URL structures, and sitemaps submitted through Search Console.

Finally, do not build the ongoing verification workflow around deprecated APIs or generic search hacks. GitHub’s REST API and DBLP’s search API are appropriate durable inputs for repository and publication verification. By contrast, Google’s Custom Search JSON API is already on a deprecation path for existing customers, with a migration deadline of January 1, 2027, so it should not become a core dependency.

What 2IA is today

The public site today is best understood as a research publication framework, not a directory. Home, Start Here, About, and Methodology all describe a civil-liberties/public-intelligence project concerned with surveillance, metadata, algorithmic risk, public records, and corrections. The Research Archive explicitly says it is organized into issue hubs, dossiers, investigations, field guides, toolkits, case studies, trackers, briefings, corrections, and public-records libraries.

The current public crawl therefore yields no AI/IA organization profiles to normalize. The public archive also still contains a default WordPress archive entry for May 2026, which is a strong signal that the site is still early in its build-out.

That inconsistency matters. Before 2IA tries to become the canonical directory for a field, it needs to become canonical about itself. The domain currently communicates one editorial identity on-page, a different one in at least one search result, and a third string on a corrections page. At minimum, this calls for title normalization, consistent <title>/H1/brand usage, stable canonicals, and Search Console monitoring before the organization directory goes live.

The good news is that 2IA’s current editorial infrastructure is unusually strong for a future directory. Its Methodology page already lays out primary-source preference, confidence labels, public-records handling, AI-use disclosure, minimization, correction workflows, and right of reply. Its Corrections page makes visible correction handling and reply rights explicit, and its Privacy Policy states that the theme does not add analytics, tracking pixels, external fonts, CDN assets, or unnecessary cookies by default. Those policies should be kept and repurposed, not discarded.

The directory model 2IA should adopt

The strongest way to make 2IA definitive is to adopt a profile-plus-ledger model rather than a simple link list. Each organization should have one canonical profile URL, one normalized identity record, one short authoritative description, and one visible verification state. The page should answer a small set of durable questions: who the organization is, what it does, where it is based, when it was founded, how to contact or join it, what its official web/social identities are, what it is known for, and what evidence was used to verify those claims. That approach matches 2IA’s current methodology far better than a generic roundup page would.

A good inclusion rule is: include durable institutions with a material role in AI research, development, safety, governance, standards, infrastructure, or field-building. Exclude one-off events, conference editions, temporary working groups without stable identity, unaffiliated product pages, and personal blogs. When naming is ambiguous, prefer the organization’s own official brand as link text and store aliases as alternate names. That matters already for entities such as Ai2, GovAI, the UK AI Security Institute, and NIST’s CAISI, all of which have meaningful short names or recent naming changes that users will search for.

A useful publication and browsing taxonomy would look like this:

graph TD
    A[AI organizations] --> B[Frontier labs]
    A --> C[Safety and evaluations]
    A --> D[Governance and policy]
    A --> E[Standards and benchmarks]
    A --> F[Academic institutes]
    A --> G[Open-source communities]
    A --> H[Public-sector bodies]

    B --> B1[OpenAI]
    B --> B2[Anthropic]
    B --> B3[Google DeepMind]
    B --> B4[Ai2]
    B --> B5[xAI]
    B --> B6[Mistral AI]

    C --> C1[CAIS]
    C --> C2[METR]
    C --> C3[ARC]
    C --> C4[CHAI]
    C --> C5[Stanford Center for AI Safety]

    D --> D1[GovAI]
    D --> D2[Ada Lovelace Institute]
    D --> D3[IAPS]
    D --> D4[Partnership on AI]

    E --> E1[MLCommons]
    E --> E2[ISO/IEC JTC 1 SC 42]
    E --> E3[GPAI]
    E --> E4[AAAI]
    E --> E5[ACM SIGAI]

    F --> F1[Stanford HAI]
    F --> F2[Mila]
    F --> F3[Vector Institute]
    F --> F4[Amii]
    F --> F5[ELLIS]

    G --> G1[Hugging Face]
    G --> G2[LAION]
    G --> G3[LF AI and Data]

    H --> H1[UK AISI]
    H --> H2[NIST CAISI]

The site architecture should mirror that taxonomy. A definitive directory should have stable profile pages under a path such as https://2ia.org/org/{slug}/, plus browse pages such as https://2ia.org/focus/ai-safety/, https://2ia.org/country/united-kingdom/, https://2ia.org/type/research-institute/, and https://2ia.org/updated/. Current 2IA archive structures are chronological and editorial; they are not a good primary navigational model for an entity directory.

Priority organizations to add

The table below combines the two most important missing-org deliverables: a prioritized seed list and paste-ready link text plus short descriptions. The URLs are exact official URLs to use as starting points; where a second official page is especially helpful, it is included too.

First-wave seed set

Second-wave coverage set

After those twenty, the next additions I would queue are xAI, Mistral AI, Amii, Stanford HAI, Stanford Center for AI Safety, and ACM SIGAI. xAI and Mistral matter because they are now prominent frontier-model organizations; Amii matters because it rounds out Canada’s three-institute national AI strategy; Stanford HAI and Stanford Center for AI Safety add major academic and policy capacity; and ACM SIGAI strengthens the professional-society layer.

The most important naming-normalization rules are straightforward. Use the official brand where one clearly exists, and store aliases separately. That means Ai2 rather than “Allen Institute for AI” as link text, GovAI with “Centre for the Governance of AI” as an alternate name, The AI Security Institute with a note that it was previously called the AI Safety Institute, and Center for AI Standards and Innovation with a note that it now occupies the role many users will still search for as the U.S. AI Safety Institute.

Metadata, taxonomy, and navigation

The technical target should be a site that search engines, LLMs, and human readers can all interpret in the same way. Google’s Organization structured data documentation explicitly recommends using organization markup for administrative details such as name, address, contact information, and business identifiers, while Schema.org provides the finer-grained properties a directory like this needs, including sameAs, foundingDate, ContactPoint, CollectionPage, ItemList, BreadcrumbList, and Dataset.

In practice, every profile page should render three things simultaneously: a human-readable page, a machine-readable Organization entity, and a visible editorial verification status. The minimal profile schema should contain official name, alternate names, official website, canonical social URLs, country, founding date, contact or membership URL, focus tags, and last-reviewed date. Because Google recommends clear canonicalization and consistent internal linking to the canonical page, every organization should have one stable slug and one preferred URL, even if there are sort, filter, or querystring variants elsewhere on the site.

A good starting JSON-LD template for an organization profile page is:

{
  "@context": "https://schema.org",
  "@graph": [
    {
      "@type": "Organization",
      "@id": "https://2ia.org/org/{slug}/#org",
      "name": "{official_name}",
      "alternateName": ["{alias_1}", "{alias_2}"],
      "url": "{official_website}",
      "sameAs": [
        "{official_x_url}",
        "{official_linkedin_url}",
        "{official_github_url}"
      ],
      "foundingDate": "{YYYY-MM-DD or YYYY}",
      "description": "{authoritative_1_to_2_sentence_description}",
      "location": {
        "@type": "Place",
        "address": {
          "@type": "PostalAddress",
          "addressCountry": "{ISO_3166_1_alpha_2}"
        }
      },
      "contactPoint": {
        "@type": "ContactPoint",
        "url": "{official_contact_or_membership_url}",
        "contactType": "{contact|membership|press|general}"
      },
      "knowsAbout": ["{tag_1}", "{tag_2}", "{tag_3}"]
    },
    {
      "@type": "WebPage",
      "@id": "https://2ia.org/org/{slug}/#page",
      "url": "https://2ia.org/org/{slug}/",
      "name": "{official_name}",
      "about": {
        "@id": "https://2ia.org/org/{slug}/#org"
      },
      "dateModified": "{last_reviewed_date}"
    },
    {
      "@type": "BreadcrumbList",
      "@id": "https://2ia.org/org/{slug}/#breadcrumb",
      "itemListElement": [
        {
          "@type": "ListItem",
          "position": 1,
          "name": "Organizations",
          "item": "https://2ia.org/org/"
        },
        {
          "@type": "ListItem",
          "position": 2,
          "name": "{official_name}",
          "item": "https://2ia.org/org/{slug}/"
        }
      ]
    }
  ]
}

This design is directly aligned with Google’s Organization guidance plus Schema.org’s organization, sameAs, contact, founding-date, and breadcrumb vocabulary.

Browse pages should use CollectionPage and ItemList, not blog-like archives. For example, a page such as https://2ia.org/focus/ai-safety/ should explicitly declare itself as a collection page whose main entity is a list of organizations.

{
  "@context": "https://schema.org",
  "@type": "CollectionPage",
  "@id": "https://2ia.org/focus/ai-safety/",
  "url": "https://2ia.org/focus/ai-safety/",
  "name": "AI safety organizations",
  "mainEntity": {
    "@type": "ItemList",
    "itemListOrder": "https://schema.org/ItemListOrderAscending",
    "numberOfItems": "{count}",
    "itemListElement": [
      {
        "@type": "ListItem",
        "position": 1,
        "url": "https://2ia.org/org/center-for-ai-safety/"
      },
      {
        "@type": "ListItem",
        "position": 2,
        "url": "https://2ia.org/org/metr/"
      }
    ]
  }
}

For a site that wants to be citable by researchers and reusable by others, 2IA should also publish a machine-readable directory export and mark it up as a Dataset.

{
  "@context": "https://schema.org",
  "@type": "Dataset",
  "name": "2IA AI organization directory",
  "description": "Structured directory of AI, AI safety, governance, standards, academic, public-sector, and open-source organizations.",
  "url": "https://2ia.org/data/organizations/",
  "license": "{dataset_license_url}",
  "distribution": [
    {
      "@type": "DataDownload",
      "encodingFormat": "application/json",
      "contentUrl": "https://2ia.org/data/organizations.json"
    },
    {
      "@type": "DataDownload",
      "encodingFormat": "text/csv",
      "contentUrl": "https://2ia.org/data/organizations.csv"
    }
  ]
}

On navigation, the current site should add a first-class organizations layer rather than forcing users through editorial chronology. The top navigation should become something like Organizations, Tags, Countries, Types, Projects, Publications, Updates, Methodology, Corrections. Every org page should show “related organizations,” “same-country organizations,” and “same-tag organizations,” because strong internal linking helps users and helps search engines discover the pages that matter. Google’s own documentation emphasizes concise, relevant internal linking and crawlable site structures, along with sitemap submission and Search Console monitoring.

Update and verification plan

The current 2IA methodology is already the right philosophical model for ongoing directory maintenance. The key move is to turn it into a repeatable verification pipeline. Each organization record should have a visible status such as Confirmed, Likely, Needs review, Disputed, Updated, or Corrected, mirroring the confidence-label logic described on the current Methodology page. Each material edit should be logged the same way the site already envisions corrections and right-of-reply handling.

The best production workflow is straightforward. Use official sites as the default source of truth for name, website, description, contact path, and membership path. Use the organization’s own footer, about page, or existing sameAs markup to find canonical social URLs. Use GitHub’s REST API to verify whether an official GitHub organization exists and which repositories are current. Use DBLP’s search API to verify publications, venues, and author affiliations for research-oriented institutions. Use LinkedIn and Google Scholar as human QA layers, not as the system of record. That keeps the pipeline aligned with official sources while still being resilient when official pages are sparse.

A practical discrepancy policy should flag, at minimum, these cases: official website URL mismatch; redirect chains that point to a different canonical site; organization name mismatch between site and social accounts; founding year mismatch across official site and profile pages; headquarters country mismatch; archived or dormant social links; renamed or absorbed organizations; and publications whose current affiliations no longer match the organization profile. The directory should show these as transparent notes rather than silently “fixing” them. That fits 2IA’s existing commitment to visible uncertainty and public correction.

The crawl cadence should be risk-based. Frontier labs, government AI institutes, and standard-setting bodies change quickly enough that they should be rechecked every two weeks. Well-funded nonprofits, benchmark groups, and active open-source institutions should be rechecked monthly. Academic centers and slower-moving institutes can usually be rechecked quarterly unless a rename, merger, or major publication release is detected. XML sitemaps, sitemap indexes, and Search Console should be part of the operating loop so that fresh updates are discoverable and indexing issues are visible.

The 90-day launch sequence should look like this:

gantt
    title Recommended 90-day launch sequence for 2IA directory
    dateFormat  YYYY-MM-DD
    section Foundation
    Scope and inclusion rules          :a1, 2026-05-20, 10d
    Brand and canonical cleanup        :a2, 2026-05-20, 14d
    Data model and taxonomy            :a3, 2026-05-24, 14d
    section Build
    Profile template and JSON-LD       :b1, 2026-06-03, 14d
    Browse pages and faceted nav       :b2, 2026-06-10, 21d
    Dataset export and sitemap setup   :b3, 2026-06-20, 14d
    section Content
    Seed 100-150 anchor orgs           :c1, 2026-06-01, 35d
    Verification ledger and QA         :c2, 2026-06-15, 28d
    section Launch
    Search Console, validation, fixes  :d1, 2026-07-10, 10d
    Public launch                      :d2, 2026-07-22, 1d
    section Ongoing
    Biweekly and monthly recrawls      :e1, 2026-07-23, 60d

One last implementation warning is important: if you want a programmatic search layer for discovery or QA, do not make Google’s Custom Search JSON API a foundational dependency. Google’s documentation says existing customers have until January 1, 2027, to migrate to an alternative solution. In contrast, GitHub REST and DBLP search remain appropriate durable interfaces for core verification tasks.

Open questions and limitations

The most important limitation is the starting point itself: the public 2ia.org site does not currently expose AI/IA organization listings, so the “table of current listings” is an empty-state audit rather than a normalized inventory. That conclusion is high-confidence based on the public pages reviewed, but it means the real work begins with scope definition and data-model design, not cleanup of existing entries.

The prioritized list above is intentionally a high-confidence seed set, not a claim that these are the only organizations worth including. A definitive directory will eventually need a broader second and third layer, especially for regional institutes, standards-adjacent groups, open-source collectives, and public-interest organizations that work on AI without being AI-only institutions. The right long-term answer is a transparent inclusion policy plus public corrections and nominations, not a fixed one-time list.

The other open strategic question is branding. Today, the root site is unmistakably framed as “Two Identities Of Anonymous,” with a civil-liberties mission. If the goal is to turn 2ia.org itself into the definitive AI organizations directory, that is a significant brand pivot. A lower-risk option is to keep the current editorial/research identity intact and launch the directory at a clearly separated top-level path such as https://2ia.org/org/ with its own landing page and navigation system.