Methodology

This page explains how entries get into the catalog, how openness is described, what counts as evidence, and how the numbers on the site are calculated. The same rules are published in the repository as ELIGIBILITY.md, OPENNESS.md, and CONTENT_REVIEW.md.

Eligibility

An entry is eligible when a documented basis connects its accountable entity to the United States. There are four bases:

• U.S. headquarters — the organization's documented headquarters or principal executive office is in the United States.
• U.S. nonprofit or lab — the entity is a U.S. nonprofit, university laboratory, or research institute.
• Documented U.S. control — the entity is documented as primarily controlled by a U.S. entity, for example a research unit wholly owned by a U.S.-headquartered company.
• U.S.-governed project — for software, datasets, and models: the documented governing or maintaining entity is U.S.-based, such as a U.S. foundation or company.

Some cases do not qualify on their own. A foreign organization with only a U.S. sales office is not eligible on that basis. Organizations with headquarters in two countries, and parent–subsidiary arrangements, get a written assessment on the entry that cites the documents it relies on. Project eligibility rests on documented governance, never on the names or assumed nationalities of contributors. A U.S. fine-tune or wrapper of foreign base weights does not make the base model American; the relationship is recorded as provenance.

There are no exceptions for prominent organizations. Candidates whose basis cannot be documented stay in an internal review queue and are not published. International collaboration is described accurately where it exists; non-U.S. models or dependencies may appear as clearly labeled provenance context, never as catalog members.

Each entry shows its eligibility status (Eligible, Pending review, or Excluded), the basis, a short explanation, the sources, and the date it was assessed. Only eligible entries are published.

Openness

Openness is described per artifact with a checklist suited to what the artifact is. A language model, a compiler, and a dataset expose different things, so the catalog does not force every artifact onto a single scale built for model weights.

• PublicDocumented as available to the general public. Access conditions and license terms may still apply.
• PartialSome of it is available, or access requires approval or is otherwise restricted.
• Not publicDocumented as not publicly available.
• UnknownNot yet assessed, or the evidence is insufficient.
• Not applicableDoes not apply to this kind of artifact.

Availability is kept separate from permission. "Publicly downloadable" does not by itself mean unrestricted, open source, commercially usable, or redistributable. Each release lists its actual licenses and what they apply to.

Model-disclosure tiers (USASI rubric v0.2)

For model releases only, the checklist determines a disclosure tier; the highest tier additionally requires a sourced, dated system-rights review. The tiers are USASI's own editorial categories, versioned as rubric v0.2. They are not an external certification, and they are not the Open Source Initiative's Open Source AI Definition or any other outside standard. The tier is computed from the recorded evidence and review; it is never inherited from a family or organization.

Training-data information completeness, actual data access, and reuse permission are assessed separately. v0.1 data statuses mixed access and disclosure; those historical notes remain visible as legacy evidence but are excluded from v0.2 tier calculations. New assessments remain unknown until checked. This may lower a computed label without claiming the underlying model changed.

The Open Source AI Definition 1.0 (external site: opensource.org) does not require every original training datum to be downloadable. It requires sufficiently detailed data information, complete training/run/preprocessing code and parameters under qualifying terms. The catalog does not claim OSI certification. A fine-tuning script or inference SDK alone is not complete training code; downloadable data without provenance does not establish complete information.

Component license filters apply the selected license to the selected component in the same license record. An MIT documentation license does not make custom-licensed weights MIT. Multiple distinct license records are not assumed to be alternatives or cumulative terms without review.

Tiers are cumulative: every reviewed open-system release also counts as open-stack, and every open-stack release also counts as open-weight. When tiers appear as columns, the columns overlap and must not be added together.

Hosted product/API is a separate label used on organization pages. It describes how a documented product is delivered, such as a hosted API or assistant app. It is not an artifact tier, and it says nothing about other releases from the same organization. An organization is never labeled "open" because it has released one open-weight model.

Source standards

Each claim cites the source that specifically supports it. In order of preference: the applicable license file; model and dataset cards; official documentation and release notes; the project's repository; the organization's own pages; and regulatory filings for facts such as headquarters. Reputable reporting is used only for the specific claim it supports and is labeled as a news report. An organization's homepage is not adequate evidence for a detailed technical claim.

Every source records its title, publisher, URL, publication date if the page states one, and the date an editor read it. The catalog does not cite search snippets, reproduce source articles, or invent URLs. It does not publish speculative versions, acquisitions, partnerships, staffing, funding, valuations, benchmark scores, or power-capacity figures.

Families and releases

A model family record (for example, a named line of models) summarizes the line and links to its releases. A model release record describes one specific published release. Licenses, availability, checklists, and tiers attach only to releases, because they can differ between releases in the same family. Family overviews are never counted as releases, and a family never carries its own tier.

Software, datasets, and evaluation tools are project records.

Review levels

USASI uses four review levels and does not merge them into a single "verified" badge, because each describes different work.

• Automated checks passed. Every published page passes schema validation, source and date checks, and a link check before it is published. This shows the page is well formed and cites sources; it does not show that each sentence matches its source.
• Source-checked. A separate fact-check pass opened the cited sources and compared each sentence with them, correcting or removing what did not match. In this catalog that pass is currently carried out with AI assistance, and its report is kept in the project's public repository. Pages say when this was done.
• Reviewed by a person. An individual reviewed the page before publication. USASI does not currently claim this level for any page, and the daily news is published without individual human review, as the news section explains.
• Tested hands-on. USASI ran a specific configuration and recorded the date, environment, and result. No record or finder entry currently claims this; finder results say "Not tested by USASI" until a dated test exists.

Dates

• Released — the documented release date of an artifact, at the precision the evidence gives (a year, a month, or a day). Unknown if not documented.
• Entry updated — the date the record was last substantively edited.
• Last reviewed — the date an editor last checked the record's evidence.
• Fact reviewed / effective — optional dates attached to a specific claim, license or availability assessment. An absent date is unknown; the record-wide review is not substituted.
• Source accessed / fetched — retrieval dates; an automated fetch is not an editorial review.
• Site generated — the date the website was built, shown in the footer.

A build never changes editorial, fact, or source dates. "Recently reviewed" on the homepage means recently checked, not newly released.

Uncertainty

Unknown means the item has not been assessed or the evidence is insufficient. It is never a polite way of saying "no." Not public means the evidence documents that something is not available. Where an entry's facts are contested or ambiguous, the record says so in plain language rather than choosing a side.

How counts work

Every total on the site is catalog coverage: the number of published records in this catalog. It is not a census of American AI, and a zero means only that no matching record has been published here.

• Organization records include research units and subsidiaries that have their own records. They are also counted separately as "units and subsidiaries," so a parent and its unit are never presented as two independent companies.
• Model families and model releases are counted separately.
• Artifacts are linked to every organization named on the artifact record, so a release maintained by two organizations appears under both. Per-organization columns are therefore never summed.
• Drafts and archived historical records are excluded from all counts, search, exports, and the sitemap. Archived records keep a clearly labeled historical page so links do not break.

Every number on the comparison workspace links to exactly the records it counts.

Equal structural representation

The two directories — organizations, and open artifacts — have the same prominence in navigation, on the homepage, and in search. Equal prominence does not mean equal numbers: the catalog lists what it can verify, and one directory may be longer than the other. Homepage selections are editorial, explained on the page, alphabetical or date-ordered by default, and never influenced by tips or sponsorships.

News

Latest news items are short, dated updates about organizations and artifacts in the catalog. They are researched and written by an AI assistant working only from primary sources (filings, official announcements, model cards, release notes); each item cites those sources and links to the entries it affects. Since October 1, 2026, a daily automated run finds and publishes new items without individual human review before publication. Every item must pass the site's automated checks, which reject items without sources, with dates in the future, without a related published entry, or containing funding amounts, valuations, staffing or user counts, benchmark scores, or superlatives. Nothing reported only by other outlets is used, and items are never generated by rewriting other publications. Because items are not individually reviewed first, please report any error with the correction link; corrections are made and recorded in the changelog. An event date records when something happened; a published date records when USASI wrote about it. The news list is ordered by event date, newest first, so an item written later about an earlier event appears at the date of the event.

Reading the badges

Status badges always pair an icon with a text label, so no information depends on color. Numbered markers after a statement link to the source list at the bottom of the entry. Links marked with an arrow lead to external sites; everything else stays on USASI.

AI on GitHub: source, ranking, and exclusions

AI on GitHub ranks the most-starred open-source AI projects on GitHub. It is a separate, machine-collected list, not part of the catalog's reviewed directories: entries are not checked for U.S. eligibility, openness, or sources the way catalog records are, and a listing is not an endorsement. Where a project's repository is also linked from a record in Open Models & Tools, its page links to that record.

Source. The list comes from the public GitHub REST search API (/search/repositories), called without an account or token. A refresh searches each of these GitHub topics for repositories with at least 500 stars, sorted by stars: artificial-intelligence, machine-learning, deep-learning, llm, large-language-models, ai-agents, agents, generative-ai, llmops, rag, transformers, pytorch, diffusion, stable-diffusion, computer-vision, nlp, reinforcement-learning, mlops, ai, chatgpt, openai, local-llm, inference, speech-recognition, text-to-speech, multimodal, llama, gpt, neural-network, tensorflow, ai-agent, mcp, claude-code, and vector-database. Because some well-known projects carry none of these topics, it also searches repository names and descriptions for "llm", "machine learning", "deep learning", "language model", "speech recognition", and "neural network". GitHub returns at most 1,000 results per search, so a very broad topic or phrase is cut at its 1,000 most-starred repositories; results from all searches are merged by repository. Searches can still miss notable projects, so a short reviewed must-include list (each entry with a one-line reason) is looked up repository by repository, and so is every GitHub repository that a record in the USASI catalog links to. Being looked up only guarantees a project is considered: it then passes through the same rules and denylist. The one exception is the "no AI signal" rule below, which does not apply to must-includes and to repositories a catalog record names as its own code, because those were reviewed as AI projects already (some have no description or topics on GitHub). Only factual repository fields are stored: name, owner and whether the owner is an organization or a personal account, the description, links, star, fork, and open-issue counts, primary language, license as GitHub detects it, topics, and creation and last-push dates. Each page shows the date of the fetch. The refresh is meant to run daily; between runs the counts shown are from the last fetch.

Ranking. Projects are ranked by stars, highest first, with ties in alphabetical order. Stars measure popularity on GitHub, not quality, safety, security, maturity, or how widely a project is used, and they can be inflated. Open issues as GitHub reports them include pull requests.

What is left out, and why. The list is for working AI projects: software, models, and tools people run or build on. These are excluded by published rules that read only the repository's own name, description, and topics:

• forks, archived repositories, and repositories GitHub has disabled;
• curated link lists ("awesome" lists, or descriptions that say the repository is a list or collection of links, papers, or resources), because they collect other projects rather than being one;
• courses, tutorials, books and book companions, interview preparation, roadmaps, and lecture notes, because they are learning material rather than projects (the Learn section covers learning);
• collections of leaked or extracted system prompts, and jailbreak prompts, because they republish other companies' prompts or exist to bypass safety measures; tools that detect or defend against jailbreaks stay;
• deepfake and face-swap tools, because their main use is impersonating real people; deepfake detectors stay;
• repositories whose name, description, and topics carry no AI signal at all.

A short reviewed denylist covers projects the rules miss; each entry names its reason. Rules can misfire in both directions. Descriptions are the projects' own words, shown as plain text; they are third-party claims, not USASI's. The list includes projects from anywhere in the world. It shows whether an owner is an organization or a personal account and never guesses anyone's country.

Hero images. Each project page has a generated image: abstract artwork drawn from the repository's name (the same name always draws the same picture), with the project's name, owner, owner type, language, and license. It leaves out rank and star counts, which change daily; the page shows the current ones. It is not the project's logo, and no project logos or avatars are used.

AI on Hugging Face: source, ranking, and exclusions

AI on Hugging Face ranks the most-liked models and datasets on the Hugging Face Hub. Like AI on GitHub, it is a separate, machine-collected list, not part of the catalog's reviewed directories: entries are not checked for U.S. eligibility, openness, or sources the way catalog records are, and a listing is not an endorsement. Where a model or dataset is also linked from a record in Open Models & Tools, its page links to that record and the record links back; where the same owner and name are listed in AI on GitHub, the two pages link to each other.

Source. The lists come from the public Hugging Face Hub API (/api/models and /api/datasets, sorted by likes), called without an account or token. A refresh takes the 5,000 most-liked models and the 1,000 most-liked datasets, stopping early where likes fall below 100. Every Hub model and dataset that a record in the USASI catalog links to is also looked up one by one, so catalog items are always considered whatever their likes; they then pass the same rules and denylist. Whether an author is an organization or a personal account comes from the Hub's own organization and user profile endpoints; when neither answers, the page says nothing about it. Only factual fields are stored: the id and author, likes, downloads in the last 30 days as the Hub reports them, the task (pipeline tag) and library, the license from the Hub's license tag or the card's license field, language tags, base-model tags, whether access is gated, creation and last-modified dates, and, for models, a parameter count only when the Hub's safetensors metadata gives one. The only prose is a dataset's short description, the excerpt of its card that the Hub API returns, shown as plain text; model cards are not copied. Each page shows the date of the fetch. The refresh is meant to run weekly.

Ranking. Models and datasets are ranked separately by likes, highest first, with ties in alphabetical order. Likes measure popularity on the Hub, not quality, safety, licensing terms, or how widely something is used, and they can be inflated. Downloads cover the last 30 days and count file requests, not people.

What is left out, and why. These are excluded by published rules that read only the item's own id and Hub tags:

• private, disabled, or removed repositories;
• items tagged not-for-all-audiences, or whose name or tags mark them as explicit content;
• deepfake, face-swap, and undressing tools (deepfake detectors stay);
• fine-tunes that remove safety training, named or tagged "uncensored", "abliterated", or the like, consistent with AI on GitHub leaving out tools that strip models' safety training;
• jailbreak and leaked-prompt collections; datasets and models for detecting, classifying, or evaluating jailbreaks stay.

A short reviewed denylist covers what the rules miss, such as datasets of scraped personal data, explicit content without the tag, and spam or duplicate uploads; each entry names its reason. Rules can misfire in both directions.

Derived models. Quantized re-uploads (GGUF, AWQ, GPTQ, and other formats of another model) stay in the list, because many people run models that way; their pages say "Quantized version of" the base model, from the Hub's base-model tags, and link to the base model's page when it is listed. Fine-tunes, adapters, and merges are labeled the same way, and a base model's page lists the derived models that are in the list. Gated items stay too, marked "Access requires accepting terms on Hugging Face"; a few authors review each request by hand.

Hero images. Each page has a generated image: abstract tile and lattice artwork drawn from the item's kind and id (the same id always draws the same picture), with its name, author, task, and license. It leaves out rank, likes, and downloads, which change daily; the page shows the current ones. It is not the item's or Hugging Face's logo, and no logos or avatars are used.

Jobs (historical editions only)

USASI no longer publishes a job board. Older editions of The State of American AI still include frozen job figures collected from employers' public feeds as of each edition's cutoff. Those numbers are snapshots; they are not refreshed. USASI does not employ, represent, recruit for, or endorse the organizations it catalogs. Applications take place on each employer's official site.

Organization records may still list an official careers URL when the employer publishes one. The careers explainer describes roles from government occupational guides and cited employer postings; it is a learning page, not a listing board.