Methodology / Codebook

Codebook

This is the data dictionary for the AI Trace record: every field in the three tables that make up the record, every value each enum can take, and what those values mean. The definitions are the ones our editorial standards use. The same definitions are published as a machine-readable schema at /datapackage.json, and the pages, the public API, and dataset releases are all generated from this one source.

Tables

The record is three tables. A company has many practices; a practice has many evidence rows. Ids are stable and are the citation keys; a company slug is the current URL and can change.

Companies `companies.csv`

One row per company profile. The id is the stable identifier; the slug is the current URL path and can change when a company is renamed.

Primary key id · parent_company_id references companies.id

Fields of the Companies table
Field	Type	Definition
`id`Company id	string (uuid)	Stable unique identifier for the company. Cite by id: https://www.aitrace.org/r/company/{id} resolves to the current page even after a rename.
`name`Name	string	Company display name.
`slug`Slug	string	URL-safe company identifier, the path segment of the company page. Unique at any one time; may change when a company is renamed, so it is not a stable citation key.
`website`Website	string (uri), nullable	Primary company website URL.
`description`Description	string, nullable	Short description of the company shown on its profile.
`industry_tags`Industry tags	array, nullable	Array of industry tags assigned to the company.
`aliases`Aliases	array, nullable	Other names the company is known by; used to match searches and reports to the profile.
`ai_status`AI status	stringCompany AI status values	Overall company AI status: confirmed_use, suspected_use, no_known_use, or under_review. Set by a moderator on the company record.
`parent_company_id`Parent company id	string (uuid), nullable	Id of the parent company when this company is a subsidiary. Null for top-level companies.
`last_verified_at`Last verified	datetime, nullable	Updated whenever a moderator publishes or edits a practice or a cited source under this company. It records the most recent moderator action on the company's record; it does not mean every claim was re-checked on that date. Null when no moderator action has been recorded.
`created_at`Created	datetime	Timestamp the company record was created.
`updated_at`Updated	datetime	Timestamp the company record was most recently updated.

Practices `practices.csv`

One row per documented AI practice. Every structured field supports an explicit "unknown" or "unclear" value, and every descriptive field may be left null, when sources do not support a specific value. We do not guess.

Primary key id · company_id references companies.id

Fields of the Practices table
Field	Type	Definition
`id`Practice id	string (uuid)	Stable unique identifier for the practice. Cite by id: https://www.aitrace.org/r/practice/{id} resolves to the current page.
`company_id`Company id	string (uuid)	Id of the company the practice belongs to.
`category`Category	stringPractice category values	Practice category, one of seven values; "Other" is the catch-all.Completeness weight: 5 points
`subcategory`Subcategory	string, nullable	Optional free-text subcategory.Completeness weight: 5 points
`summary`Summary	string	One-sentence summary of the practice.
`details`Details	string, nullable	Longer prose description.
`products_affected`Products affected	array, nullable	Array of product or service names touched by this practice. Empty when none are named.
`status`Status	stringPractice status values	Editorial status: verified, reported, disputed, or retracted.
`date_practice_started`Date started	string, nullable	Earliest known date the practice began. A partial date (a year, or a year and month) is allowed; null when sources do not say.Completeness weight: 5 points
`discontinued_at`Discontinued	date, nullable	Date the practice was discontinued. Null when still active. Discontinuation is an explicit flag, not a deletion: the entry stays in the record.
`last_reviewed_at`Last reviewed	datetime, nullable	Timestamp of the most recent moderator review of this practice.
`deployment_scope`Deployment scope	stringDeployment scope values	Access conditions for the AI system. The value "unknown" is the sentinel default and earns zero points in the completeness score. Orthogonal to visibility; never infer one from the other.Completeness weight: 10 points
`deployment_regions`Deployment regions	string, nullable	Free-text geographic scope (examples: "US only", "60+ countries", "global"). Null when sources do not clarify. For scoring purposes, the literal string "unknown" is treated the same as null.Completeness weight: 5 points
`user_base_affected`User base affected	string, nullable	Free-text scale and population affected (example: "all 500M monthly users", "enterprise customers only"). Null when sources do not clarify. For scoring purposes, the literal string "unknown" is treated the same as null.Completeness weight: 10 points
`ai_function`AI function	arrayAI function values	Array of AI functions; the primary function is listed first. Sentinel default is ["other"]. An array containing only "other" earns zero points in the completeness score.Completeness weight: 10 points
`ai_input`AI input	string, nullable	Free-text description of what the AI system receives as input.Completeness weight: 10 points
`ai_output`AI output	string, nullable	Free-text description of what the AI system produces.Completeness weight: 10 points
`labor_impact_type`Labor impact type	stringLabor impact type values	Labor consequence of the AI system. Sentinel default is "unclear", which earns zero points. Setting "automates_prior_task" requires explicit two-sided evidence at the specific company.Completeness weight: 10 points
`prior_human_process`Prior human process	string, nullable	Free-text description of what humans did before the AI system existed. Null when labor_impact_type is new_capability or when sources do not clarify. Populating this field earns its 10 points even when labor_impact_type remains unclear.Completeness weight: 10 points
`labor_context`Labor context	string, nullable	Free-text nuance about the labor situation that does not fit the enum cleanly (example: "editors still review before publish"). Documenting a gap here earns its 5 points even when labor_impact_type is unclear.Completeness weight: 5 points
`visibility`Visibility	string, nullableVisibility values	Whose experience includes the AI output. Null is the sentinel and earns zero points. Orthogonal to deployment_scope; both fields must be set independently based on what the sources support.Completeness weight: 5 points
`completeness_score`Completeness score	integer	Derived 0 to 100 editorial completeness score, maintained automatically whenever the practice is written. Purely additive; zero is a valid score. Sentinel values ("unknown", "unclear", a sole ["other"], null) earn zero points for their respective fields. A low score due to honest unknowns is acceptable. It is never authored by hand and is not an editorial judgment.
`created_at`Created	datetime	Timestamp the practice record was created.
`updated_at`Updated	datetime	Timestamp the practice record was most recently updated.

Evidence `evidence.csv`

One row per cited source. Every claim in a practice is traceable to at least one row here; a published practice has at least one accepted source, and at least one Tier 1 or Tier 2 source.

Primary key id · practice_id references practices.id

Fields of the Evidence table
Field	Type	Definition
`id`Evidence id	string (uuid)	Evidence row unique identifier.
`practice_id`Practice id	string (uuid)	Id of the practice this source supports.
`source_type`Source type	stringSource type values	Type of source: news_article, company_disclosure, regulatory_filing, community_report, academic_paper, social_media_post. Each type maps to a tier.
`source_url`Source URL	string (uri)	Canonical URL of the source.
`archived_url`Archived copy	string (uri), nullable	Wayback or archive URL, when captured. Null when the page could not be archived (for example, the publisher blocks archiving or the page was gone before capture).
`source_title`Title	string	Title of the article, document, or post.
`source_publisher`Publisher	string, nullable	Publisher or platform hosting the source.
`source_date`Date published	date, nullable	Publication date of the source.
`excerpt`Excerpt	string, nullable	Evidence excerpt relevant to the claim.
`verified`Moderator verified	boolean	Whether a moderator has manually verified this source. A source is marked moderator-verified only when three conditions are met simultaneously: the source is from the company itself, the URL is on a recognizable official domain, and the content directly confirms the specific practice being documented. This flag is applied at the source level, not the practice level.
`created_at`Created	datetime	Timestamp the evidence row was created.

Records `records.csv`

One public record about a company: a regulator action, a lawsuit, a conduct and harm finding, an incident, a lobbying spending record, a government contract record, a safety evaluation record, a training data statement, a safety statement, a model release, a user-data terms record, an acceptable use policy record or a licensing deal record, each quoting the words of its source. The records export in JSON serves one object per record, and the company route serves the company's records when include=records is asked for. A published record can carry fields that failed their check; each such field is shown as null.

Primary key record_id · company_slug references companies.slug

Fields of the Records table
Field	Type	Definition
`record_id`Record id	string (uuid)	Stable unique identifier for the record. The export names this field record_id and the company route names it id; the value is the same.
`company_slug`Company slug	string	The slug of the company the record is about, as the companies table holds it. Served by the export only, since the company route serves one company's records.
`kind`Kind	stringRecord kind values	The kind of record, which sets its fields and its section on the company page. Served by the export and the company route.
`source_class`Source class	stringSource class values	Where the quoted words come from: the company itself, a regulator or another public body, a report by an outlet, or an independent evaluator. Served by the export and the company route.
`record_date`Record date	string, nullable	The date of the action, report or statement, taken from the record's date field. A partial date (a year, or a year and month) is allowed; null when the source gives no date or the date failed its check. Served by the export and the company route.
`outlet`Outlet	string, nullable	The name of the outlet for a record whose source class is attributed_report, the name of the register for a lobbying spending record, or the evaluation source of a safety evaluation record (institute, independent or self), and null for every other record or when none is stored. Served by the export and the company route.
`finding_seen_on`Page last seen	date, nullable	The date the page holding the record's quoted words was last fetched. Null when that date is not known. Served by the export and the company route.
`created_at`Created	datetime	Timestamp the record was created. Served by the export only.
`updated_at`Updated	datetime	Timestamp the record was most recently updated. Served by the export only.
`fields`Fields	object	The record's fields by name, one key for each field of the record's kind, each holding a field object with the value, quote and source_url of the record fields table, or null. A null field means the field failed its check or was not stated. Served by the export and the company route.

Record fields `record_fields.csv`

One row per field of a record that is not null, in the records export in CSV, which repeats the record's own columns on every row. A record whose fields are all null has one row with the field, value, quote and source_url cells empty, so this table has no primary key. In JSON the value, quote and source_url of a field form the field object under a record's fields, in the export and in the company route.

Primary key · record_id references records.record_id · company_slug references companies.slug

Fields of the Record fields table
Field	Type	Definition
`record_id`Record id	string (uuid)	Id of the record the row belongs to, the record_id of the records table. A column of the CSV file.
`company_slug`Company slug	string	The slug of the company the record is about. A column of the CSV file.
`kind`Kind	stringRecord kind values	The kind of the record. A column of the CSV file.
`source_class`Source class	stringSource class values	The source class of the record. A column of the CSV file.
`record_date`Record date	string, nullable	The record date of the record, as in the records table. A column of the CSV file.
`outlet`Outlet	string, nullable	The outlet of an attributed report, the register of a lobbying spending record or the evaluation source of a safety evaluation record, as in the records table. A column of the CSV file.
`finding_seen_on`Page last seen	date, nullable	The date the page holding the record's quoted words was last fetched, as in the records table. A column of the CSV file.
`field`Field	string, nullable	The field's name, one of the fields of the record's kind (for example regulator, case_status or release_text; the record kind definitions list every kind's fields). Empty on the one row of a record whose fields are all null. A column of the CSV file; in JSON the name is the key under the record's fields.
`value`Value	string, nullable	The field's value. In the CSV file it is text: a list of values is joined by a semicolon and a space, and a yes or no value reads true or false. In JSON it is a string, a number, true or false, or a list of those, and null when the stored value is none of these.
`quote`Quote	string, nullable	The words of the source, quoted exactly, that support the value. A column of the CSV file and a key of the JSON field object.
`source_url`Source address	string, nullable	The address of the page the quote was taken from, as stored. A column of the CSV file and a key of the JSON field object.

None-found lines `none_found_lines.csv`

One line per company and record kind where the latest search found no record, while the company holds no published record of that kind. The none_found export serves one row per line, and the company route serves the company's lines when include=records is asked for. Each line states, with dates, where the search looked and found none, and which sites it did not search. A line with no sentence the company page shows is not served.

Primary key company_slug, kind · company_slug references companies.slug

Fields of the None-found lines table
Field	Type	Definition
`company_slug`Company slug	string	The slug of the company the line is about. Served by the export only, since the company route serves one company's lines.
`kind`Kind	stringRecord kind values	The kind of record the search looked for. Served by the export and the company route.
`decided_at`Decided	datetime	Timestamp the line was decided. Served by the export and the company route.
`sentences`Sentences	string	The line's sentences that the company page shows, in stored order. In the CSV file, the text of each sentence, joined by a newline. In JSON, a list of sentence objects whose keys are the sentences fields below.
`sentences.type`Sentence type	stringNone-found sentence type values	The sentence's none-found sentence type. A key of each sentence object in JSON.
`sentences.site`Sentence site	string, nullable	The site the sentence is about, or null when none is stored. An own page sentence carries the site of the page, though its text names the address. A key of each sentence object in JSON.
`sentences.address`Sentence address	string, nullable	The address the sentence rests on when one is stored: the page read for an own page sentence, the home page read for a home page sentence, the address asked or watched for the other types. Null when none is stored. A key of each sentence object in JSON.
`sentences.date`Sentence date	string, nullable	The date the sentence states as year, month and day (the last date of the span for a watched sentence), or null. A key of each sentence object in JSON.
`sentences.name`Sentence name	string, nullable	The company name the search looked for, or null. A key of an asked, a newest, a watched or a home page sentence object in JSON, and absent from an own page sentence.
`sentences.start`Sentence start	string, nullable	The first date of the span a watched sentence states, as year, month and day, or null. A key of a watched sentence object in JSON, and absent from the other types.
`sentences.count`Sentence count	integer, nullable	The number of newest notices the site answered with, none of which names the company, or null. A key of a newest sentence object in JSON, and absent from the other types.
`sentences.looked_for`Sentence looked for	string, nullable	What the home page was searched for a link to, such as a privacy policy, terms or legal page, or null. A key of a home page sentence object in JSON, and absent from the other types.
`sentences.sitemap`Sentence sitemap	string, nullable	The address of the sitemap read for a home page sentence, or null when none was read. A key of a home page sentence object in JSON, and absent from the other types.
`sentences.text`Sentence text	string	The sentence as the company page shows it, with its dates written out. A key of each sentence object in JSON; the CSV file carries this text in its sentences column.
`not_searched`Not searched	array, nullable	The sites on the kind's list of sites that earned no sentence, in stored order, each an object whose keys are the not_searched fields below. Served in JSON only, by the export and the company route.
`not_searched.site`Not-searched site	string, nullable	The site that earned no sentence, or null. A key of each not-searched object in JSON.
`not_searched.reason`Not-searched reason	string, nullableNot-searched reason values	Why the site earned no sentence, or null. A key of each not-searched object in JSON.
`rendered_not_searched`Not searched, as shown	string, nullable	The not-searched list as the company page shows it, one entry per site with its reasons, or null when every site earned a sentence. A column of the CSV file and a key of each line in JSON.

Coverage statements `/api/v2/companies/[slug]`

One statement per date for a company that holds no published record of a record kind the site shows and at least one none-found line the company page shows. The company route serves the statements as its coverage key when include=records is asked for, newest date first, and serves null for any other company and when the read of the company's records fails; the company page shows the same statements. Each statement names the date the record kinds were searched, the sites searched and the sites whose feeds or sitemaps were read, and says that nothing matched the record kinds on that date. It describes the searches only, not whether the company uses AI.

Primary key pass_date

Fields of the Coverage statements table
Field	Type	Definition
`pass_date`Pass date	date	The date the statement's none-found lines were decided, as year, month and day in UTC: the date part of their decided_at. Lines of different kinds decided on different dates give one statement per date.
`kinds`Kinds	arrayRecord kind values	The record kinds whose latest none-found line was decided on the pass date, in the order of the company page.
`searched`Sites searched	array	The sites of the statement's asked sentences, each once, in first-seen order. Empty when no asked sentence is shown.
`watched`Feeds or sitemaps read	array	The sites of the statement's watched sentences, whose feeds or sitemaps were read, each once, in first-seen order. Empty when no watched sentence is shown.
`not_searched`Not searched	array	The sites on the kinds' lists of sites that earned no sentence, one object per site and reason, in first-seen order, each with the not_searched fields below.
`not_searched.site`Not-searched site	string	The site that earned no sentence. A key of each not-searched object.
`not_searched.reason`Not-searched reason	stringNot-searched reason values	Why the site earned no sentence. A key of each not-searched object.
`statement`Statement	array	The statement's sentences as the company page shows them, in order: the date of the search and the company, the sites searched when there are any, the feeds or sitemaps read when there are any, and the sentence that nothing matched the record kinds on that date.

Enum values

Stored values are what the API and release files carry. Labels are what the site renders. Where a value is marked sentinel, it is the explicit record that we looked and the sources do not say; a sentinel earns zero points in the completeness score, and an entry with honestly unknown fields is more defensible than one filled with plausible guesses.

Company AI status `ai_status`

The company-level status shown on the company badge. A moderator sets it on the company record; it is set independently of the status of any one practice.

Values of Company AI status
Stored value	Label	Definition
`confirmed_use`	Confirmed AI Use	The company badge reads Confirmed AI Use. Descriptive, not a judgment: the site tracks AI use and does not rate it.
`suspected_use`	Suspected AI Use	The company badge reads Suspected AI Use.
`no_known_use`	No Known AI Use	The company badge reads No Known AI Use.
`under_review`	Under Review	The company badge reads Under Review. A company record created by the automated discovery pipeline carries this status until a moderator reviews it.

Practice status `practice_status`

The editorial status of one practice entry. Verified, Reported and In re-verification come from the label rule, computed from the stored sources and claims; Disputed and Retracted come from the dispute and correction process.

Values of Practice status
Stored value	Label	Definition
`verified`	Verified	A practice is marked "Verified" when the label rule's conditions all hold: its corroboration level is 80 or more, every claim is verified against its source, and no source is on hold. The corroboration level scores the best source tier present (Tier 1 90, Tier 2 70, Tier 3 30) and adds 10 when two or more accepted sources come from distinct publishers, to a cap of 100.
`reported`	Reported	A practice is marked "Reported" when the label rule's conditions are not all met. Its sources are shown as they stand.
`withheld`	In re-verification	A practice is marked "In re-verification" when it was Verified and its verification is being re-read. The reason is stated on its page, and the record is kept.
`disputed`	Disputed	If an entry is disputed, it carries a "Disputed" badge until the dispute is resolved. Contested information is not silently removed.
`retracted`	Retracted	An entry that turns out to be materially wrong is marked Retracted rather than deleted: the record of what we got wrong stays public.

Practice category `practice_category`

The kind of AI practice. One of seven values; "Other" is the catch-all when no specific category fits.

Values of Practice category
Stored value	Label	Definition
`creative_generation`	Creative Generation	AI used to generate creative material.
`productivity_automation`	Productivity Automation	AI used to automate or speed up work tasks.
`data_analysis`	Data Analysis	AI used to analyze data.
`content_moderation`	Content Moderation	AI used to moderate content.
`recommendation_system`	Recommendation System	AI used to recommend items to people.
`customer_service`	Customer Service	AI used in customer service.
`other`sentinel	Other	The catch-all when no specific category fits.

Deployment scope `deployment_scope`

Access conditions for the AI system: who can access it. Orthogonal to visibility; never infer one from the other. The value "unknown" is the sentinel default and earns zero points in the completeness score.

Values of Deployment scope
Stored value	Label	Definition
`all_users`	All Users	Available to all users of the product or service.
`opt_in`	Opt-in	Available to users who choose to turn it on.
`pilot`	Pilot	Deployed to a limited pilot group.
`internal_tool`	Internal Tool	Used inside the company by its own staff.
`offered_not_mandated`	Offered, Not Mandated	Offered to staff or users without being required.
`unknown`sentinel	Unknown	The sentinel used when sources do not clarify; it earns zero points in the completeness score.

AI function `ai_function`

The technical mechanism: what the model does, independent of labor impact. Stored as an array with the primary function listed first. The value "other" is the sentinel used when no specific function fits; an array containing only "other" earns zero points in the completeness score.

Values of AI function
Stored value	Label	Definition
`generating`	Generating	Produces new content.
`classifying`	Classifying	Assigns items to categories.
`matching`	Matching	Pairs or links items to one another.
`summarizing`	Summarizing	Condenses longer material.
`recommending`	Recommending	Ranks or suggests items.
`detecting`	Detecting	Finds patterns, anomalies, or specific content.
`predicting`	Predicting	Estimates future or unobserved values.
`automating`	Automating	Carries out a task end to end. Shares a word with the labor impact value automates_prior_task but describes the mechanism, not the labor consequence; the two are set independently.
`other`sentinel	Other	The sentinel used when no specific function fits.

Labor impact type `labor_impact_type`

Labor consequence of the AI system, fully independent of the AI function. The value "unclear" is the sentinel and earns zero points. The value "automates_prior_task" requires explicit two-sided evidence at the specific company: one side documents the prior human process, the other confirms that AI now performs that task.

Values of Labor impact type
Stored value	Label	Definition
`augments_existing`	Augments Existing Work	AI assists work that humans continue to do.
`automates_prior_task`	Automates Prior Task	AI performs a task that humans at the company previously performed. Requires two-sided evidence; where only one side is confirmed, or where the prior human process is an industry-wide assumption rather than a company-specific fact, the value is unclear.
`new_capability`	New Capability	AI enables something that was not previously done by humans at the company.
`unclear`sentinel	Unclear	The sentinel used when sources do not support a specific classification; it earns zero points.

Visibility `visibility`

Whose experience includes the AI output. Null is the sentinel and earns zero points. Orthogonal to deployment scope; both fields must be set independently based on what the sources support.

Values of Visibility
Stored value	Label	Definition
`consumer_facing`	Consumer-facing	Customers or the public experience the AI output.
`internal`	Internal	Only people inside the company experience the AI output.
`both`	Both	Both customers and internal staff experience the AI output.

Source type `source_type`

The kind of cited source. Each type maps to a tier: Tier 1 (Company Disclosure, Regulatory Filing) is material from the company itself or filed with a government and can substantiate a claim alone; Tier 2 (News Article, Academic Paper) is reporting from established outlets and peer-reviewed research; Tier 3 (Social Media, Community Report) is never sufficient alone.

Values of Source type
Stored value	Label	Definition
`company_disclosure`	Company Disclosure	Tier 1. Official statements, press releases, engineering blogs and documentation on official domains.
`regulatory_filing`	Regulatory Filing	Tier 1. Filings with the SEC or another regulator, and pages a regulator publishes on its own website, such as an order, a decision or a notice of an inquiry.
`news_article`	News Article	Tier 2. Reporting from established outlets with editorial standards and named authorship.
`academic_paper`	Academic Paper	Tier 2. Peer-reviewed academic research.
`social_media_post`	Social Media	Tier 3. Social media posts; screenshots are required.
`community_report`	Community Report	Tier 3. Community reports and forums.

Provenance status `provenance_status`

Authorship of the current value of a tracked practice field. Rendered on each practice page as the provenance line; not a column in the release files. The tracked fields are summary, details, category, subcategory, products affected, visibility, labor impact type, deployment scope, deployment regions, user base affected, AI function, AI input, AI output, prior human process, and labor context.

Values of Provenance status
Stored value	Label	Definition
`ai_drafted`	AI drafted	The current value was drafted by the automated research pipeline and accepted by a human reviewer without edits.
`human_edited`	Human edited	A human reviewer wrote or changed the current value.
`human_verified`	Human verified	A human reviewer verified the current value against the cited sources.

Record kind `record_kind`

The kind of a public record about a company. Each kind has its own fields and its own section on the company page; only the kinds listed here are published. Each definition names the kind's fields, the keys of a record's fields object.

Values of Record kind
Stored value	Label	Definition
`regulator_action`	Regulator actions	An action a regulator took concerning the company, such as an inquiry, an order or a fine, with the words of the regulator's page or of a report that names the regulator. Its fields are regulator, respondent, action_type, action_text, amount, jurisdiction, action_status and action_date.
`lawsuit`	Lawsuits	A case in a court that names the company, with the words of the court record, the court, the parties, the claims and the status of the case as the record states it. Its fields are case_name, court, docket_number, party, case_status, case_text, claim_text, ruling_text, filed_date and status_date.
`conduct_finding`	Conduct and harm findings	A conclusion a court, a government inquiry or an official hearing reached about what the company did, with the words of the body's page. Its fields are body, respondent, body_type, finding_text, finding_status, jurisdiction and finding_date.
`incident`	Incidents	A harm reported in connection with an AI system of the company, with the words of the page that reports it. Its fields are system_name, harm_text, incident_date, region and report_date.
`lobbying_spending`	Lobbying spending	A filing in a public lobbying register that states what was spent on lobbying for the company, with the register's words: the client, the filer, the period, the amount as the register states it and the issues lobbied on. Its fields are client_name, registrant_name, filing_period, amount_basis, amount, spending_text, issues_text and filing_date.
`government_contract`	Government contracts	A public notice of a contract a government body awarded to the company, with the notice's words: the awarding body, the title, the value as the notice states it and the award date. Its fields are awarding_body, supplier, contract_title, award_text, value and award_date.
`safety_evaluation`	Safety evaluations	A published evaluation of a model of the company, with the words of the page that publishes it: the evaluating body, the developer, the model and the finding. The page is a public institute's, an independent evaluator's on a fixed, measured list, or the company's own. Its fields are evaluating_body, developer_name, model_name, evaluation_text and evaluation_date.
`training_data`	Training data	What the company states about the data a model of its was trained on, with the words of the company's own page. Its fields are model_name, statement, acquisition_methods, data_cutoff, opt_out_offered and statement_date.
`safety_statement`	Safety statements	A dated statement or commitment by the company about the safety of its AI, with the words of the company's own page or of a government page that lists the company as a signatory. Its fields are commitment_name, statement, statement_type, addressee and statement_date.
`model_release`	Model releases	A model the company released, with its date, its access and its documents, in the words of the company's own page. Its fields are model_name, release_text, access_type, documents and release_date.
`user_data_terms`	User-data terms	What the company's terms of service or privacy policy state about whether the content users give its AI products trains models, and the route to opt out, with the words of the company's own page in dated versions. Its fields are policy_name, service_name, training_use, training_text, opt_out_text and version_date.
`acceptable_use_policy`	Acceptable use policies	A use of the company's AI products that its usage policy bans, or an enforcement figure the company reports for that policy, with the words of the company's own page. Its fields are policy_name, service_name, record_part, policy_text, enforcement_figure, reporting_period and policy_date.
`licensing_deal`	Licensing deals	A deal the company announces with another party to license content or to work together, with the words of the company's own announcement: the other party, the content licensed and the announcement date. Its fields are company_party, other_party, content_text, deal_text and announcement_date.

Source class `source_class`

Where the words a record quotes come from. Every record carries one class, and the company page shows its label above the record.

Values of Source class
Stored value	Label	Definition
`company_statement`	Company statement	The company's own words on its own page. Such a record states what the company says; it is not a finding about the company.
`official_record`	Official record	The words of a page a regulator, a court or another public body publishes. The company page names the regulator, the court, the body or the register, or the site of the page when the record names none; a government page that lists the company as a signatory reads as that body listing the company.
`attributed_report`	Attributed report	The words of a report by a named outlet about the company. The company page shows it with the name of the outlet or, when no name is stored, the host of the source.
`independent_evaluation`	Independent evaluation	The words of a page an independent evaluator on a fixed, measured list publishes about a company's model; the company page names the evaluator.

None-found sentence type `sentence_type`

The kind of search one sentence of a none-found line reports. A none-found line states, with dates, where the search for records of one kind about a company looked and found none.

Values of None-found sentence type
Stored value	Label	Definition
`asked`	Asked	On the date the sentence gives, the site's own search was asked for the company by name and answered zero results.
`newest`	Newest notices	On the date the sentence gives, the site answered with its newest notices, as many as the sentence gives, and none names the company. The site does not search by name, so the sentence never states zero results.
`watched`	Watched	No item whose title, summary or address names the company appeared in the site's feed or sitemap between the two dates the sentence gives.
`own_page`	Own page	On the date the sentence gives, the page it names was read and no record of the kind was drafted from it.
`home_page`	Home page	On the date the sentence gives, the company's home page was read and held no link to the page the kind looks for; when the sentence names a sitemap, the sitemap held no entry for one either.

Not-searched reason `not_searched_reason`

Why a site on the list of sites for a record kind earned no sentence in a none-found line. The line lists each such site once with its reasons; the stored value is also the text the company page shows.

Values of Not-searched reason
Stored value	Label	Definition
`failed fetch`	failed fetch	The site's page, search or feed could not be fetched.
`robots.txt`	robots.txt	The site's robots.txt file does not allow the fetch, and the search obeys it.
`answered with results`	answered with results	The site's search answered with results, so it cannot support a sentence that it found none.
`results did not name the company`	results did not name the company	The site's search answered with results, and none of them names the company.
`unproven`	unproven	The site's answer could not be shown to be a true answer of zero results.
`no dated item`	no dated item	The site's feed or sitemap holds no dated item, so no span of dates can be stated for it.
`press`	press	The site is a press outlet, and a press outlet earns no sentence in a none-found line.
`no channel`	no channel	The site offers no search, feed or sitemap that the search can read.
`not polled`	not polled	The site's feed or sitemap was not read for the span the line covers.

Statement type `statement_type`

The kind of statement a safety statement record quotes. The value of the statement_type field.

Values of Statement type
Stored value	Label	Definition
`signed_commitment`	Signed commitment	A voluntary commitment the company signed, or a government page that lists the company as a signatory to one.
`safety_framework`	Safety framework	A safety framework the company published.
`policy_promise`	Policy promise	A promise the company made in a published policy.

Access type `access_type`

How a released model can be reached, as the company's page states it. The value of the access_type field of a model release record, which is a list and can hold more than one access type.

Values of Access type
Stored value	Label	Definition
`open_weights`	Open weights	The model's weights are published for anyone to download.
`api`	API	The model can be used through a programming interface the company offers.
`product`	Product	The model can be used through a product of the company.

Case status `case_status`

The status of a lawsuit as the court record states it. The value of the case_status field; the company page shows it with the court record's words beside it.

Values of Case status
Stored value	Label	Definition
`filed`	Filed	The case was filed, and the court record states no later outcome.
`dismissed`	Dismissed	The court dismissed the case.
`settled`	Settled	The parties settled the case.
`judgment`	Judgment	The court entered a judgment in the case.
`appealed`	Appealed	A party appealed a decision in the case.

Body type `body_type`

The kind of official body whose conclusion a conduct and harm finding record quotes. The value of the body_type field.

Values of Body type
Stored value	Label	Definition
`court_judgment`	Court judgment	The judgment of a court.
`government_inquiry`	Government inquiry	The conclusion of an inquiry by a government body.
`official_hearing`	Official hearing	The conclusion of an official hearing.

Training use `training_use`

What a user-data terms record states about training models on user content. The value of the training_use field.

Values of Training use
Stored value	Label	Definition
`trains_on_content`	Trains on user content	The terms state that user content is used to train models.
`does_not_train`	Does not train on user content	The terms state that user content is not used to train models.
`trains_unless_opted_out`	Trains on user content unless the user opts out	The terms state that user content is used to train models unless the user opts out.

Record part `record_part`

Which part of an acceptable use policy a record quotes. The value of the record_part field.

Values of Record part
Stored value	Label	Definition
`prohibited_use`	Prohibited use	A use the policy bans.
`enforcement`	Enforcement	An enforcement figure the company reports for the policy.

Amount basis `amount_basis`

What the figure of a lobbying spending record is, as the register states it. The value of the amount_basis field.

Values of Amount basis
Stored value	Label	Definition
`income`	Lobbying firm's income from the client	The figure is the lobbying firm's income from the client for the period.
`expenses`	Client's own lobbying expenses	The figure is the client's own lobbying expenses for the period.

Completeness score

Each practice carries a completeness score from 0 to 100. It is computed, not guessed: it reflects actual non-default field values per the weights below, is maintained automatically whenever a practice is written, and is never authored by hand. It measures how much the cited evidence supports, not how good or bad the practice is. A low score due to honest unknowns is acceptable.

Completeness score weights by field
Field	Points
Category `category`	5
Subcategory `subcategory`	5
Date started `date_practice_started`	5
Deployment scope `deployment_scope`	10
Deployment regions `deployment_regions`	5
User base affected `user_base_affected`	10
AI function `ai_function`	10
AI input `ai_input`	10
AI output `ai_output`	10
Labor impact type `labor_impact_type`	10
Prior human process `prior_human_process`	10
Labor context `labor_context`	5
Visibility `visibility`	5
Total	100

Machine-readable schema

The same dictionary is published as a Frictionless Data Package at /datapackage.json: one Table Schema per table, with field names, types, enum constraints, primary keys, and foreign keys. Dataset releases ship this file beside the CSVs it describes, so a release can be validated against the published schema. The package carries a version once the first dataset release exists; until then the field is absent rather than invented.

For citation forms and the data license, see How to cite. For what the record does not cover, see Known limitations.

Tables

Companies companies.csv

Practices practices.csv

Evidence evidence.csv

Records records.csv

Record fields record_fields.csv

None-found lines none_found_lines.csv

Coverage statements /api/v2/companies/[slug]

Enum values

Company AI status ai_status

Practice status practice_status

Practice category practice_category

Deployment scope deployment_scope

AI function ai_function

Labor impact type labor_impact_type

Visibility visibility

Source type source_type

Provenance status provenance_status

Record kind record_kind

Source class source_class

None-found sentence type sentence_type

Not-searched reason not_searched_reason

Statement type statement_type

Access type access_type

Case status case_status

Body type body_type

Training use training_use

Record part record_part

Amount basis amount_basis