Overview

What is this dataset?

The AGORA (AI GOvernance and Regulatory Archive) dataset is a living collection of AI-relevant laws, regulations, standards, and other governance documents from the United States and around the world. AGORA includes document text, metadata, summaries, and thematic tags to enable deep analysis of the global AI governance landscape.

How do I get it?

The dataset is available on Zenodo.

Which ETO products use it?

The AGORA dataset powers ETO's AGORA tool.

What are its sources?

The AGORA dataset is an original Emerging Technology Observatory resource. Document text and some metadata are taken from official sources. Other metadata, summaries, and tags are produced by ETO analysts and annotators.

What are its main limitations?

• The dataset's focus on documents that directly address AI means that many AI-relevant laws, regulations, and norms are excluded by design. In particular, laws of general applicability - for example, securities regulation, civil rights law, or common-law tort doctrines - are excluded. Read more >>
• Many documents within AGORA's scope are not yet included. AGORA’s nominal scope is broader than the set of documents collected to date. In particular, the current dataset skews toward U.S. law and policy. New data are added regularly and we aim to broaden coverage over time. Read more >>
• Some data is machine-generated and may contain errors. We use machine learning to draft summaries of AGORA documents. The machine output is included in the dataset once generated; it may take some time after that for a human annotator to review. We prominently flag unreviewed machine-generated output in the dataset. Read more >>
• Annotators exercise judgment in applying AGORA’s scope and taxonomy, raising the risks of inconsistent screening and tagging. We try to mitigate these risks in several ways. Read more >>

What are the terms of use?

This dataset is subject to ETO's general terms of use. If you use it, please cite us.

Note that the dataset includes full text of AGORA documents, taken from sources such as government websites and repositories. Given its nature, we believe all of this material is open to non-commercial use consistent with our general terms of use, but we make no warranties.

Does it contain sensitive information, such as personally identifiable information?

No.

How do I cite it?

Please cite the "Emerging Technology Observatory AGORA dataset," including the link.

If you use AGORA to access the data, you can cite that tool instead.

Structure and content

The AGORA dataset consists of the csv tables documents, segments, collections, and authorities and a folder of text documents called fulltext.

documents

This table includes core metadata, summaries, and thematic tags for AGORA documents.

segments

This table includes metadata, summaries, and thematic tags for individual segments of AGORA documents.

collections

This table lists and describes the collections AGORA documents are assigned.

authorities

This table lists and describes the authorities associated with AGORA documents.

fulltext

The "fulltext" folder includes individual files with the full text of each AGORA document for which full text has been collected. Each file is labeled with the AGORA ID of the corresponding document.

Sources and methodology

Scope

AGORA's nominal scope

The AGORA dataset includes laws, regulations, standards, and similar documents that directly and substantively address the development, deployment, or use of artificial intelligence technology. The intent of this scoping definition is to encompass the large majority of documents created by lawmakers, regulators, and standard-setters in direct response to advances in modern machine learning and related technologies.

Applying subjective elements of this definition, such as “directly and substantively,” inevitably involves judgment. When screening documents for inclusion in AGORA, we try to constrain this judgment by defining heuristics.

Critically, the requirement that documents “directly” address artificial intelligence generally excludes laws predating the rise of modern machine learning, even if they are broad enough in scope to bear on AI. We draw this line to ensure that AGORA’s scope is manageable in practice and to reinforce the dataset’s emphasis on policies created in response to 21st century developments in AI, rather than the entire set of policy documents that may affect individual sectors and governance writ large. Note, however, that more recent documents that tailor these broad laws to the specific context of AI would qualify for inclusion in AGORA. For example, while the Civil Rights Act of 1964 would not be included in AGORA, a related federal regulation or guidance document applying the Act to racially discriminatory AI is within AGORA’s scope.

Current coverage

AGORA’s nominal scope is broader than the set of documents collected to date. In particular, the current dataset skews toward U.S. law and policy. New data are added regularly and we plan to broaden coverage over time.

For now, AGORA aims to include the following documents (to the extent they are within AGORA's scope) with a lag of no more than a few months:

• United States federal documents - all enacted and proposed federal laws since 2020; all enacted and proposed federal regulations since 2020; all enacted executive orders since 2020; other agency documents of major popular or scholarly interest.
• United States state documents - currently collected ad hoc; not necessarily comprehensive or representative.
• Other documents - currently collected ad hoc; not necessarily comprehensive or representative.

Our next priorities are to broaden coverage of U.S. state documents, with the aim of including all enacted, in-scope state laws on a going-forward basis, and to broaden coverage of Chinese central government documents and major corporate commitments. We will update this section as we make progress toward these goals.

Collection, screening, and initial enrichment

Candidate documents for inclusion in AGORA are currently collected manually or using semi-automated means (e.g., saved queries against larger datasets) from a wide range of official and unofficial sources, reflecting the decentralized, largely ad-hoc status of current AI governance tracking. The most common sources include official, general-purpose regulatory compilations, such as the Congress.gov service for United States federal legislation and the Federal Register for United States federal regulation. Human screeners review these sources (periodically, in the case of sources that update) and assess each document against the AGORA scoping definition.

Screeners first determine whether documents are within AGORA's scope. Only in-scope documents are included in the AGORA dataset. For documents determined to be in scope, screeners locate the authoritative text of the document (for example, on the official website of the United States Congress or a state legislature) and use it to populate basic metadata such as title and date of introduction. Whenever feasible, they also archive a plaintext copy of the document and divide it into shorter segments (often, but not always, corresponding to sections, subsections, or similar divisions already present in the text) for granular annotation.

Annotators also identify “packages,” or larger, thematically diverse documents containing AI-related portions amidst other, AI-unrelated material. A typical example is the annual National Defense Authorization Act (NDAA) in the United States, a massive, largely AI-unrelated law with some diverse AI-related provisions sprinkled throughout in recent years. NDAAs and other such packages are divided into conceptually discrete AGORA documents, corresponding to sections, subsections, or other subdivisions in the packages, according to standing guidance (reproduced in the appendices).

Finally, screeners identify a canonical AGORA authority for the document and assign it to one or more thematic collections. AGORA authorities and collections are listed in the authorities and collections tables, respectively.

Significance ratings

During the screening process, AGORA screeners assign each document a significance rating. This rating is used as a rough basis for sorting documents during the subsequent annotation process and in the AGORA interface. It should not be relied on for rigorous analysis.

A significance rating is a screener's subjective assessment of how important the document is (if not yet in effect, assuming it were). Screeners use the following guidelines:

• Lower significance: Has a minor impact on a limited group of people; makes slight changes to the existing rights, duties, or resources of individuals or organizations. Of niche interest.
• Moderate: Materially changes the status quo for a sizable number of people and/or organizations, such as businesses in a particular economic sector or a large government agency. Interesting to those who follow AI developments specifically and/or to stakeholders in the particular sectors or groups affected.
• Major: Affects rights, duties, or resources for a large and diverse number of people and/or organizations, or has a highly disruptive effect on the rights, duties, or resources of a narrower group. Spurs significant activity in response and attracts widespread attention across a general audience.
• Foundational: Defines a broad and far-reaching regime of new or altered rights, duties, or resources. Significantly disrupts the status quo, spurring major changes across the public and private sectors. Widely acknowledged as a major development in the overall trajectory of AI governance at the national or global level.

Provisional annotation using a language model

After screening, a language model processes the full text of each in-scope document and generates draft annotations for human review. We use an "off-the-shelf" commercial LLM (accessed via API) and ETO-developed prompts.

At present, the language model is used only to generate draft summaries of each document as a whole and each segment in the document. We have found that machine-generated summaries of AGORA documents are generally reliable and useful. Regardless, we flag unreviewed machine-generated output in the dataset, and affected documents can easily be removed, if desired, by filtering on the flag fields.

In the future, we may also use language models to generate provisional thematic tags for segments.

Annotation and validation

Using the basic metadata and authoritative text compiled during the screening process, ETO annotators develop summaries and thematic tags for each in-scope document - or review and revise the machine-generated drafts, as available.

For tagging, the annotators use an extensive codebook that includes general guidance and specific definitions for each thematic tag. Tags are applied segment by segment. A custom-built Airtable interface structures the annotator workflow and facilitates quick and accurate annotation.

AGORA’s summaries are meant mainly to help users skim and sift, rather than as an analytic resource in themselves; the codebook provides brief instructions for short- and long-form summaries, but significant discretion is left to annotators.

After initial annotation, a second annotator (designated the “validator”) reviews each document in full and discusses any disagreements with the initial annotator. Note that these are not fixed roles; each AGORA annotator serves as initial annotator on some documents and as validator on others.

Once all issues identified in validation have been resolved, the document’s record, consisting of validated metadata, short and long summaries, and thematic tags, is marked complete.

The AGORA thematic taxonomy

AGORA includes a thematic taxonomy that is inspired by scholarly and policy literature, but intended to be useful to a wide range of potential users and reasonably intuitive to both those users and AGORA annotators. The taxonomy was drafted by an interdisciplinary team with training in law, data engineering, public policy, political science, AI governance, and quantitative and qualitative social science methods, with input from potential users in government, academia, and the private sector, and has been refined iteratively based on annotator and user feedback.

The taxonomy consists of discrete concepts (“tags”) organized into five domains:

• Risk factors governed: The characteristics addressed by the document that affect AI systems’ propensity to cause harm, closely related to ethical concerns. AGORA’s risk tags are adapted from the risk categories outlined in the widely-used AI Risk Management Framework issued by the U.S. National Institute of Standards and Technology.
• Harms addressed: The potential harmful consequences of the development or use of AI that this document means to prevent. Per AGORA’s definitions, “harms” are the consequences of “risks” (prior bullet). For example, an AI system that is insecure or biased (risk characteristics) might end up causing harm to physical health or financial loss (harms). AGORA’s harm tags are adapted from categories of harm developed by CSET researchers for use with the AI Incident Database.
• Governance strategies: The means provided in the document to address, assess, or otherwise act with respect to the development, deployment, or use of AI. These are essentially the proposed solutions to the policy problems or goals articulated. AGORA’s governance strategy tags were developed iteratively by the AGORA leadership team over the first several months of annotation.
• Incentives for compliance: The types of incentives provided for people, organizations, etc., to comply with the requirements of the document. This small group of tags covers positive and negative incentives commonly seen in AI-related statutes and regulations, such as subsidies and fines, respectively.
• Application domains addressed: Any economic or social sectors, such as healthcare or defense, specifically addressed in the document as contexts for AI development, deployment, or use. AGORA’s application tags are adapted from the TINA industry taxonomy used in prior analyses of AI-related investment and are comparable in granularity to 2-digit NAICS codes, with one exception: given significant attention to the use of AI by governments in particular, the code for government applications of AI is divided into subcategories, allowing for more specific analysis.

Annotators read each AGORA document in full, then decide whether each of the 77 codes in the AGORA taxonomy applies at any point in the document, based on the definition and (where available) examples and keywords provided in the codebook. In deciding, annotators are instructed to consider only the operative text of each document; to focus on what the document explicitly states or clearly and directly implies; and to ignore material unrelated to artificial intelligence.

Known limitations

• The dataset's focus on documents that directly address AI means that many AI-relevant laws, regulations, and norms are excluded by design. In particular, laws of general applicability - for example, securities regulation, civil rights law, or even common-law tort doctrines have implications for modern AI. However, this body of often older, AI-relevant general law is a potentially unbounded set, and the applicability of any particular document may be debated. Accordingly, we exclude these documents from AGORA’s scope. Though this makes the dataset less comprehensive of the overall AI governance landscape, it reduces uncertainty over the bounds of the resource. Also, AI-focused implementations of more general laws - such as agency regulations or guidance explaining how broad existing authorities will be applied to AI - are in scope for the dataset, reducing the impact of the exclusion. Nonetheless, this limitation means analyzing the dataset on its own cannot replace careful region-, context-, and sector-specific analysis of regulations relevant to particular groups, organizations, or AI use cases, though it can enhance such analysis.
• Many other documents within AGORA's scope are not yet included. AGORA’s nominal scope is broader than the set of documents collected to date. In particular, the current dataset skews toward U.S. law and policy. We are experimenting with automation and seeking more resources to increase processing volume and multi-language coverage. In the meantime, we have focused annotators’ efforts on especially high-profile or consequential documents (e.g., adopted rather than proposed policies) and on sets of documents of particular interest to current AGORA stakeholders.
• Some data is machine-generated and may contain errors. We currently use machine learning to draft summaries of AGORA documents. The machine output is included in the dataset once generated; it may take some time after that for a human annotator to review. We have found that the machine-generated summaries of AGORA documents are generally reliable and useful. Regardless, we flag unreviewed machine-generated output in the dataset, and affected documents can easily be removed, if desired, by filtering on the relevant fields.
• Annotators exercise judgment in applying AGORA’s scope and taxonomy, raising the risks of inconsistent screening and tagging. To mitigate these risks, we provide detailed conceptual definitions, along with examples and decision heuristics. Further, we implement a dual annotation process (initial annotation followed by validation and reconciliation) for each document, with disagreements elevated to AGORA leadership and resolved according to defined procedures. (Records pending validation are included in the dataset, but can easily be excluded, if desired, by filtering on the validation fields in the relevant tables.) Finally, we maintain a searchable central repository of prior questions and answers and make it accessible to all AGORA annotators online.
• Some documents are thematically tagged as a whole, rather than segment-by-segment; in these cases, it may not be immediately clear which parts of the documents justify which tags.
• AGORA fulltext files are unofficial copies of the underlying documents, collected by our screeners for research purposes and users' convenience. These files have been reformatted and may be out of date or otherwise diverge from official versions, particularly in the case of documents that were not finalized at the time text was collected (e.g., bills of Congress not yet signed into law). Use AGORA fulltext with caution. For official text, visit the original sources specified in the document's AGORA metadata.

Maintenance

How are the data updated?

Records are periodically added to the public AGORA dataset (and the web interface) using an automated script. We add records as soon as they are screened, then update them as they are annotated and validated. This means that some records in the dataset often change over time, and many include unvalidated and/or machine-generated content. If desired, this content can be disregarded by filtering on the relevant fields in the data tables.

Credits

• Schema, codebook and taxonomy: Zach Arnold, Daniel Schiff, Kaylyn Jackson Schiff
• Editors: Lindsay Jenkins, Ashley Lin, Konstantin Pilz
• Annotators: Zoe Borden, Eileen Chen, Ogadinma Enwereazu, Ari Filler, Diya Kalavala, Mallika Keralavarma, Kieran Lee, Sophia Lu, Anjali Nookala, Maya Snyder, Jayan Srinivasan, Alina Thai, Julio Wang
• Documentation: Zach Arnold, Daniel Schiff

Major change log

10/1/24

Initial release