This is the initial public draft of NIST AI 300-1, Guidance and Templates for Public-Facing AI Documentation: An AI Standards "Zero Draft", published by NIST in July 2026 in the NIST Trustworthy and Responsible AI series and authored by Razvan Amironesei and Jesse Dunietz. It runs to 54 pages and is available free of charge at https://doi.org/10.6028/NIST.AI.300-1.ipd.
The document is the first published output of the AI Standards Zero Drafts pilot, a mechanism by which NIST drafts a preliminary standard and hands it to the private-sector standardisation process rather than maintaining it as a NIST framework. The topic page is NIST AI 300-1 — Guidance and Templates for Public-Facing AI Documentation.
The Zero Drafts mechanism
NIST announced the AI Standards Zero Drafts project in March 2025. As the foreword describes it, NIST collects input on "topics with a science-backed body of work" and uses it to develop "zero drafts—preliminary, stakeholder-driven drafts of standards that are as thorough as possible," which "will then be submitted via the private sector–led standardization process as proposals for voluntary consensus standards." Two topics were selected from community input; one is documentation of AI datasets and models for transparency between AI stakeholders, and this document is its initial public draft.
The intended path is explicit: the resulting document "will be submitted to INCITS/AI, the private sector–led committee that represents the United States in ISO/IEC JTC 1/SC 42," the subcommittee for cross-sectoral AI standards within ISO and IEC. If INCITS/AI proposes it as a new SC 42 project and SC 42 accepts, "the future of the document will be subject to the usual consensus processes of standards development." NIST states that it "does not expect to maintain the document further, and while the agency expects to contribute heavily through the ISO standardization process, NIST will be just one voice among many in SC 42 and INCITS/AI."
To ease that transition, the draft is written in ISO/IEC drafting conventions: "shall," "required" and "requirement" denote provisions mandatory for conformance; "should," "guidance" and "may" denote recommendations; "can" denotes a possibility and is not normative. The foreword adds a disclaimer that these conventions carry no regulatory force: "any usage of 'shall' or stated 'requirement' does not reflect any regulatory intent, nor does it indicate a rule or directed action from the U.S. federal government. Rather, use of 'shall' or 'requirement' indicates only what constitutes conformity."
NIST states its motivation as stakeholder demand, "most notably from industry": that standardising AI documentation practices "would lower the time, effort, and expertise required for both creating and consuming documentation, which would facilitate building on prior research and development efforts and ultimately accelerate innovation and adoption."
Scope
The stated scope is narrower than AI documentation generally. The document "provides guidance on objectives for and approaches to documenting AI datasets and models with the intent of releasing the resulting documentation artifacts publicly," supplying recommendations for documentation processes, requirements for artifacts "in the form of templates against which conformity of documentation artifacts can be assessed," and a mechanism for extending the templates via profiles.
It applies to any organisation regardless of size, type, or nature that provides or uses products or services utilising AI systems, and to any type of AI model, "including generative AI and large language models but also predictive machine learning models and others." It applies to public-facing documentation of datasets and models "whether or not the datasets and models themselves are publicly available."
Two exclusions are stated. Whole AI systems are out of scope: the document "does not address documenting entire AI systems, which may consist of large, complex arrangements of components that rely on or interact with a given dataset or model," because "NIST assesses that documentation practices at this level are less mature, so they have been left to future work." And an AI model "is considered to include only the model architecture and parameters"; components typically shipped or served with a model — the draft's example is "guardrail classifiers that operate on model output" — fall outside it. A reviewer note states that NIST "has not identified a principled basis for scoping in certain system components, such as those sometimes referred to as part of a model, while scoping out others," and invites input on how model documentation could be expanded without extending to entire systems. That boundary determines whether a documentation standard can reach the safety scaffolding around a model, which is where much of the system-card and model-documentation literature places its content.
Structure and content
The draft has six clauses and an annex. Clause 3 sets out the dimensions along which documentation processes and artifacts produce outcomes, grouped as process-driven (organisational resource demands, internal collaboration, internal decision-making, release speed), artifact-driven (reuse and integration, malicious use or attacks), and jointly driven (accountability, privacy preservation, proprietary information protection, legal obligations, trust).
Clause 4 gives eight qualities of documentation artifacts — comprehensibility, informativeness, correctness, judiciousness, freshness, interoperability, findability and availability, maintainability — and seven items of process guidance: define documentation objectives, provide organisational support, keep processes manageable, provide guidelines, document continuously, distribute documentation work, and incorporate audience input.
Clause 4.4 names four trade-offs the draft treats as unavoidable. The first, informativeness against judiciousness, is the disclosure question in its general form: "There is often information that would be useful to some interested parties, but which other interested parties, possibly including the originating organization, would prefer or even be obliged to protect," and "some information can be useful to both legitimate interested parties and malicious attackers, so its release can require careful judgment." The others weigh informativeness against comprehensibility and maintainability, informativeness against freshness, and continuous documenting against process manageability.
Clause 5 supplies the two templates. Conformity turns on three rules: a field included in an artifact "shall be populated with information matching the field description and guidance"; information matching a field's description "shall" appear under that field; and all fields designated required "shall be present and populated." Fields marked recommended "should be populated under most circumstances, but are not required to be." Both templates consist of root fields only, with most designated optional; the granular subfields and more prescriptive designations are deferred to the Annex A default profiles.
The dataset template has seven root fields: Identifying Descriptors (required), Intended Use, Usage Rights and Restrictions, Composition and Provenance (recommended), Evaluation, Maintenance and Monitoring, and Dataset Governance (recommended).
The model template has eight: Identifying Descriptors (required), Intended Use, Usage Rights and Restrictions, Design, Training, Evaluation, Maintenance and Monitoring, and Governance (recommended). Only one field in each template is required. Training-data information "should be structured according to the dataset template in Clause 5.2, or should reference documentation artifacts that follow that template," linking the two.
Clause 6 defines profiles: documents "that can be created by any interested party to adapt the templates to particular contexts of use (e.g., an AI application domain, a technical category of dataset or model, a risk level, or a geographic area with specific policy requirements)." A profile "shall not change or contradict the structure, requirements, and guidance of the templates in Clause 5, except to elevate a root field's designation" — optional to recommended or required, or recommended to required — and may add subfields, add root fields, and set conditional designations. The worked example is that "a profile for medical datasets would likely make a privacy policy field required, while a profile for astronomical imagery datasets would likely make the field optional."
Draft status and open questions
The Introduction is a placeholder: "(To be completed with the final draft)." Three reviewer notes solicit input on specific unresolved choices: whether "documentation artifact" is the right term for a record describing a product or service; how far model documentation should extend beyond architecture and parameters; and whether the templates should carry inline example content or filled-in examples for representative datasets and models, NIST recording that "stakeholder input has indicated a desire for concrete examples."
The foreword states that "NIST will publish a final and complete version of this document based on feedback received." Contemporaneous trade reporting characterised the draft as "an initial and potentially final version"; no such phrase appears in the document, whose only statement in that direction is that NIST does not expect to maintain the document after it enters the ISO process.
Comments are invited on any aspect of the document by email to ai-standards+doczd@nist.gov, with an optional commenting template that NIST "strongly encourages" but does not require; marked-up documents, bulleted lists, and reaction letters are accepted. Submissions "will become part of the public record and may be subject to public disclosure." NIST states it "will consider input received by September 16, 2026," with later input still potentially informing subsequent standardisation contributions. Organisations are invited to host listening sessions. A closing instruction asks that use of AI assistants in preparing feedback be disclosed, so that "the output prioritizes actionable advice and references that go beyond what could easily be elicited from AI assistants."
Provenance
Published in July 2026; the PDF cover carries the month only. NIST's AI Standards page states the initial public draft was released on July 29, 2026, and the Zero Drafts pilot page gives "July 2026" without a day; an Inside AI Policy item of August 3 gives July 30 from a lede whose body was not retrievable. The July 29 date on NIST's own page is the most authoritative day-level date available and is used here. Retrieved and text-extracted from https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.300-1.ipd.pdf on August 5, 2026; the DOI prefix 10.6028 is NIST's registered prefix and both NIST programme pages link the same DOI under the same title.
Relationships
- depends-on: NIST AI Risk Management Framework 1.0 — the preceding NIST AI framework, and the reference point for NIST's documentation and transparency work.
- related: NIST AI 300-1 — Guidance and Templates for Public-Facing AI Documentation — the topic page for the document and its process.
- related: ISO/IEC 42001 — AI Management System — the ISO/IEC AI management-system standard developed in the same JTC 1/SC 42 committee the zero draft targets.
- related: AI System Cards, System Card Due Diligence, AI Transparency — the documentation practices the templates formalise.
- related: National Institute of Standards and Technology (NIST), NIST CAISI (Center for AI Standards and Innovation).