The DoD Responsible AI Toolkit, explained

The Responsible Artificial Intelligence (RAI) Toolkit is a self-assessment resource the Chief Digital and Artificial Intelligence Office (CDAO) built so that teams inside the Department can check an AI project against the Department's AI ethical principles at any stage of its life. The CDAO released it publicly on November 14, 2023 as a key deliverable of the June 2022 Responsible AI Strategy and Implementation Pathway. The public site has since retitled it the AI Assurance Toolkit, and the pages now say Department of War where they once said Department of Defense. The structure underneath, a six-step assessment called SHIELD, a risk guide called DAGR, and a list of tools, is the same.

The five principles the toolkit operationalizes

The toolkit exists to apply five principles. The Department adopted them on February 24, 2020, following recommendations the Defense Innovation Board had delivered the previous October after 15 months of consultation. The release, now held in a historical collection on war.gov, says the principles apply to both combat and non-combat functions and build on an ethics framework grounded in the Constitution, Title 10 of the U.S. Code, the law of war, and existing treaties. The five are Responsible, Equitable, Traceable, Reliable, and Governable. Traceable asks that relevant personnel understand the technology and its development processes, with transparent and auditable methodologies, data sources, and documentation. Reliable asks that AI capabilities have explicit, well-defined uses, with safety, security, and effectiveness tested across their life cycles. The release named the DoD Joint Artificial Intelligence Center as the focal point for coordinating implementation.

The post on the DoD AI ethical principles gives the full text of all five, the May 26, 2021 implementing memo with its six tenets, the June 2022 pathway, and the Department's 2025 and 2026 changes. This post covers the toolkit those documents produced.

From the 2022 pathway to the 2023 toolkit

The CDAO's 2023 release ties the toolkit to its parent document. The Responsible AI Strategy and Implementation Pathway, the release says, was signed into adoption by Deputy Secretary of Defense Kathleen Hicks in June 2022 and outlines more than 60 lines of effort across the Department for operationalizing the AI Ethical Principles through the RAI implementation tenets, which include fostering a trusted RAI ecosystem of technical tools and guidance, educational resources, and thought leadership. The toolkit's own executive summary counts 64 lines of effort and calls the toolkit the pathway's "cornerstone."

The release describes what was delivered. The toolkit is built on numerous other RAI and AI ethics frameworks, assessments, and toolkits. It provides a voluntary process that identifies, tracks, and improves the alignment of AI projects to RAI best practices and the DoD AI Ethical Principles while capitalizing on opportunities for innovation. The user is guided through tailorable and modular assessments, tools, and artifacts across the AI product lifecycle. The CDAO said it hoped to bring transparency to the Department's approach by making the toolkit public, at rai.tradewindai.com, and that the version would be continually improved.

The executive summary names the sources the toolkit was built on: the earlier Responsible AI Guidelines and Worksheets from the Defense Innovation Unit, the National Institute of Standards and Technology (NIST) AI Risk Management Framework, and IEEE 7000. The summary also describes itself as a Minimum Viable Product, with organizational and operational guidance planned for later versions.

What the toolkit contains

The executive summary lists five design features. The toolkit is modular and tailorable, so users filter content by labels and see only what applies to their project. The content is aligned to a RASCI matrix that clarifies who is Responsible, Accountable, Supporting, Consulted, and Informed for each item, so that a program manager does not need full expertise or have to locate experts alone. The toolkit is holistic, meaning it covers evaluating and improving AI systems and not only engineering them. Every element can be traced back to the AI Ethical Principles. And a tools list offers more than 70 tools, mostly open-source, with the note that they are not endorsed and still need normal software approval before use.

The SHIELD assessment

SHIELD is the core. The executive summary lists it as six sequential activities, and a webinar hosted by the Cybersecurity and Information Systems Information Analysis Center (CSIAC) on November 5, 2024, presented by Andrew Brooks with the CDAO as source, describes each step:

The output of the process is the Statement of Concern (SOC), which the executive summary defines as a short factual statement of projected risk. SHIELD routes users to the tools list for mitigations. The webinar describes the toolkit as meant for personnel across an AI system's life cycle to check alignment and address whatever concerns the assessment surfaces.

DAGR

The second central piece is DAGR, the Department's AI Guide on Risk. The executive summary describes it as a resource for identifying and evaluating sources of AI risk that underpins SHIELD and the Statements of Concern. The summary says DAGR was to be built into a Department-specific AI risk management framework in fiscal year 2024 in collaboration with NIST. The CSIAC webinar calls it a Department-specific risk assessment resource included in the toolkit.

The tools list and the planned additions

The toolkit's home page describes the tools list as a collection of tools and resources to help design, develop, and deploy assured AI applications; the executive summary counts more than 70 tools, mostly industry-standard and open-source, and says a SHIELD assessment routes users to them. The executive summary lists what is planned but not yet in the toolkit: data, model, and system card templates; AI assurance project management tools; harms and impact analysis templates; an acquisition toolkit with standardized contract language; and a human-machine teaming red-teaming guidebook. What those card templates would document is the subject of the posts on system cards and model cards and datasheets.

Who it is for and how to use it

The Tradewinds toolkit page describes the toolkit as a self-evaluation resource that helps teams assess their AI projects at any stage of development, and lists the people it expects to use it: program managers, AI developers, acquisition teams, legal and compliance staff, and end users, who assess a project collaboratively to confirm it meets performance goals and risk tolerance. The page says the toolkit is accessible in both Impact Level 2 and Impact Level 5 environments and that the CDAO will update it continuously based on advances and user feedback. What the impact levels mean is covered in the post on DoD cloud impact levels.

The toolkit's home page explains the mechanics. The public version is a web application at rai.acqbot.com, under the CDAO's AI Assurance Division. Anyone can read the content; a user who creates an account can save projects, collaborate, track progress, and join or create an organization workspace for a team. The page links a separate Impact Level 5 version on acqbot.mil for projects that require a secure environment. Help content covers filters, the responsibility matrix, PDF export, and team collaboration.

Two points follow for a contractor. First, the toolkit is public, so a vendor can run its own project through SHIELD before a government team does, and arrive with Statements of Concern already written and mitigations already mapped to the tools list. Second, the assessment is designed to be shared across roles, which means the government program manager, the contracting officer's representative, and the vendor's engineers can be looking at the same project record.

How the toolkit relates to the NIST AI RMF

The toolkit and the NIST framework are different kinds of document that cite each other. The NIST AI Risk Management Framework is a voluntary framework of outcomes organized into four functions, Govern, Map, Measure, and Manage, with a Generative AI Profile, NIST AI 600-1, that names 12 risks of generative AI. The toolkit is a procedure: six steps that produce Statements of Concern for one project and route the team to mitigations. The executive summary lists the NIST framework among the toolkit's sources, and says DAGR, the risk guide that feeds the Statements of Concern, was to be developed into a Department-specific AI risk management framework in fiscal year 2024 in collaboration with NIST.

For a vendor that already documents its system against the NIST framework, the mapping is mostly a matter of order. The Set Foundations and Hone Operationalizations steps ask for the legal, policy, and ethical basis and the assessment criteria, which is the Govern and Map work of the framework. Evaluate Status is the Measure work. Improve and Innovate and Detect via Continuous Monitoring are the Manage work. A project record that already has a NIST profile should be able to fill the SHIELD steps from it; a project that has run SHIELD has the raw material for a profile. Neither document replaces the other, and the toolkit's own pages do not claim that completing SHIELD satisfies any NIST or regulatory requirement.

How the toolkit ties to acquisition

The toolkit was never a contract clause. The CDAO's release calls it voluntary, and the toolkit's home page describes a self-assessment tool. The acquisition connection is one of intent and gaps.

The intent is on the toolkit's own pages. Acquisition teams are among the named users, and the acquisition toolkit with standardized contract language is on the planned list. The toolkit's inventory of planned artifacts, data and model and system card templates and a test-and-evaluation posture through the Detect step, reads as a description of deliverables a program office could ask for.

The gap is documented by the Government Accountability Office (GAO). In GAO-23-105850, published June 29, 2023, GAO found that the Department had designated AI a top modernization area and was spending considerably on it, but had not issued department-wide guidance on how its components should approach acquiring AI, and that the military services also lacked AI acquisition-specific guidance. GAO noted that the CDAO oversees the Tradewind AI marketplace and that several of its resources stress intellectual property and data rights in contract negotiations. GAO made four recommendations, one to the Department to prioritize department-wide AI acquisition guidance and one each to the Army, Navy, and Air Force to follow with service-specific guidance. The Department concurred. As GAO's page records it, the Department issued a Data, Analytics, and AI Adoption Strategy in November 2023 and a federated AI construct implementation plan in December 2024, and as of May 2025 a decision aid framework appendix was under internal review, with the recommendation marked open and partially addressed. GAO's page lists the three service recommendations as open, with the latest service updates dated November 2024 to March 2025.

No public source found for this post states standard RAI requirements that appear in Department AI contracts. The toolkit's executive summary says contract language is planned; the post on the DoD AI ethical principles, linked above, describes the January 9, 2026 Department AI strategy memo and its direction on "any lawful use" language for AI service contracts. A vendor should read the solicitation in front of it rather than assume either document.

The Department's naming on the toolkit pages

The toolkit pages reflect the Department's current naming. The Tradewinds page says the toolkit operationalizes the "DoW AI Ethical Principles," and its disclaimer says the CDAO will use feedback to accelerate responsible adoption of AI within the Department of War. The executive summary at rai.acqbot.com uses Department of War and DoW throughout its visible text; the only places DoD appears are in the addresses of linked documents. The CDAO's own homepage describes its mission as accelerating DoW adoption of data, analytics, and artificial intelligence. The 2020 release that adopted the principles, kept on war.gov in a historical collection, still reads Department of Defense, as does the archived 2023 toolkit release. The post on the DoD AI ethical principles covers the September 5, 2025 executive order behind the secondary name and notes that statutory references to the Department of Defense remain in force.

For a document writer, the practical rule is to match the page being cited: DoD for the 2020 principles and the 2023 release, DoW or Department of War for the current toolkit pages, and the toolkit's current title, the AI Assurance Toolkit, when pointing a reader to it. The CDAO's 2023 release and the CSIAC webinar remain the public record of the Responsible AI Toolkit name.

What to check

• Which version of the toolkit does a program office or solicitation reference, the 2023 Responsible AI Toolkit or the current AI Assurance Toolkit, and does the assessment record carry a date?
• Has the project been run through all six SHIELD steps, and are the Statements of Concern written as short factual statements of projected risk with named mitigations from the tools list?
• Which of the five principles does each Statement of Concern trace to, since the toolkit is built so that every element traces back to them?
• Who holds each RASCI role for the project, and does the vendor's team appear in the matrix for the items it is responsible for?
• Does the plan for the Detect step name what will be monitored after fielding, how often, and who acts on a performance drop?
• Does the solicitation itself state RAI deliverables or contract language, given that the toolkit is voluntary and its acquisition toolkit is still listed as planned?

Azcuna's ELLMENT, the E-2D Large Language Model Entity, was proposed to NAVAIR PMA-231 under Navy SBIR Phase I topic DON26BZ01-NV010. The prototype, built at private expense on publicly available doctrine, answers with a page-level citation and refuses when the documents do not support an answer. The proposal has been submitted.

Questions people ask

What is the DoD Responsible AI Toolkit?

The Responsible Artificial Intelligence (RAI) Toolkit is a self-assessment resource from the Chief Digital and Artificial Intelligence Office (CDAO). The CDAO's November 14, 2023 release called it a voluntary process that identifies, tracks, and improves alignment of AI projects to RAI best practices and the DoD AI Ethical Principles, through tailorable and modular assessments, tools, and artifacts across the AI product lifecycle. The public site now titles it the AI Assurance Toolkit.

What is the RAI Toolkit from CDAO built around?

The core is the SHIELD assessment, six sequential activities: Set Foundations, Hone Operationalizations, Improve and Innovate, Evaluate Status, Log for Traceability, and Detect via Continuous Monitoring. The toolkit's executive summary says SHIELD routes users to a tools list of more than 70 mostly open-source tools and produces Statements of Concern, short factual statements of projected risk. The Department's AI Guide on Risk, DAGR, underpins SHIELD and the Statements of Concern.

What are the five DoD AI ethical principles?

The Department adopted five principles on February 24, 2020, following Defense Innovation Board recommendations: Responsible, Equitable, Traceable, Reliable, and Governable. The release says they apply to combat and non-combat functions and build on the existing ethics framework of the Constitution, Title 10, the law of war, and treaties. The toolkit's executive summary says each element of the toolkit can be traced back to these principles, which the site now calls the DoW AI Ethical Principles.

What is the RAI Strategy and Implementation Pathway?

The CDAO's 2023 release describes the Responsible AI Strategy and Implementation Pathway as the document Deputy Secretary of Defense Kathleen Hicks signed into adoption in June 2022, outlining more than 60 lines of effort for operationalizing the AI Ethical Principles through the RAI implementation tenets. The toolkit's current executive summary counts 64 lines of effort and calls the toolkit the pathway's cornerstone. The toolkit was a deliverable of that pathway.

Who is the DoD RAI Toolkit for?

The Tradewinds page lists program managers, AI developers, acquisition teams, legal and compliance staff, and end users, who work through the assessment together. The executive summary says the toolkit is aligned to a RASCI matrix of Responsible, Accountable, Supporting, Consulted, and Informed roles so that a program manager does not need full expertise or have to find experts alone. The CDAO made the toolkit public to bring transparency to the Department's approach, and anyone can use the public version.

Is the Responsible AI Toolkit mandatory for DoD AI contracts?

The CDAO's release calls the toolkit a voluntary process, and the toolkit's home page describes a self-assessment tool. The current page lists an acquisition toolkit with standardized contract language among planned additions, not existing ones. GAO reported on June 29, 2023 that the Department had not issued department-wide guidance on acquiring AI, and as of May 2025 GAO's recommendation to the CDAO remained open and partially addressed. A specific solicitation can still require RAI artifacts.

Where is the DoD RAI Toolkit hosted?

The 2023 release pointed to rai.tradewindai.com. That address now redirects to the Tradewinds AI Assurance Toolkit page, which links to the toolkit itself at rai.acqbot.com and says the toolkit is accessible in Impact Level 2 and Impact Level 5 environments. The rai.acqbot.com home page links a separate Impact Level 5 version for projects that need a secure environment. Users can create an account to save projects and work in an organization workspace. Tradewinds is the CDAO's marketplace site, and the toolkit page says CDAO will update the toolkit continuously.

What is DAGR in the RAI Toolkit?

DAGR is the Department's AI Guide on Risk, which the toolkit's executive summary describes as a resource for identifying and evaluating sources of AI risk that underpins the SHIELD assessment and the Statements of Concern it produces. The summary says DAGR was to be built into a Department-specific AI risk management framework in fiscal year 2024 in collaboration with NIST. A CSIAC webinar of November 5, 2024 on DAGR and SHIELD called DAGR a DoD-specific risk assessment resource included in the toolkit.

Why do the toolkit pages say Department of War?

The public toolkit pages on tradewindai.com and rai.acqbot.com, and the CDAO's own ai.mil homepage, now use Department of War and the abbreviation DoW, and refer to the DoW AI Ethical Principles. The February 2020 release that adopted the principles is kept on war.gov in a historical collection and still reads Department of Defense. The substance of the principles and the toolkit's steps has not changed with the name on the page.

Sources

1. 1Responsible AI Toolkit (CDAO Public Affairs, November 14, 2023) · Chief Digital and Artificial Intelligence Office, archived by the Internet Archive
2. 2AI Assurance Toolkit: Executive Summary · Chief Digital and Artificial Intelligence Office
3. 3AI Assurance Toolkit · Chief Digital and Artificial Intelligence Office
4. 4AIA Toolkit · Tradewinds (Chief Digital and Artificial Intelligence Office)
5. 5Responsible Artificial Intelligence (RAI) Toolkit: DAGR and SHIELD · Cybersecurity and Information Systems Information Analysis Center (DTIC)
6. 6DOD Adopts Ethical Principles for Artificial Intelligence · U.S. Department of War (release of February 24, 2020)
7. 7Chief Digital and Artificial Intelligence Office · Chief Digital and Artificial Intelligence Office
8. 8Artificial Intelligence: DOD Needs Department-Wide Guidance to Inform Acquisitions (GAO-23-105850) · U.S. Government Accountability Office

Where this shows up in our work

• CALIBER, the market testbed for AI agents measures from an AI agent's trades whether it is gullible, herds, overpays, or can be steered.
• ELLMENT, the offline LLM for E-2D crews answers a crew's question from the aircraft's own manuals and cites the page, with no network.