The 2016 White House AI report asked agencies for fairness evidence and grantmakers to review grant terms

The October 2016 report on artificial intelligence asked agencies using AI for consequential decisions to verify efficacy and fairness, and asked grantmaking agencies to review grant terms. Later OMB policy set binding practices for agencies’ own AI and did not require them in grant criteria.

Looking Back revisits past weeks one at a time, up to the most recent, and asks what it meant for the banks, health systems and public agencies we serve. The series was written in 2026, with the benefit of hindsight, and each article says which week it looks back on.

Two-minute summaryThe dek and the first sentence of each section, quoted from the article

The October 2016 report on artificial intelligence asked agencies using AI for consequential decisions to verify efficacy and fairness, and asked grantmaking agencies to review grant terms. Later OMB policy set binding practices for agencies’ own AI and did not require them in grant criteria.

  1. The week The report came from the council’s Subcommittee on Machine Learning and Artificial Intelligence.
  2. The deploying agency: evidence before reliance The recommendation did not define consequential decisions.
  3. The funding agency: grant terms as a lever The review had a purpose.
  4. The evidence: what testing and disclosure the report described The report recorded a researcher’s argument.
  5. The judges: who assesses the AI It said that program provides for temporary assignment of personnel between the federal government and state and local governments, universities and other eligible organizations. [1] That program could move expertise in both directions, including to the state and local bodies Recommendation 17 concerned.
  6. How the advice reached different public bodies A state agency receiving federal funds could face grant terms if its federal funder adopted them. [1] It could also adopt its own standards.
  7. Objections, in order The report was advice. [1] But its two recommendations described real authorities.
  8. What a program director could check today For a federal agency, the first check is the AI use case inventory that later policy required.
  9. With hindsight Federal policy picked up Recommendation 16’s direction through an executive order and two OMB memoranda.

The week

In October 2016, the National Science and Technology Council’s Committee on Technology issued a report titled Preparing for the Future of Artificial Intelligence. [1] Its cover letter is dated October 12, 2016. [1]

The letter said the report surveyed the current state of AI, its existing and potential applications, and the questions it raised for society and public policy. [1] It also said the report made recommendations for specific further actions by federal agencies and other actors. [1]

The report came from the council’s Subcommittee on Machine Learning and Artificial Intelligence. The letter said that subcommittee was chartered in May 2016. [1] The Committee on Technology reviewed the report and concurred with its contents. [1]

The report followed a public process. On June 27, 2016, the Office of Science and Technology Policy published a request for information on artificial intelligence in the Federal Register. [2] Responses were due by July 22, 2016. [2] The request said OSTP wanted a view of AI across all sectors, to recommend directions for research and identify challenges and opportunities. [2]

The request also described workshops. It said OSTP was co-hosting four public workshops in 2016, on legal and governance issues, AI for public good, safety and control, and social and economic implications. [2] The report’s letter later said the outreach included five public workshops and that the request received 161 responses, published on September 6, 2016. [1]

The letter named two companion pieces. A companion document laid out a strategic plan for federally funded AI research and development. [1] In the coming months, it said, the Administration would release a follow-on report on AI-driven automation, jobs and the economy. [1]

The report contained 23 numbered recommendations, collected at its end. [1] They ranged across open data, agency capacity, regulation, transportation, research, workforce, fairness, safety, international engagement, cybersecurity and weapons. [1]

Two recommendations addressed consequential decisions about individuals. Recommendation 16 said agencies that use AI-based systems to make or support such decisions should take extra care to ensure efficacy and fairness, based on evidence-based verification and validation. [1]

Recommendation 17 addressed grantmaking. Agencies that make grants to state and local governments supporting AI-based systems for consequential decisions should review grant terms. [1] The aim was to ensure that products or services bought with federal grant funds produce results in a sufficiently transparent fashion and are supported by evidence of efficacy and fairness. [1]

Other recommendations addressed the government’s own capacity. Recommendation 5 asked agencies to draw on senior technical expertise when setting regulatory policy for AI-enabled products. [1] Recommendation 6 asked them to use personnel assignment and exchange models to broaden the workforce’s view of technology. [1]

The report placed these recommendations in a section on fairness, safety and governance. [1] That section said a commonly cited example at the workshops was apparently biased risk prediction tools used by some judges and prison officials, as detailed in a news investigation. [1] It said the article presented evidence suggesting that one commercial risk scoring tool generated racially biased scores. [1]

The same section discussed hiring. It said a model trained on biased past decisions could perpetuate that bias when used to screen job applicants. [1]

The October 2016 AI report spends much of its length surveying the field. Its advice on decisions about people sat in two recommendations near the end. [1] One asked agencies that use AI for consequential decisions about individuals to verify efficacy and fairness with evidence. The other asked grantmaking agencies to review grant terms for AI that state and local governments buy. [1]

Our argument is that the report named two separate levers, and later federal policy pulled only the first. OMB’s 2024 and 2025 memoranda set binding practices for agencies’ own AI. [4] [5] The 2024 memorandum said agencies were not required to put those practices into criteria for federal financial assistance. [4] We take the report’s governance advice in four parts:

  1. The deploying agency: what the report asked of an agency using AI on individuals.
  2. The funding agency: what it asked of an agency paying for AI that others use.
  3. The evidence: what kinds of testing and disclosure the report described as support for fairness claims.
  4. The judges: who the report said needed to be in the room to assess AI.

The parts do not overlap. The first two are the actors and their levers, the third is what counts as proof and the fourth is the capacity to weigh it. After them, we look at how the advice reached different public bodies, take the objections in order, and end with what a program director could check today.

This is a reading of public documents. Agencies and grantees should confirm current obligations with their counsel and grants officers.

The deploying agency: evidence before reliance

What Recommendation 16 asked

Recommendation 16 addressed federal agencies that use AI-based systems to make, or provide decision support for, consequential decisions about individuals. [1] It asked them to take extra care to ensure the efficacy and fairness of those systems. [1] The basis was to be evidence-based verification and validation. [1]

The recommendation did not define consequential decisions. It also did not say who would check the evidence. [1] It was advice from a council of the Executive Office of the President, not a directive with a deadline. [1]

Two tests, not one

The recommendation named efficacy and fairness together. [1] They are different tests. Efficacy asks whether a system does what it claims, such as predicting an outcome accurately. Fairness asks whether its errors and outcomes fall unevenly across groups.

A system can pass one and fail the other. The report’s own example made that point. It described a hypothetical model to spot false names that could misclassify distinctive names from a small minority group, even with representative data, because that group is less numerous. [1] The model could look accurate overall and still fail one group.

Decision support counted

The recommendation reached systems that provide decision support as well as systems that decide. [1] That matters because agency staff can treat a score as advice while relying on it. The report did not exempt a system because a person made the final call. [1]

Later OMB policy took the same view. M-25-21 says a high-impact determination is possible whether or not there is human oversight for the decision or action. [5] In our reading, that sentence carried forward the 2016 report’s inclusion of decision support.

Why the deploying agency is the natural owner

The deploying agency controls the use. It knows the population, the decision and the consequences of an error. It can design tests that match the real use, because it holds the real data. [1]

The report’s fairness discussion said that even developers may not understand a system well enough to prevent unintended outcomes. [1] That places the burden on the user of the system to test it in its own setting, rather than rely on a developer’s assurance.

What a deploying agency would need

An agency following Recommendation 16 would need a list of its AI systems used for consequential decisions. It would need, for each, a statement of what the system is supposed to do and a test of whether it does. It would need a test of whether errors fall unevenly.

None of those items was required in 2016. [1] They became requirements later, in the order and memoranda discussed in the hindsight section. [3] [4] [5]

What counts as consequential

The 2016 report did not list consequential decisions. [1] Later policy did. M-24-10 listed purposes presumed to be rights-impacting. [4] In law enforcement contexts, the list included risk assessments about individuals, predicting recidivism and determinations related to sentencing, parole, probation, bail and pretrial release. [4]

That list tracked the report’s own examples closely. The report had described risk tools used in sentencing, bail, assignment and parole. [1] In our reading, the 2024 list gave Recommendation 16 the definition it lacked. [1] [4]

M-25-21 used a different term and a general test. AI is high-impact, it says, when its output serves as a principal basis for decisions or actions with a legal, material, binding or significant effect on rights or safety. [5] Its Section 6 lists categories of use presumed to be high-impact. [5] An official who decides that a use in those categories is not high-impact must document that in writing to the chief AI officer. [5]

Errors in both directions

A consequential decision can go wrong two ways. A hypothetical benefits screening tool could flag an eligible applicant for review, delaying help. It could also clear an ineligible one. Each error has a different victim and a different cost.

Recommendation 16’s call for evidence of efficacy and fairness reaches both. [1] An evidence file would need to show the rate of each kind of error, and whether either fell unevenly across groups. A single accuracy figure would hide that split.

Monitoring after deployment

Evidence gathered before deployment can go stale. Data changes, populations change and the use of a tool drifts. The 2016 report did not set a monitoring cycle. [1]

Later policy did. EO 13960 required that agencies’ AI applications be regularly tested against its principles, with mechanisms to supersede, disengage or deactivate applications whose outcomes are inconsistent with intended use. [3] M-25-21 requires ongoing monitoring designed to detect changes to a system after deployment or changes to the context of use or associated data. [5]

In our reading, those requirements finished a thought the 2016 report began. Verification and validation are not one-time events for a system that keeps making decisions about people. [1] [3] [5]

Training the people who use the output

A score is read by a person. The 2016 report discussed training for AI practitioners and students. [1] It did not address the caseworker who reads a score. M-25-21 does. It requires sufficient and periodic training, assessment and oversight for operators of AI, so they can interpret and act on its output. [5]

Who accepts the risk

Evidence needs an owner who decides whether it is enough. The 2016 report did not name one. [1] M-25-21 does. It says agency policies must let agency heads delegate responsibility and accountability for risk acceptance to appropriate officials throughout the agency. [5]

The independent reviewer’s comments go to the person accepting the risk when that determination is made. [5] That creates a record: who accepted which risk, on what evidence, after what review. The report recorded that one workshop speaker compared AI to the delegation of authority to administrative agencies, which is bounded by oversight and limits. [1] In our reading, a dated risk-acceptance record supplies one such bound. [5]

The funding agency: grant terms as a lever

What Recommendation 17 asked

Recommendation 17 addressed federal agencies that make grants to state and local governments in support of AI-based systems for consequential decisions about individuals. [1] It asked them to review the terms of those grants. [1]

The review had a purpose. Products or services bought with federal grant funds should produce results in a sufficiently transparent fashion and be supported by evidence of efficacy and fairness. [1]

Why the report reached for grants

The report’s fairness section had described tools used by some judges in sentencing and bail hearings, and by some prison officials in assignment and parole decisions. [1] In our reading, most such decisions are made by state and local bodies, not federal agencies. A federal agency has no direct authority over a county court’s choice of tool.

What a federal agency can control is the money it gives. Recommendation 17 pointed to that control. [1] In our reading, the report recognized that the decisions it worried about were made below the federal level and that grant terms were the federal reach into them.

How a grant term would work

A grant term is a condition the recipient accepts with the money. A term under Recommendation 17 might require the recipient to obtain evidence of efficacy and fairness before buying a tool with grant funds. [1] It might require the vendor to allow testing. It might require the recipient to publish how the tool reaches results.

The report did not draft such terms. [1] It asked agencies to review their grants. The design of the terms was left to each grantmaking agency.

What the lever could and could not reach

Grant terms reach only what grant funds pay for. A county that bought a tool with its own money would be outside them. A tool bought with federal funds, then renewed with local funds, might fall out of the terms at renewal.

Grant terms also reach only the recipient. A state agency that passes funds to a county would need to carry the terms down. The report did not address subawards. [1] Those limits do not make the lever weak. They define where it works.

Procurement is not the same lever

Procurement and grants are easy to confuse. In procurement, the federal agency buys a tool for its own use. In a grant, the agency pays a recipient that buys and uses the tool itself.

EO 13960 reached procurement. Its scope included agencies’ procurement of AI applications and AI developed by third parties on behalf of agencies. [3] M-25-21 went further, with a section of recommendations on procurement to support the minimum practices for high-impact AI. [5]

Those provisions cover tools the federal government uses. A tool a county buys with federal grant funds is the county’s tool. Recommendation 17 was aimed at that gap. [1]

A hypothetical set of grant terms

A grantmaking agency reviewing its terms under Recommendation 17 might consider a short set of conditions. The report did not draft these; they are an illustration of what its words could support. [1]

  • Before buying an AI tool for consequential decisions with grant funds, the recipient obtains evidence of the tool’s efficacy and fairness for its population.
  • The vendor agrees to allow testing, including black-box testing with synthetic inputs where the code is not available.
  • The recipient publishes enough about the tool’s results for the public to evaluate its risk of bias.
  • The recipient keeps a human review and appeal route for people affected.
  • The recipient reports annually on use and any changes.

Each condition maps to something the report or later federal policy described. [1] [5] Each would also need staff to review what came in.

Data rights

M-25-21’s procurement section raised a point that grant terms could also address. It said agencies should ensure contracts retain sufficient rights to government data and to improvements to that data. [5] It said contract terms should protect such data from being used to train or improve a vendor’s commercial offerings without the agency’s express permission. [5]

A local recipient using a vendor’s tool on residents’ records faces the same question. Who may use the data the tool processes, and for what. A grant term could require the recipient to settle that in its contract. The 2016 report did not address data rights. [1]

Monitoring a recipient

Grant terms need monitoring. A federal program office would need to read what recipients submit, compare it with the terms and act on gaps. That draws on the same scarce expertise the report discussed for agencies. [1]

The report’s own account of uneven capacity suggests where monitoring would be thinnest. Agencies charged with reducing poverty and increasing mobility, it said, had more modest capabilities. [1] In our reading, those agencies fund local programs of the kind where consequential decisions about individuals are made.

The difference between the two levers

The first lever acts on the federal agency’s own operations. The second acts on someone else’s operations through a contract. [1] The first can be enforced through management. The second needs terms, monitoring and the willingness to withhold money.

We think that difference helps explain why later policy moved on the first and not the second. An OMB memorandum can direct agencies’ own practices. Changing grant criteria involves program statutes, recipients’ capacity and each agency’s grant rules. The 2024 memorandum’s footnote reflected that difference. [4]

The evidence: what testing and disclosure the report described

The report’s list of approaches

The report described several technical approaches to accountability for complex algorithmic decisions. [1] A system can be tested in the wild by presenting it with situations and observing its behavior. [1] It can be subjected to black-box testing with synthetic inputs, which allows testing of scenarios that might not occur naturally. [1]

Some or all technical details of a system’s design can be published so analysts can replicate it. [1] In some cases, the report said, it is possible to publish information that helps the public evaluate a system’s risk of bias while withholding other information as proprietary or private. [1]

Extensive testing as the main control

The report recorded a researcher’s argument. Because of the opacity of AI algorithms, extensive testing was the most effective way to minimize unintended outcomes, by listing bad outcomes and building tests to rule them out. [1]

It gave an example. A photo captioning model had applied an offensive label to some photos of dark-skinned faces, despite testing that showed high accuracy across all photos. [1] The report said extensive testing of how human faces were labeled, with review by people who could recognize unacceptable results, was one way to prevent it. [1]

Interpretable models

The report also pointed to interpretability. It said an interpretable model that helps people understand a decision allows them to interrogate the assumptions and processes behind it. [1] It called research to improve interpretability one example of making technology an aid to accountability. [1]

Data quality as evidence

The report tied fairness to data. It said that in the criminal justice system, some of the biggest concerns with big data were the lack of data and the lack of quality data. [1] If data is incomplete or biased, it said, AI can exacerbate problems of bias. [1]

That makes data part of the evidence. A fairness claim about a model trained on incomplete records is a claim about those records too. An agency would need to document what the data covered and what it left out.

Ethics and engineering

The report said ethical training for AI practitioners was necessary but not sufficient. [1] It said ethical training needs to be augmented with the technical capability to take precautions as a system is built and tested. [1]

It also described AI as a field that could be as much art as science, and said some stakeholders suggested a need to grow AI into a more mature engineering field. [1] Recommendation 19 asked AI and safety professionals to continue progress toward a mature field of AI safety engineering. [1]

Testing without the code

A public body buying a commercial tool may not have its source code or model. The 2016 report’s black-box testing described one answer: present the system with synthetic inputs and observe its behavior. [1]

M-25-21 later adopted the same idea. It says an agency without access to the underlying code, models or data must use alternative test methods. [5] Its examples are querying the AI service and observing outputs, or giving evaluation data to the vendor and obtaining results. [5] In our reading, that turned the report’s description into a required fallback for federal agencies. [1] [5]

Disclosure and trade secrets

The report recognized a tension between transparency and proprietary interests. It said it is sometimes possible to publish information that helps the public evaluate a system’s risk of bias while withholding other details as proprietary or private. [1]

That middle path matters for grant terms. A term demanding full source code could deter vendors. A term demanding published bias-relevant results could be met without exposing trade secrets. The report pointed to the second. [1]

Feedback from the people affected

Evidence can also come from the people a system decides about. M-25-21 requires agencies to give end users and the public an option to submit feedback on a high-impact use, where appropriate, and to use that feedback in decisions about the AI. [5]

The 2016 report recorded calls at its workshops for greater transparency when AI tools are used for public purposes. [1] It said transparency concerns extended to the potential for some form of explanation of any AI-based determination. [1] Public feedback is one way to learn whether explanations are understood.

What the evidence would look like in a file

An agency file supporting a fairness claim under Recommendation 16 might hold four things. A statement of the decision and population. Test results on real or representative cases. Results broken out by group. A description of the data and its gaps.

The report did not prescribe that file. [1] The later OMB memoranda came close to it. M-24-10 required agencies to assess whether rights-impacting AI produced significant disparities in performance across demographic groups. [4] M-25-21 required an AI impact assessment that summarizes the data used and its fitness for purpose. [5]

The judges: who assesses the AI

Expertise at the senior level

Recommendation 5 asked agencies to draw on technical expertise at the senior level when setting regulatory policy for AI-enabled products. [1] It said effective regulation needs collaboration among agency leadership, staff who know the regulatory framework, and technical experts with knowledge of AI. [1]

The recommendation asked agency leadership to recruit the technical talent needed, or identify it among existing staff. [1] It asked for sufficient technical seats at the table in regulatory policy discussions. [1]

Hiring and exchange

Recommendation 6 asked agencies to use the full range of personnel assignment and exchange models, such as hiring authorities, to build a workforce with more diverse perspectives on technology. [1] The report named the Intergovernmental Personnel Act Mobility Program as one example. [1]

It said that program provides for temporary assignment of personnel between the federal government and state and local governments, universities and other eligible organizations. [1] That program could move expertise in both directions, including to the state and local bodies Recommendation 17 concerned.

A community of practice

Recommendation 4 asked the subcommittee to develop a community of practice for AI practitioners across government. [1] It asked agencies to develop and share standards and best practices for using AI in government operations. [1]

Uneven capacity

The report was candid about capacity. It said the federal government’s capacity to foster and harness innovation varies widely across agencies. [1] It said some agencies, particularly those charged with reducing poverty and increasing mobility, have more modest relevant capabilities. [1]

It gave a comparison. It said the National Institutes of Health had an R&D budget of more than $30 billion, while the Department of Labor’s was $14 million. [1] Agencies running benefits and grants were, by the report’s own account, among those with less capacity to judge AI.

Regulation through existing regimes

The report’s regulation section set a starting point for the judges. It said that where a risk falls within an existing regulatory regime, policy discussion should start by asking whether existing rules already address it or need adapting. [1]

It recorded that the general consensus of commenters was that broad regulation of AI research or practice would be inadvisable at the time. [1] Commenters, it said, called for existing regulation to be adapted as necessary. [1] That approach needs judges who know both the regime and the technology, which is what Recommendation 5 asked for. [1]

An independent reviewer

M-25-21 added a specific judge. Its AI impact assessment must include the results of independent review by a reviewer within the agency who was not involved in development. [5] The reviewer’s comments must be included in the documentation and shared with the official accepting the risk. [5]

That requirement answers a question the 2016 report left open: who checks the evidence. [1] [5] The answer is a person inside the agency but outside the project.

The appeal as a check

The last judge is the person who hears an appeal. M-25-21 requires agencies to ensure that people affected by AI-enabled decisions have access to timely human review and a chance to appeal negative impacts, when appropriate. [5] It allows agencies to extend an existing appeals process to cover decisions made with AI. [5]

An appeals process is also a source of evidence. A rise in successful appeals from one group can show a fairness problem that pre-deployment tests missed. A director who reads appeal outcomes alongside test results has two views of the same system.

Why the judges matter to both levers

The deploying agency needs experts to design tests and read results. The funding agency needs experts to write grant terms and review what recipients submit. In our reading, the judges are the constraint on both levers.

A grant term that requires evidence of fairness is only as useful as the reviewer who reads that evidence. A program office without technical staff could end up relying on a vendor’s summary. The report’s workforce recommendations addressed that gap, though without a mechanism. [1]

How the advice reached different public bodies

Federal agencies using AI directly

A federal agency with AI in benefits, enforcement or screening was the direct audience of Recommendation 16. [1] In 2016 the recommendation imposed no deadline. [1] The obligation came later with EO 13960 and the OMB memoranda. [3] [4] [5]

Federal grantmaking agencies

In our reading, agencies funding criminal justice, workforce, health and social service programs were the audience of Recommendation 17. [1] The recommendation asked them to review their grants. [1] Later OMB policy encouraged, but did not require, them to consider the minimum practices in grant criteria. [4]

State agencies

A state agency receiving federal funds could face grant terms if its federal funder adopted them. [1] It could also adopt its own standards. The report’s advice did not bind states. [1] A state program director reading it in 2016 had a description of risks and a list of testing approaches.

Counties and cities

Local bodies, in our reading, make many of the decisions in the report’s criminal justice examples. [1] The report’s account of uneven capacity concerned federal agencies; it did not assess local capacity. [1] Federal grant terms were the route by which federal expectations would reach them. [1]

Courts

The report’s examples included judges using risk scores in sentencing and bail. [1] Courts are not executive agencies, and executive memoranda do not govern them. Grant terms on funds a court system receives were, in our reading, the closest federal executive route to those tools.

Vendors

Vendors selling AI to public bodies would feel both levers. A federal buyer applying Recommendation 16 would ask for testing access. A grant recipient bound by terms under Recommendation 17 would ask for evidence of efficacy and fairness. [1] The report’s note that some bias information could be published while other details stayed proprietary offered vendors a middle path. [1]

Objections, in order

1. The report was advice, so its levers were never real

The report was advice. [1] But its two recommendations described real authorities. Agencies control their own systems and set terms on their grants. The question was whether anyone would use those authorities, and later policy used the first. [3] [4] [5]

2. Grant terms would burden small recipients

That is a fair concern. A small county could lack the staff to obtain and review fairness evidence. The report itself described uneven capacity. [1] But the alternative was no evidence at all for consequential decisions bought with federal funds. A term could be scaled to the decision’s stakes.

3. Fairness has no single definition

The report did not define fairness. [1] Later OMB policy did not adopt one metric either, but required assessment of disparities in performance across demographic groups. [4] An agency could set its own measures, document them and test against them. The absence of a single definition did not prevent testing.

4. Testing cannot prove a system is fair

Testing finds the failures it looks for. The report said as much when it described building specialized tests for listed bad outcomes. [1] That is a limit of the evidence, not a reason to skip it. A file that records what was tested also records what was not.

5. Federal rules should not reach local decisions

There is a federalism objection. Recommendation 17 did not try to regulate local decisions. It asked federal agencies to review the terms of their own money. [1] A recipient that rejects the terms can decline the funds.

6. Human review makes the AI advisory

The report reached decision support as well as decisions. [1] M-25-21 later said a high-impact determination is possible with or without human oversight. [5] A person who reviews a score and never departs from it may add little. The objection assumes review that the record may not show.

7. AI governance slows useful services

The report itself took this concern seriously. It said that where regulatory responses threaten to raise compliance costs or slow beneficial innovation, policymakers should consider adjusting them to lower costs without harming safety or market fairness. [1]

That advice applies to grant terms too. A term scaled to the stakes of a decision costs less for low-stakes uses. The report’s concern was consequential decisions about individuals, not every use of AI. [1]

What a program director could check today

An inventory of consequential uses

For a federal agency, the first check is the AI use case inventory that later policy required. EO 13960 directed annual inventories of agencies’ non-classified and non-sensitive AI use cases. [3] M-25-21 says agencies must continue updating their annual inventory. [5] The check is whether every consequential use is on it.

The minimum practices file

For each high-impact use, a director could check for the file M-25-21 requires. That includes pre-deployment testing, an AI impact assessment, independent review and access to human review and appeal where appropriate. [5] Those items are the evidence Recommendation 16 asked for, now in required form. [1] [5]

Grant programs that fund AI

For a grantmaking office, the check is which programs fund AI used in consequential decisions. M-24-10 said agencies were encouraged to consider the minimum practices when choosing financial assistance criteria. [4] A director could record whether the office has considered it, and why it chose its terms.

Recipients’ evidence

For a state or local recipient, the check is whether it holds evidence of efficacy and fairness for each AI tool used in consequential decisions, whatever its funding source. [1] The report’s list of testing and disclosure approaches gives a starting set. [1]

Who reviews the evidence

Finally, a director could check who reads the evidence. The report asked for technical expertise at senior levels. [1] M-25-21 requires an independent reviewer within the agency who was not involved in development. [5] A recipient could name its equivalent.

An exit plan for each tool

A director could also check that each high-impact tool can be switched off. M-25-21 says agencies must have a plan to discontinue high-impact AI that is not performing at an appropriate level. [5] It says a non-compliant use case must be safely discontinued. [5] A recipient using a tool bought with grant funds could keep the same plan.

Appeals against tests

A last check joins two records. A director could compare the groups that appeal AI-supported decisions, and the groups that win, with the groups the pre-deployment tests examined. [5] A gap between them shows where testing should go next. That comparison needs no new authority, only both records in one place.

With hindsight

Federal policy picked up Recommendation 16’s direction through an executive order and two OMB memoranda. On December 3, 2020, the President signed Executive Order 13960, Promoting the Use of Trustworthy Artificial Intelligence in the Federal Government, published at 85 FR 78939 on December 8, 2020. [3]

The order set principles agencies must follow when designing, developing, acquiring and using AI. [3] They included that AI be accurate, reliable and effective, understandable, responsible and traceable, regularly monitored and transparent. [3] The order also directed each agency to prepare an annual inventory of its non-classified and non-sensitive AI use cases. [3]

The order’s scope covered AI used to advance agencies’ missions, enhance decision making or provide a specified benefit to the public. [3] It reached AI developed by third parties on behalf of agencies and agencies’ procurement of AI. [3] It did not mention grants to state and local governments as a channel for its principles. [3]

On March 28, 2024, OMB issued M-24-10. [4] It required agencies, by December 1, 2024, to implement minimum practices for safety-impacting and rights-impacting AI, or stop using non-compliant AI. [4] Among the purposes presumed rights-impacting were law enforcement risk assessments and determinations related to sentencing, parole, bail and pretrial release. [4]

For rights-impacting AI, M-24-10 required agencies to assess whether a model produced significant disparities in performance across demographic groups, and to mitigate disparities that lead to unlawful discrimination or harmful bias. [4] In our reading, that turned Recommendation 16’s evidence-based verification of fairness into a requirement. [1] [4]

M-24-10 addressed grants in a footnote. It said agencies were not required to incorporate the minimum practices into criteria for granting federal financial assistance. [4] It said they were encouraged, consistent with law, to consider the practices when choosing such criteria. [4] That footnote is the closest we found in later policy to Recommendation 17, and it stopped at encouragement. [1] [4]

On April 3, 2025, OMB issued M-25-21, which rescinded and replaced M-24-10. [5] It directed agencies to implement minimum risk management practices for high-impact AI. [5] It defined AI as high-impact when its output serves as a principal basis for decisions or actions with a legal, material, binding or significant effect on rights or safety. [5]

M-25-21 requires pre-deployment testing, an AI impact assessment, independent review, human oversight and access to timely human review and appeal where appropriate. [5] It gave agencies 365 days from its issuance to document implementation of the minimum practices. [5] It lists the AI in Government Act of 2020, the Advancing American AI Act and Executive Orders 13960 and 14179 among the laws and policies its requirements are consistent with. [5]

We found no reference to grants or federal financial assistance in M-25-21’s text. [5] As checked on November 16, 2026, M-25-21 was posted on the White House website. [5] We did not find a later OMB document in our sources that requires the practices in grant criteria.

What held up from October 2016 was the first lever. Agencies’ own consequential AI moved from advice to required testing, assessment and review. [1] [4] [5] What did not hold up, as of our check, was the second. The grant route the report named for state and local tools remained voluntary in federal policy. [1] [4]

Spotted an error? Tell us and we will check it and correct the article where needed.

  • artificial intelligence
  • algorithmic fairness
  • federal grants
  • OMB
  • state and local government
  • AI governance