What the White House Published in 2022

The Office of Science and Technology Policy (OSTP), a White House office, published the document Blueprint for an AI Bill of Rights: Making Automated Systems Work for the American People in October 2022. NITRD recorded the publication announcement on October 4 of that year. While the title can be translated as "Blueprint for an AI Bill of Rights," the original name more precisely identifies the historical document.

The Blueprint gathers five principles and associated practices to guide the design, use, and deployment of automated systems. Its stated purpose was to help protect civil rights, civil liberties, and privacy, and to promote democratic values in the US context. The initiative followed a public participation process led by the OSTP; it was not a law passed by Congress nor a list of legal requirements automatically applicable to all organizations.

The document was accompanied by From Principles to Practice, a technical manual with steps and examples for incorporating these ideas into policies, processes, and technological design. This distinction matters: the Blueprint formulates general principles, while the manual seeks to facilitate their practical application. Both texts serve to understand a 2022 governance proposal, not to certify that a system meets a legal obligation.

Which Automated Systems It Referenced

The proposed scope went beyond chatbots and generative AI. The document spoke of automated systems capable of significantly affecting the rights, opportunities, or access of the public to essential needs and services. It included technologies with artificial intelligence, but the concept was not limited to a specific commercial application or current generative models.

Examples of relevant areas include employment, credit, health, education, justice, and benefits. For instance, a system might rank job applications, provide a score for a credit evaluation, or help determine which cases receive priority attention in a service. These are illustrative hypothetical situations: the point is that a tool can influence the outcome even if it does not make the final decision on its own.

The guide was formulated for the US context and considered both public and private organizations. The weight of each principle depends on the specific use, the people affected, and the possible consequences. A tool that recommends songs and another that influences access to an essential resource do not pose the same level of risk nor require identical safeguards.

Safe and Effective Systems

The first principle, Safe and Effective Systems, recommends that systems be designed with input from affected communities, stakeholders, and specialists in the relevant field. Participation can help discover risks that are not evident to those developing the tool: for example, that an interface is inaccessible or that the process does not reflect the real conditions in which it will be used.

The proposal also calls for testing before deployment, identification and mitigation of risks, and subsequent monitoring. Evaluating a tool only in laboratory conditions does not necessarily reveal how it will behave when data, environment, or operators change. Monitoring allows for the detection of degradation or unintended effects and determines whether the system needs adjustment.

A possible measure, according to the principle, is not to deploy or to withdraw the system when risks cannot be adequately mitigated. This does not equate to promising absolute security: the guide proposes controls to reduce risks and review performance, not a guarantee that there will never be errors. It also invites examining foreseeable uses different from the stated purpose and avoiding irrelevant or harmfully reused data.

Protection Against Algorithmic Discrimination

The second principle, Algorithmic Discrimination Protections, addresses unjustified differences in treatment or impact associated with characteristics protected by law or the circumstances of affected groups. It recommends examining both input data and outcomes: a seemingly neutral variable can act as a proxy for another characteristic, and an acceptable average performance can hide errors concentrated in certain groups.

Suggested practices include equity evaluations from the design phase, data appropriate to the purpose, accessibility for people with disabilities, testing before and after deployment, and corrective measures. In a hypothetical example, a hiring tool might produce different results for comparable groups; reviewing its results and the process's functioning would help identify a disparity requiring analysis.

Removing a variable is not enough to demonstrate that bias has been eliminated: other variables may be related to it, and data quality, design, and context of use also influence outcomes. The Blueprint proposes continuous evaluation and mitigation, not a single formula or an automatic conclusion about the legality of each result. It also does not create crimes, sanctions, or new enforceable obligations on its own; the application of anti-discrimination laws depends on their own terms and circumstances.

Illustration of people and data profiles connected to an automated system, with symbols of fairness and privacy.
Editorial illustration by BytePreview

Privacy and Control Over Data

The Data Privacy principle recommends protections integrated from the design phase and enabled by default. It proposes limiting collection to data necessary for the purpose and context, explaining how they are used, and ensuring that interface decisions do not hide options or push toward invasive practices. The practical idea is to review what information is needed before collecting it, rather than accumulating it just in case.

The document also proposes asking for permission and respecting decisions regarding collection, use, access, transfer, and deletion when appropriate and to the extent possible. It does not present consent as a universal solution: its use must make sense and be understandable to the person. If it is not possible to adequately rely on consent, it contemplates other privacy-by-design safeguards.

For sensitive areas such as health, employment, education, justice, and finance, the guide suggests stricter protections and limiting data and inferences to necessary functions. This does not equate to declaring a general right to suppression applicable to all situations, nor to establishing a universal consent requirement. The framework emphasizes that the response depends on the context, the data processed, and the applicable norms.

Notice and Explanation of Decisions

The fourth principle, Notice and Explanation, combines two distinct needs. Notice of use should allow individuals to know that an automated system is intervening, who is responsible for it, and what role the automation plays. An explanation of the outcome, in contrast, helps understand how and why the system contributed to the result affecting a person.

The guide recommends clear, timely, accessible, and comprehensible information, as well as notifying of important changes in the use or functions of the system. If a tool contributes to ranking applications and a person makes the final decision, the explanation should clarify that participation, rather than implying that the machine decided on its own.

According to the proposal, the explanation must be valid and useful for its purpose and for the recipient. A person needing to correct data may require different information than an operator needs to review the system. The principle also does not reduce transparency to publishing technicalities: it seeks for the information to serve in understanding, detecting errors, or challenging a result when appropriate.

Human Alternatives and Resolution Paths

The fifth principle is called Human Alternatives, Consideration, and Fallback. It proposes a human alternative where appropriate, considering reasonable expectations in the context, accessibility, and possible impact. It does not propose an unconditional right to speak with a person in every automated interaction; it recognizes that the relevance of an alternative depends on the case and that other norms might require it in specific circumstances.

It also recommends accessible paths for a person to review an error, request consideration, and challenge or appeal the effects of a system. A nominal review would not be sufficient if the reviewer lacks time, information, training, or capacity to correct the result. The principle therefore includes operator training and effective escalation and backup mechanisms.

In a hypothetical situation, an application might be blocked by an incorrect data point. A useful route should allow identifying the problem and requesting a human review within a timeframe appropriate to the consequences. The guide also distinguishes between a review to correct failures and a complete alternative to the use of the system: they are related tools, but not identical.

Its Value as a Guide and What It Did Not Create Legally

The legal limit is expressed in the document's own legal notice: the Blueprint is a white paper non-binding and not constituting US government policy. It also does not replace, modify, or direct the interpretation of existing laws, regulations, or other instruments. The notice states that it is not binding guidance for the public or federal agencies and that its principles do not require compliance on their own.

The text further declares that it does not create legal rights, benefits, or defenses enforceable in courts. Therefore, it is not a law nor a compliance certification. Adopting its recommendations also does not guarantee compliance with other existing obligations. Conversely, the Blueprint's non-binding nature does not eliminate laws that are applicable to an organization, activity, or jurisdiction.

Its value lies in offering a common framework of questions for designers, deployment managers, authorities, and affected communities. It can help structure a risk assessment or a conversation about transparency and redress, but it does not replace specific legal analysis. The guide itself warns that context matters and that some sectors may need specific orientations.

Illustration of a person facing a robot and data profiles, with scales and a shield with a padlock.
Editorial illustration by BytePreview

How to Read It Today: 2025 and 2026 Federal Context

The Blueprint should be read as a document published in 2022, not as an automatic statement of current federal policy in November 2026. Its non-binding status was part of the original text. Furthermore, since then, distinct executive instruments and federal plans have been issued, which must be identified separately rather than treated as versions of the Blueprint.

On January 23, 2025, an executive order pushed for the development of an action plan and ordered the review of actions linked to Executive Order 14110 of 2023. The plan presented on July 23, 2025, grouped its actions into three pillars: innovation, AI infrastructure, and international diplomacy and security. These instruments have their own dates, purposes, and scopes; they do not retroactively convert the 2022 document into a legal norm.

On March 20, 2026, the White House presented a national legislative framework with recommendations directed to Congress. Presenting proposals does not certify that they have been approved as law. In September 2026, another executive order indicated that the executive branch should use "Super Intelligence" and "SI" instead of "Artificial Intelligence" and "AI" in certain documents and communications, without requiring changes to historical documents already issued. This terminological change does not equate to creating the rights of the Blueprint nor to replacing all applicable laws on its own.

The Five Principles as Practical Questions

Its questions are not an official compliance test and do not replace legal or sectoral requirements; they turn general principles into concrete points for discussion about a system and its effects.

The Blueprint’s five principles

A concise overview of the five principles and the main safeguard each recommends.

The Blueprint’s five principles
Safety and effectivenessTest, monitor, and address risks
Protection from discriminationAssess disparities and reduce harm
Privacy and data controlProtect privacy and limit data use
Notice and explanationMake automated involvement understandable
Human review and alternativesEnable appropriate review and recourse

The White House OSTP published the Blueprint in 2022.

The principles are recommendations in a non-binding document, not a compliance test.

Sources and references5

A Guide to Ask Questions, Not a Norm to Apply Automatically

For users, students, and technical teams, the Blueprint remains useful as a framework for formulating questions: what decision receives automated support, who might be affected, what data intervenes, and what happens if there is an error. Its approach allows comparing risks and safeguards without assuming that all uses of automation are equivalent.

To know which obligations apply in a real case, one must identify the applicable instrument, its date, the issuing authority, and the relevant jurisdiction. An obligation may proceed from a law, regulation, contract, or policy distinct from the Blueprint. This verification is especially important when employment, credit, health, benefits, or other essential services are at stake.

The most precise reading combines both ideas: the 2022 document offers substantive recommendations on security, equity, privacy, transparency, and human review; its own declaration limits its legal force. Understanding this difference avoids presenting a historical proposal as current federal policy and, at the same time, allows leveraging its questions to analyze how automated systems are designed and used.