Blueprint for an AI Bill of Rights

Blueprint for an AI bill of rights making automated systems work for the american people october 2022

The framework recognizes that law enforcement and national security activities may require adjusted application of its principles to protect sensitive information and must remain subject to existing judicial, privacy, and civil liberties oversight. The document creates no enforceable legal rights or defenses against the United States or any other entity.

Section title foreword

Automated systems have created serious harms through bias and discrimination in healthcare, hiring, credit, and surveillance, but these outcomes are not inevitable. The White House has identified five principles to direct construction and implementation of automated systems that protect the public while advancing civil rights and democratic values as foundational American principles.

About this framework

This framework consists of five principles and practices for directing how automated systems should be designed, built, and deployed to protect American rights, civil liberties, and privacy. It provides a national values statement and toolkit applicable across sectors, with a technical handbook offering concrete implementation steps for organizations at all levels.

Listening to the american public

The framework was shaped through a year-long process involving public listening sessions, panel discussions, and formal requests for information from impacted communities, industry stakeholders, technology developers, and policymakers on algorithmic and data-driven harms. The core findings were that artificial intelligence has transformative potential and that preventing technology harms is both necessary and achievable.

Section title blueprint for an AI bill of rights safe and effective systems

People should be protected from unsafe or ineffective automated systems, which must be developed with consultation from diverse communities and experts, undergo pre-deployment testing and ongoing monitoring, and may not be deployed or must be removed if risks are identified. Systems must not be designed to endanger safety and should be protected from inappropriate data use and unintended foreseeable harms, with independent evaluation and public reporting of safety measures.

Algorithmic discrimination protections

Section title data privacy

Individuals should receive built-in privacy protections and have control over their data. Designers and developers should obtain permission before collecting, using, accessing, transferring, or deleting personal data, limit collection to what is necessary, and use privacy-by-design defaults rather than privacy-invasive settings. Consent should only apply where it can be meaningfully given through brief, plain language requests.

Notice and explanation

Individuals should receive clear notification that automated systems are in use, understand their role in outcomes affecting them, and obtain explanations of how and why decisions were made. Designers and developers must provide accessible plain language documentation identifying the responsible party, describing overall system functioning, and explaining outcomes in ways proportionate to risk. Notification should be kept current and people should be informed of significant changes.

Section title human alternatives, consideration, and fallback

Individuals should be able to opt out of automated systems in favor of human alternatives where appropriate and have timely access to human review and remedy through fallback and escalation processes when systems fail, produce errors, or are appealed. In sensitive domains including criminal justice, employment, education, and health, automated systems should include meaningful human oversight, appropriate operator training, and human consideration of adverse or high-risk decisions. Human governance processes should be accessible, equitable, and effective, and their timeliness, accessibility, and effectiveness should be publicly reported where possible.

Section title

Rights, opportunities, or access

The framework protects civil rights, civil liberties, privacy, freedom of speech, and voting in both public and private contexts; equal opportunities in education, housing, credit, employment, and programs; and access to critical services including healthcare, financial services, safety, social services, truthful goods and services information, and government benefits. Implementation measures should be proportionate to the extent and nature of potential harm.

Relationship to existing law and policy

Some protections outlined in this framework are already mandated by the U.S. Constitution or existing laws, such as requirements for judicial oversight of government surveillance and data seizure, human review of criminal investigations, and civil rights protections against discrimination.

Section title

Relationship to existing law and policy

The framework acknowledges that existing regulatory requirements cover medical devices and sector-specific privacy protections, and implementing additional protections may require new legislation. Departures from the framework's principles may be justified when needed to comply with current laws, address specific operational constraints, or balance competing public interests, particularly in law enforcement contexts. The framework serves as guidance for moving principles into practice across government and the private sector, adapting elements of global privacy principles relevant to automated systems without restricting them to a single domain such as privacy or ethics alone.

Definitions

Using this technical companion

The framework comprises five principles and associated practices guiding development, implementation, and operation of automated systems to protect public rights in artificial intelligence contexts. For each principle, the technical companion offers examples and concrete actions for communities, industry, governments, and others to incorporate protections into policy and technological design. The companion serves as a reference for anyone affected by or involved in creating, deploying, evaluating, or governing automated systems.

Hy this principle is important

This section provides a summary of the problems that each principle addresses and protects against, including illustrative examples.

Hat should be expected of automated systems

Ow these principles can move into practice

Safe and effective systems

You have the right to protection from systems that are unsafe or ineffective. Developers must consult diverse communities and experts to identify risks, conduct evaluations before deployment and afterwards, and mitigate unsafe outcomes including those beyond intended uses. Systems must not be designed to endanger safety and should protect against foreseeable harms from unintended uses.

Safe and effective systems why this principle is important

This section explains problems that the safe and effective systems principle addresses, including technology deployed without proven effectiveness, systems that fail to work as intended or promised, and harm caused by unintended uses or purposeful misuse. Many preventable harms already result from inadequately deployed protections, and the section advocates for expanded, proactive safeguards to build public confidence while establishing clear standards for developers. This section provides real-world examples of system failures: a sepsis prediction model that underperformed clinical claims and triggered excessive false alerts; automated content moderation that silenced counter-speech from Black users; tracking devices designed for lost items but misused for stalking despite partial manufacturer fixes; and a policing algorithm that perpetuated unfair geographic targeting due to feedback loops from historical arrest data.

Safe and effective systems why this principle is important

This section illustrates harms from deepfake nudification tools that disproportionately target women through non-consensual intimate imagery, and from AI camera systems used to evaluate driver safety that penalized drivers for events outside their control, resulting in unfair bonus ineligibility.

Safe and effective systems what should be expected of automated systems

Automated systems must proactively protect the public from harm, avoid using irrelevant or inappropriate data, and demonstrate safety and effectiveness through extensive pre-deployment testing mirroring real-world conditions and comparing performance against existing human-driven procedures as a baseline and minimum ongoing standard, with public consultation across all development phases and early engagement of diverse affected communities and relevant experts including civil rights and privacy specialists (though government and law enforcement applications may require confidential engagement). Before and during deployment, potential risks must be identified and mitigated, focusing on meaningful impact to rights and opportunities across affected communities including those not direct users, and risks from intentional misuse or consultation feedback. Systems intentionally designed to violate safety must not be developed or used; those with unintended safety violations must not be used until risks are mitigated, and ongoing risk mitigation may require rollback or substantial modification after launch.

Safe and effective systems what should be expected of automated systems

Automated systems must undergo ongoing monitoring with continuous performance evaluation, harm assessment, updates, and machine learning retraining to maintain acceptable performance over time under changing conditions, along with fallback mechanisms and manual human verification of predictions beyond operator judgment. Organizations must establish clear governance structures with specific responsibility assigned high enough to decide on resources, mitigation, and rollback, ensure decision-makers are aware of systems with meaningful impact on rights and opportunities, and may conduct independent ethics review before deployment. Data used in automated systems must be relevant (demonstrated through research-backed causal evidence or reasonable domain expectations rather than historical connection alone), high-quality, representative, with measured and limited entry errors, and suited to the specific task, with particular quality scrutiny for predicted outcomes or labels; documentation must justify each data attribute and source, explain appropriateness and legal compliance, and provide overall descriptions for high-dimensional or derived attributes.

Safe and effective systems what should be expected of automated systems

Organizations developing or deploying automated systems must track derived data as potentially high-risk inputs, validate them against collateral consequences, and provide extra oversight for data from criminal justice, finance, employment, and housing when reused in new contexts, reusing such data only where legally authorized and where benefits to impacted individuals outweigh risks after reasonable mitigation measures are applied. Independent evaluators including researchers, journalists, auditors, and ethics boards must have timely and genuine access to deployment-ready systems and associated data samples to perform evaluations, with protections ensuring access cannot be revoked without reasonable justification. Entities responsible for automated systems must provide regularly updated reports in plain language and machine-readable format covering system overview and goals, data sources and quality issues, public consultation results, risk assessments and mitigation steps, performance testing results including accuracy and demographic impact analysis, ongoing monitoring procedures and results, and independent evaluation procedures and findings.

Safe and effective systems how these principles can move into practice

Federal agencies under Executive Order 13960 must follow nine principles when designing AI systems: lawfulness and respect for national values, purposefulness and performance, accuracy and effectiveness, safety and security, understandability, responsibility and traceability, regular monitoring, transparency, and accountability. The National Institute of Standards and Technology is developing a voluntary risk management framework through a consensus-driven collaborative process to help organizations incorporate trustworthiness considerations including accuracy, explainability, reliability, privacy, robustness, safety, security, and bias mitigation into AI design, development, use, and evaluation. Industry provides risk assessments, auditing mechanisms, dashboards for monitoring, and documentation procedures to mitigate AI safety and efficacy risks, while the Office of Management and Budget encourages meaningful stakeholder engagement in program and service design.

Safe and effective systems how these principles can move into practice

The Department of Energy, Department of Defense, and Intelligence Community have established frameworks and councils for ethical AI use in their specific contexts. The National Science Foundation funds research on safe, secure, and effective AI through multiple programs supporting trustworthy algorithms, safe autonomous systems, cybersecurity, formal verification of automated systems, and accountable software development. Idaho Code Section 19-1910 requires pretrial risk assessments to be validated as free of bias against protected classes before use, with public access to supporting documents and records, prohibiting trade secret claims from preventing discovery in criminal cases.

Why this principle is important

Automated systems can produce unequal outcomes and reinforce existing unfairness, particularly when data reflects historical biases, affecting decisions across hiring, lending, healthcare, and criminal justice. While some organizations have already begun testing and developing standards to address algorithmic bias, more comprehensive safeguards are needed to protect all people from discrimination in their interactions with automated systems, especially in underserved communities. Automated systems used for loan pricing, hiring, student risk assessment, criminal recidivism prediction, and other purposes have demonstrated racial and ethnic disparities, including charging higher rates to applicants from historically Black colleges, rejecting qualified women applicants, and misclassifying students and individuals based on demographic characteristics.

Why this principle is important

Automated systems including sentiment analyzers, search engines, advertisement delivery, airport body scanners, remote proctoring, and healthcare algorithms have shown bias against religious minorities, racial groups, gender minorities, transgender people, people with disabilities, and other populations, leading to discrimination in content moderation, search results, job advertising, security screening, education, and healthcare.

What should be expected of automated systems

Developers and deployers must test automated systems before sale or use to ensure they are free from algorithmic discrimination and should conduct proactive equity assessments during design and acquisition phases to review data, context, and accessibility impacts. Data used in system development must be representative of local communities where the system will be deployed and reviewed for historical biases, with sufficiently robust samples to identify and mitigate potential harms. Developers should avoid directly using demographic information in system design or deployment, and must proactively test for proxy attributes that correlate with demographic features and could cause discrimination, either removing proxies or ensuring they are not given undue weight while monitoring for resulting bias.

What should be expected of automated systems

Automated systems should be designed with accessibility for people with disabilities and tested using broad demographic measures before and after deployment to identify disparities; when disparities are found against protected groups, entities must document them and justify continued use, with disparities showing potential for discrimination or meaningful harm to be mitigated by selecting models with less adverse impact, modifying data choices, or reconsidering system use. Systems must be regularly monitored to catch discrimination arising from unforeseen interactions, system changes after deployment, or changes in context or data, with the deploying entity examining whether the system causes discrimination when used through assessments performed regularly or whenever unusual patterns occur using methods suited to available demographic information, and higher-impact systems requiring more frequent monitoring.

What should be expected of automated systems

Entities should allow independent evaluation of algorithmic discrimination in their systems, with public sector evaluations made public unless prevented by law enforcement or national security needs. Entities must provide an algorithmic impact assessment reporting design equity assessments, accessibility testing, disparity testing, remaining disparities, and mitigation efforts, made public where possible in plain language and machine-readable form.

How these principles can move into practice

Real-world applications include federal initiatives to combat lending discrimination and redlining, equal employment guidance on how AI may violate disability protections, research identifying healthcare algorithm bias against Black patients, employer best practices for evaluating hiring software, and existing accessibility standards from the Access Board, ISO, and W3C. NIST Special Publication 1270 addresses stakes, challenges, and categories of AI bias (systemic, statistical, and human), and describes broad mitigation challenges around datasets, testing, and human factors, using a sociotechnical perspective.

Data privacy

Individuals should have default privacy protections ensuring data collection matches reasonable expectations and collects only what is necessary. Designers should seek permission for data collection, use, and deletion, with consent used only when meaningful; notice should be brief and in plain language. Sensitive domains including health, work, education, and finance should have enhanced protections with restrictions on use and ethical review; surveillance technologies need pre-deployment assessment and heightened oversight, and continuous surveillance should not be used in education, work, or housing.

Data privacy why this principle is important

Data collection and surveillance by companies and government agencies have expanded without adequate legal constraints, creating risks of discrimination, mental health harms, threats to democratic processes, and unauthorized monitoring in an increasingly automated society, while the American public often lacks knowledge of or control over data harvesting that fuels automated systems affecting their lives. The United States lacks comprehensive statutory or regulatory protections for personal data across all contexts, with only scattered sectoral laws, requiring additional protections.

Data privacy why this principle is important

Provides four examples of privacy harms: insurers using social media to set rates, data breaches exposing consumers to identity theft, facial recognition in public housing shared with law enforcement, and employers tracking union organizing discussions.

Data privacy what should be expected of automated systems

Requires automated systems collecting personal data to embed privacy protections by default, conduct privacy risk assessments throughout development, minimize data collection, and limit collection to specific narrow goals with established retention timelines. Data should not move to different contexts without reassessing privacy risks and obtaining new consent where appropriate. Requires entities managing automated systems to follow privacy and security best practices that prevent data and metadata leakage beyond consented uses, using privacy-enhancing technologies and cryptography alongside conventional security protocols.

Data privacy what should be expected of automated systems

Surveillance systems must undergo assessment before and during deployment to prevent algorithmic discrimination, be strictly necessary and proportionate, use the least invasive means, and must not monitor the exercise of democratic rights or target people based on identity characteristics, with continuous monitoring prohibited in workplaces, educational institutions, and public accommodations. Data collectors seeking consent must request it for specific narrow uses, defined time periods, and particular entities, with consent expiring if any condition changes and requiring renewal before new uses or transfers, using plain language that users can understand, making requests accessible to people with disabilities, and avoiding manipulative design tactics, with refusal allowed without penalties where feasible.

Data privacy what should be expected of automated systems

Automated systems should allow individuals to access and correct their personal data, know who accesses it, and obtain consent before data sharing with other organizations. Entities must support individual use of automated systems to manage consent and data control decisions through machine-readable formats and metadata tags. Entities should allow independent evaluation of data policies and publish results publicly.

Data privacy extra protections for data related to sensitive domains

Data privacy extra protections for data related to sensitive domains

Data privacy what should be expected of automated systems

Systems collecting, using, sharing, or storing sensitive data must limit use to functions strictly necessary for the domain or required for administration unless additional optional consent (not required or incentivized as a condition of access) is obtained; any system use of sensitive data that might restrict rights or opportunities must undergo thorough ethical review in advance and periodically, and some novel uses may require human subject experimentation review. Entities must regularly audit data quality in sensitive domains and promptly correct inaccuracies, and must not sell, share, or make sensitive data or data derived from it public through data brokerage, limiting access based on necessity and proximity to the data subject. Entities in sensitive domains must regularly report in clear and machine-readable formats on data security breaches involving sensitive data, the number and outcomes of ethical pre-reviews, descriptions of any shared data and how it was assessed for sensitive risk, and ongoing risk management procedures.

Data privacy how these principles can move into practice

Federal privacy protections include the Privacy Act of 1974, which limits personal data storage to only what is relevant and necessary, grants individuals access to their information, and provides remedies including correction of inaccurate records or monetary damages, while the NIST Privacy Framework offers voluntary guidance for organizations to manage privacy risks; a state-wide biometrics moratorium resulted from a school's facial recognition surveillance plan without community input. Meaningful data agency is demonstrated through well-designed privacy controls with contextual, timely alerts, and best practices for privacy by design include transparency about how user data is used, requesting app permissions during use, and ensuring apps function when users deny or revoke permissions.

Notice and explanation

Designers and deployers should provide accessible plain language documentation of how automated systems work and their role in outcomes, identify responsible parties, and explain decisions that are technically valid, meaningful, and calibrated to risk level. Notice should be kept current and updated when the system's use or functionality changes significantly. Summary reports on system explanations and documentation quality should be made public where possible.

Notice & explanation why this principle is important

Automated systems increasingly determine opportunities across employment, credit, courts, and education, but their impact is often invisible to affected people who cannot contest or correct errors without knowing whether automation was involved or how decisions were made. System complexity is sometimes used to avoid explaining important decisions despite clear notice being necessary for safety and accountability. Cases demonstrate how lack of notice prevents both system operators and affected individuals from validating results or contesting decisions, including when data entry errors cause undetected eligibility changes in benefits systems.

Notice & explanation why this principle is important

A predictive policing system identified individuals at risk of gun violence without explaining how it made determinations, preventing police and the public from understanding its reasoning. A benefits system made invisible changes to criteria, causing denials that went undiscovered until explanations were demanded and revealed data entry errors and system flaws.

Notice & explanation what should be expected of automated systems

Notice & explanation what should be expected of automated systems

Automated systems should incorporate mechanisms for explaining their behavior, with the extent and design of explanation tailored to the assessed risk level and context, using only fully transparent models in high-consequence settings. Explanations must accurately reflect the factors influencing decisions in a scientifically supportable manner, including error ranges where appropriate. Organizations must document responsible entities, system purposes, risk assessments, and explanation accessibility in clear, plain language reporting.

Notice & explanation how these principles can move into practice

Existing laws and industry practices demonstrate implementation of notice and explanation principles: Illinois requires written notice when private entities collect biometric information; major technology companies are developing frameworks to communicate transparently about automated systems including safety and disparity assessments; federal lending laws require adverse action notices and credit score disclosures when credit decisions are denied. California requires warehouse employers to provide written descriptions of algorithmic quota systems, including task specifications and potential adverse employment consequences. Federal research agencies including NIST and the Defense Advanced Research Projects Agency conduct research on explainable AI measurement methods and techniques to improve transparency while maintaining performance.

Human alternatives, consideration, and fallback

People should be able to opt out of automated systems in favor of human alternatives where appropriate based on context and broad accessibility needs, and should have access to timely human review and remedy through fallback processes when systems fail or produce errors. Systems used in sensitive domains like criminal justice, employment, education, and health require additional safeguards including human oversight of adverse or high-risk decisions, training for operators, and public reporting on timeliness and effectiveness of human review processes.

Human alternatives, consideration, and fallback why this principle is important

Automated systems can fail, reinforce bias, or be inaccessible, yet the public often lacks convenient alternatives, resulting in delayed or lost access to rights and services; the American public deserves the ability to conveniently opt out without disadvantage and to have human fallback systems for unexpected failures, particularly in time-critical scenarios. Criminal justice, employment, education, and healthcare demand extensive human oversight as safeguards against unfair or dangerous outcomes. Human review is particularly important in areas like mail-in ballot signature matching, which works less accurately for voters with disabilities, shorter or hyphenated names, or name changes, and serves as a necessary check on automated system validity even when fallback mechanisms cannot guarantee correct outcomes.

Human alternatives, consideration, and fallback why this principle is important

Colorado's unemployment benefits system denied access when no human identity verification alternative existed; fraud detection systems have incorrectly flagged entries without providing opportunity for human explanation; hospital medication software caused patient harm when doctors feared overriding its error; automated HR systems fired workers without human review or appeal.

Human alternatives, consideration, and fallback what should be expected of automated systems

Automated systems should enable users to easily opt out in favor of human alternatives where appropriate, with clear and accessible notice and instructions assessed for effectiveness. When automated systems are part of access processes, timely human alternatives should be available without unreasonable burden, and proportionate human consideration with fallback and escalation mechanisms should be provided, tested for accessibility by users with disabilities, and structured to be no more burdensome than the automated option.

Human alternatives, consideration, and fallback what should be expected of automated systems

Fallback and escalation systems must deliver equitable outcomes compared to automated systems, especially for underserved communities, and operate on timelines appropriate to each system's context, with immediate availability for time-critical systems such as voting, building access, healthcare, and wage withholding. Human decision-makers must be able to effectively enact overruling decisions throughout system components, and staff administering or interpreting automated systems require regular training in system operation and automation bias mitigation, with ongoing assessment to prevent algorithmic discrimination and governance oversight of human-based systems.

Human alternatives, consideration, and fallback what should be expected of automated systems

Automated systems in sensitive domains such as criminal justice, employment, education, and health must use narrowly scoped data justified for specific use cases, be validated for their particular deployment context, and not make high-risk decisions without human consideration in areas like sentencing or medical care; system designers should consider waivers of confidentiality and trade secret protections to enable meaningful oversight including access to source code and documentation, with preference for fully transparent models. Organizations must publicly report regularly on human alternatives and fallback systems, including assessment of timeliness and burden, aggregate statistics on who uses human alternatives, mean and maximum wait times for requests, accessibility procedures and their evaluation results, and for sensitive domains, information on training and governance. All reporting must document system goals, data included, governance of access, and assessment of goal achievement in clear machine-readable format.

Human alternatives, consideration, and fallback how these principles can move into practice

This section illustrates practical applications of the principles through real-world examples. It describes healthcare navigators assisting people with coverage enrollment, integrated customer service systems combining automated tools with human agents for complex issues, and procedures in at least 24 states enabling voters to correct flagged ballots through contact by election officials who request alternative information or new signatures.

Appendix

This appendix illustrates the scope of automated systems covered by the Blueprint, including those affecting civil rights, opportunities, and access to critical services. Examples span content moderation, surveillance, criminal justice algorithms, voting systems, education and employment screening tools, health and financial systems, and community safety infrastructure, though the list is not exhaustive.

Appendix

The appendix continues with examples of systems impacting community safety such as traffic and grid controls, and systems managing access to benefits or penalties through decision support, fraud detection, biometric access controls, or autonomous determinations affecting services.

Appendix section title

Appendix

The section discusses two panels: one on consumer rights in AI-enabled products and services, emphasizing the need for transparency, data governance, and flexible policy to balance ownership, use rights, and community welfare while representing marginalized voices; and another on criminal justice applications of predictive policing, risk assessments, and surveillance tools, raising concerns about system validity and data bias while emphasizing that safety should be the primary focus and that transparency alone is insufficient without validity standards, early community input in design, and regulatory limits on technology type and cost.

Appendix

This section summarizes concerns raised about AI and technologies limiting equal opportunities through remote proctoring, facial recognition in education, automated hiring screening, workplace surveillance, and automated tenant screening. Panelists identified surveillance as creating chilling effects on student expression, inappropriate tenant control, and erosion of work-life boundaries that shift oversight burdens from institutions to individuals. Panelists described technical and governance protections including community input into design and use, public reporting on system elements, better notice and consent procedures, opt-out ability with human process fallback, decision explanations, institutional training, local validation of use cases, and third-party audits to maintain accountability and validity.

Appendix

This section describes panel discussions on designing technology that supports democratic values, emphasizing risks of biased data, discriminatory outcomes, and opaque decision-making in high-stakes AI applications. Panelists advocated for transparency, value-sensitive participatory design, stakeholder-targeted explanations, and public consultation, while prioritizing trust in people and engagement with impacted communities to build protections into system design.

Appendix

Panelists identified risks from technology use in fraud detection, digital ID systems for social welfare, and healthcare delivery, including feedback loops that reinforce inequality, broadband access barriers, privacy concerns, and racial biases in medical technologies. Recommended mitigations include community input in design, opt-out options with human alternatives, timely benefit payments, clear notice and explanations, and accountability to affected stakeholders. Panelists emphasized using technology to provide positive outcomes rather than deny support and centering the voices of those subjected to medical technologies.

Appendix

AHIP

Article19

Clear

Appendix

This section lists organizations and individuals who submitted responses to a request for information on artificial intelligence and participated in listening sessions conducted by the Office of Science and Technology Policy, with over 300 members of the public attending these sessions in total.

Appendix

The Office of Science and Technology Policy held meetings with various private sector and civil society organizations to gather ideas for developing the Blueprint for an AI Bill of Rights and to discuss positive use cases, potential harms, and oversight possibilities for these technologies.

(ACLU)

This section continues listing organizations from the private sector and civil society that participated in consultations with the Office of Science and Technology Policy regarding the Blueprint's development.