What is the US Blueprint for an AI Bill of Rights?

Three weeks ago, OSTP, the White House's Office of Science and Technology Policy, published a document called Blueprint for an AI Bill of Rights. OSTP's job is to give the President advice on scientific and technical questions. Its blueprint describes what people in the United States should be able to expect when software helps decide whether they get a job, a loan, an apartment or medical care.

The word "blueprint" matters. The document says plainly that it is non-binding: it is not official government policy, it changes no existing law, and it gives no one a new right they can take to court. MIT Technology Review points out the reason: OSTP cannot make laws. Any new legal duties would have to come from Congress, state legislatures or regulators. What the blueprint offers instead is a shared vocabulary and a checklist that companies, agencies and lawmakers can borrow.

What counts as an "automated system"

The blueprint deliberately avoids defining "AI" narrowly. It notes that harm can come from simple software as well as from advanced machine learning, so it uses a two-part test. A system is in scope if it is automated and if it could meaningfully affect people's rights, their opportunities, or their access to critical resources and services.

In practice, that covers areas such as employment, education, housing, credit, health care, social services and government benefits. A program that recommends songs is not the focus. A program that ranks job applicants, flags tenants as risky or sets someone's insurance price is.

The document's introduction gives its reasons in concrete terms. It mentions patient-care systems that turned out to be unsafe or biased, and hiring and credit algorithms found to repeat old inequities or introduce new ones. Alongside the five principles, OSTP published a longer handbook called From Principles to Practice, with practical steps for designers and policymakers.

A worked example: an automated hiring tool

Principles are easier to understand when you apply them to one real situation. Imagine a mid-sized company that buys a résumé-screening tool. The software reads each application, gives it a score from 0 to 100, and sends only the top scores to a human recruiter. Here is how the five principles would shape that tool if the company followed the blueprint.

None of these steps requires knowing how the scoring model works inside. They are questions anyone can ask: Was it tested? Is it fair across groups? What data does it use? Was I told? Can I reach a person?

How people are reacting

Reaction has been mixed. MIT Technology Review reports that Marc Rotenberg of the Center for AI and Digital Policy called the blueprint impressive and "clearly a starting point." Russell Wald of the Stanford Institute for Human-Centered AI said it lacks details and enforcement mechanisms. Some advocates would prefer rules closer to the European Union's proposed AI Act, which would add extra checks to the uses with the most potential for harm.

For a beginner, the useful takeaway is that the blueprint is a reference point. When you read a news story about an algorithm that denied someone benefits or rejected a job applicant, the five principles give you a structured way to ask what went wrong.

What protections does it propose?

The blueprint proposes five protections, each phrased from the point of view of the person affected. The first is protection from unsafe or ineffective systems: automated tools should be developed with input from the communities and experts concerned, tested before they are used, monitored afterward, and evaluated independently, with results published when possible. The second is protection from algorithmic discrimination, meaning unjustified differences in how a system treats people because of race, sex, age, disability or another protected trait. That calls for equity checks during design, representative data and ongoing testing for unequal outcomes.

The third protection is data privacy. Systems should collect only the data a task strictly needs, ask permission in clear language, and avoid confusing designs that push people into sharing more. The blueprint asks for extra care with sensitive areas such as health, finance, schooling, work and criminal justice, and with data about young people, and it argues against continuous surveillance in schools and workplaces.

The fourth is notice and explanation: you should know when an automated system is being used on you and be able to understand, in plain language, how it contributed to a decision. The fifth is a human alternative: where appropriate, you should be able to opt out in favor of a person, and if the system makes a mistake you should be able to reach someone who can review and fix it quickly. In areas like employment and health, the blueprint asks for human consideration before high-risk or adverse decisions. Because the document is guidance, how much of this becomes required depends on future laws and on the choices organizations make on their own.