Human-centered AI is a design approach that aims for high human control and high computer automation at the same time, instead of treating the two as a trade-off. Computer scientist Ben Shneiderman laid out the idea in a February 2020 paper, later published in the International Journal of Human-Computer Interaction, and the U.S. National Institute of Standards and Technology (NIST) and Stanford's Institute for Human-Centered AI each describe the role of people in AI systems in their own terms. It is a design principle and a risk-management habit, not a law.
Key takeaways
- Shneiderman's framework replaces a single dial from human control to automation with two separate axes, so a design can score high on both.
- Some tasks still call for full automation (air bag deployment) or full human mastery (playing the piano), and the framework says where each fits.
- NIST's AI Risk Management Framework 1.0, released January 26, 2023, treats human oversight as a defined process that organizations choose and document, not a default.
- Stanford's institute, launched in 2019, states three aims: guided by human impact, designed to augment people, and inspired by human intelligence.
- None of these documents is binding law; they are guidance that organizations may adopt.
From one dial to two axes
Older writing on automation, which Shneiderman traces to Sheridan and Verplank, pictured a single line. At one end a person does everything, at the other the computer does, and every step toward automation was assumed to cost the user some control. His paper, first posted on arXiv on February 10, 2020, argues that the two measures are independent. A design can give users strong control and also carry out a large share of the work, and the paper treats that combination as the usual target for tools meant to be reliable, safe, and trustworthy. The journal version appeared in volume 36, issue 6 of the International Journal of Human-Computer Interaction in 2020.
The same paper names three supports for trust: technical practices that make software reliable, management strategies that build a culture of safety, and independent oversight structures. In this view, a well-designed interface is necessary but not sufficient. The organization behind it and the outside reviewers who check it matter as well.
The four quadrants and a pain-pump example
The framework is easiest to see in a table. The examples come from the paper; the last column restates its design advice in plain words.
| Quadrant | Example in the paper | Design emphasis |
|---|---|---|
| High automation, high human control | Elevators, phone cameras, and the target design for complex tasks with varying conditions | Automation does the routine work while users keep clear ways to steer, correct, and stop |
| High automation, low human control | Air bag deployment, anti-lock brakes, pacemakers | No time for human intervention, so extra care in design, testing, and monitoring after release |
| Low automation, high human control | Bicycle riding, piano playing, baking | Preserve mastery and creativity; computers may coach but should not take over the activity |
| Low automation, low human control | Clocks, mousetraps, and also land mines | Simple devices, with risk depending on what they do |
The paper also works through patient-controlled analgesia, the pump that lets a recovering patient request pain medication. A fixed drip sits in the low-low corner. A fully machine-selected dose moves it to the high-automation, low-control corner. A button the patient squeezes, with a lockout period of 6 to 10 minutes between doses and total dose limits over 1 to 4 hours, gives human control with little automation. The design the paper favors keeps the button but lets sensors and machine learning choose appropriate doses while blocking overdoses, and it explains to the patient why limits exist.
How NIST handles the human role
NIST's AI Risk Management Framework 1.0, published in January 2023, is voluntary. It lists characteristics of trustworthy AI: valid and reliable, safe, secure and resilient, accountable and transparent, explainable and interpretable, privacy-enhanced, and fair with harmful bias managed. On people, it says roles and responsibilities in decision making and overseeing AI systems need to be clearly defined and that human-AI configurations can range from fully autonomous to fully manual. An AI system may act alone, defer to a human expert, or serve as an extra opinion for a human decision maker.
The framework does not require oversight everywhere. It notes that some systems, such as models that improve video compression, may not need it, while others do. One of its mapping subcategories, MAP 3.5, asks organizations to define, assess, and document their human oversight processes. NIST's page says the framework is being revised as part of the White House AI Action Plan, with no completion date given, and a generative AI profile followed on July 26, 2024.
Stanford's institute and the word "human-centered"
Stanford launched its Institute for Human-Centered AI in 2019, with Fei-Fei Li as founding director. Its website states that AI should be guided by its human impact, designed to augment rather than replace people, and inspired by human intelligence. The phrase "human-centered" therefore covers more than interface design. It spans research aims, education, and social impact, and the three statements are goals rather than testable standards.
Questions to ask of a "human-centered" claim
- Which tasks are automated, and which stay with a person?
- Can the person see what the system is doing, correct it, or switch it off?
- Who is accountable for an error, and is that written down?
- Has anyone outside the team reviewed the design or tested it?
- Does the system get monitored after release, and are problems reported?
- Is a human check present at the point where harm could occur, or only at the end?
The label itself proves nothing. A product can use the phrase and still give users no control. This article gives general information, not legal advice or a compliance opinion, and organizations with legal duties should consult qualified counsel and the primary documents.
FAQ
What does human-centered AI mean?
In Shneiderman's usage it means designing AI so that people keep meaningful control while computers automate what they do well. Other groups use the phrase more broadly for AI that serves human interests.
Is human-centered AI required by law in the United States or Canada?
The documents covered here are guidance and research proposals, not statutes. Specific laws may require human review in particular settings, so a given use should be checked against the rules that apply to it.
Does more automation always mean less human control?
No. The two-axis framework was built to show that the measures are separate. The paper gives elevators and phone cameras as examples of tools that automate a lot while leaving users in charge.
Does every AI system need a human in the loop?
Not according to NIST's framework, which says some systems, such as video compression models, may not need oversight, while others specifically require it. The decision depends on the risk of harm.
Related reading
- How education codes shape AI use in schools
- Where human and machine intelligence differ
- How conversational interfaces developed
- Researchers behind modern AI
More in AI Policy and Ethics.
Last verified: October 2026
Sources
- Shneiderman B., Human-Centered Artificial Intelligence: Reliable, Safe and Trustworthy, International Journal of Human-Computer Interaction 36(6), 2020 (arXiv preprint, February 2020)
- Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1, National Institute of Standards and Technology, January 2023
- AI Risk Management Framework overview page, National Institute of Standards and Technology, 2026
- About, Stanford Institute for Human-Centered Artificial Intelligence, Stanford University, 2026