Values and design
The hard part of AI is not making it powerful. It is deciding what power must preserve.
Use this reading map to move from broad values such as dignity, fairness, agency, and care into concrete questions a team can ask before an AI system changes somebody's options.
A direct answer
Human values do not arrive as a single setting that can be switched on inside a model. They show up in the people invited into a decision, the kinds of harm that count, the trade-offs a system is allowed to make, and the routes available when an automated answer is wrong. A useful values conversation therefore starts with a situation, a person affected by it, and a decision that could be changed.
Visual atlas
This visual passage connects questions about values, language, knowledge, emotion, mind, work, and shared futures. Return to this page for the deeper reading on artificial intelligence and human values.
Value lens
Choose a lens. The prompt below is a starting point for a design review, not a verdict about the system.
Could this outcome make a person feel reduced to a score, category, or efficiency target?
01
A system that reduces waiting time may still reduce a person's ability to explain an unusual circumstance. A system that treats everyone identically may still reproduce an uneven starting point. These are not contradictions to eliminate with a slogan. They are reasons to name the values that collide in a particular setting. Ask whose time, privacy, safety, dignity, opportunity, or voice is being protected—and who absorbs the remaining cost.
02
“Respect autonomy” becomes more useful when it produces a design choice: show a meaningful explanation, allow a person to decline an automated path, preserve a handoff to a human, or keep a record that can be challenged. “Fairness” becomes more useful when it specifies a comparison group, a measurement window, a harm threshold, and a remedy. The point is not to turn ethics into a checklist; it is to make vague approval impossible.
03
An AI system can be technically accurate and still be wrong for a person whose context was not represented. A values-aware system leaves a path to contest a result, correct missing context, and learn from recurring errors. That path matters most in high-consequence settings, but the habit belongs everywhere: a recommendation, a ranking, a content filter, a hiring screen, or an assistant drafting a decision.
A closer look
Before a team says that a feature is fair, it helps to draw the decision in plain language. Who receives an outcome, what information is considered, what information is missing, and what happens after a disagreement? That small map often reveals that two worthy values are in tension: speed against explanation, personalization against privacy, consistency against the ability to make an exception. Naming the tension is more honest than claiming that one abstract principle settled it.
A useful record does not need to be bureaucratic. It can be a one-page decision note that names the affected groups, the value at risk, the evidence the team has, the uncertainty it cannot remove, and the safeguard that changes the product. Such a note makes it easier for a designer, engineer, reviewer, or community representative to point to the same choice and ask whether the safeguard is still working after the launch date.
Values work also needs a rhythm. Revisit the decision when the model, data, audience, or consequence changes. Review complaints and near misses rather than waiting for a dramatic failure. If people repeatedly need a human override, a clearer explanation, or a way to correct their record, that is evidence about the system—not noise around it. A responsible design keeps enough traceability for those patterns to become a reason for revision.
Consider a public-benefits screening tool that improves consistency by using a fixed set of inputs. The same consistency may become unfair if a recent illness, a change in housing, or a missing document cannot be represented. The useful response is not to decide that consistency or compassion always wins. It is to decide which cases deserve a second look, what explanation a person receives, and who has the authority to correct the record. The quality of the system becomes visible in those concrete arrangements: a person can understand the result, supply context, and receive a timely review. Teams should also notice when an apparently small feature changes the balance of power. A default ranking, a confusing consent screen, or a silent data-sharing rule can shape agency just as much as an explicit decision. Good values work follows these details through the full experience, including the moment a person tries to leave, appeal, or ask for help.
Put it to use
Pick one decision an AI system will influence. Do not begin with the model. Begin with the person who receives the outcome, then make the value visible as a choice that person can notice and challenge.
Keep the boundary visible
Values language cannot certify a system as good. It can only make the choices, omissions, and accountability routes easier to inspect. The work continues after launch, when real people encounter consequences no workshop predicted.
Questions readers ask
Alignment is one way to describe the effort to make a system pursue intended goals. Human values are broader: they include the legitimacy of the goal, who defines it, what trade-offs are acceptable, and how people can challenge the result.
Usually not. Values are often plural and situated. A responsible process makes the relevant perspectives and disagreement visible instead of presenting one group's preference as universal.
Choose one decision, identify the people affected, name the value at stake, and add one observable safeguard such as explanation, consent, review, or appeal.