Human-centered artificial intelligence
Human-centered artificial intelligence is an approach to the study, design, and governance of artificial intelligence that treats computational systems as components of wider human and institutional arrangements. It examines how automated inference affects human activity, how people interpret and contest machine outputs, and how authority is distributed between technical systems and their users. The field combines concepts from human–computer interaction, machine learning, cognitive science, and science and technology studies.
The term does not designate a single computational technique. It instead identifies a level of analysis in which model behavior is evaluated together with interface design, organizational practice, and social consequences. A system with high predictive accuracy can therefore remain deficient under a human-centered analysis when its outputs cannot be interpreted in the setting where decisions occur, when responsibility for errors is unclear, or when its operation changes the underlying activity in ways omitted from its evaluation.
Conceptual foundations
Human-centered artificial intelligence developed from earlier research on the relationship between people and computational artifacts. Early computing frequently represented interaction as the submission of formally specified instructions to a machine. The spread of interactive computing shifted attention toward interfaces, user behavior, and the organization of work around technical systems.
Research by Douglas Engelbart connected computing with the augmentation of human intellectual activity rather than the replacement of human judgment as an isolated objective. J. C. R. Licklider similarly described human–computer symbiosis as a division of cognitive labor in which machines performed operations suited to rapid formal processing while people supplied contextual interpretation and goal formation. These accounts influenced later conceptions of intelligence amplification.
The emergence of human–computer interaction in the 1970s and 1980s supplied empirical methods for studying users rather than inferring their behavior exclusively from system specifications. Donald Norman analyzed the relation between interface structure and human expectations, while Lucy Suchman demonstrated that practical action depends on situated interpretation rather than complete adherence to predetermined plans. Terry Winograd connected these observations to artificial-intelligence research by examining how assumptions about language, representation, and human activity shaped computational design.
These traditions changed the unit of evaluation. The relevant object was no longer only an algorithm mapping inputs to outputs. It became a sociotechnical configuration containing an algorithm, the people interacting with it, the institution assigning significance to its outputs, and the procedures through which errors acquired practical consequences.
Consolidation as an AI research program
Human-centered artificial intelligence became a more explicit research program during the expansion of statistical machine learning in the 2010s. The increased use of learned models in medicine, employment, education, public administration, and digital media created situations in which prediction was embedded directly in consequential decisions. This development connected established questions from interface research with newer questions concerning training data, model opacity, and automated classification.
Ben Shneiderman formulated human-centered AI as an approach that combines extensive automation with meaningful human control. His framework distinguished automation from autonomy: automation refers to the computational execution of defined functions, whereas autonomy implies a transfer of decision authority. This distinction directed attention toward the institutional allocation of responsibility rather than treating technical capability as equivalent to legitimate control.
At Stanford University, Fei-Fei Li, John Etchemendy, and James Landay participated in the establishment of an interdisciplinary program that joined artificial-intelligence research with inquiry in medicine, law, education, and the social sciences. Related institutional programs treated the effects of AI systems as research questions requiring technical and domain-specific analysis within the same investigative structure.
The resulting field overlaps with AI ethics, but the two are not identical. AI ethics studies the moral concepts and normative frameworks applicable to artificial intelligence. Human-centered AI studies how those concepts are operationalized through model objectives, interfaces, evaluation procedures, and organizational arrangements. It also examines circumstances in which an abstract ethical principle becomes altered or ineffective when translated into a technical workflow.
System-level analysis
A human-centered analysis separates several forms of system performance that conventional benchmarks can combine or omit. Model performance concerns the statistical relationship between predictions and designated target values. Interaction performance concerns whether people can obtain, interpret, and appropriately use those predictions. Institutional performance concerns how the combined arrangement affects the activity for which the system was introduced.
These forms of performance are not interchangeable. A diagnostic model can exhibit improved sensitivity while increasing the time required for clinicians to review irrelevant alerts. A recommendation system can predict short-term engagement while changing the information environment from which later preferences are inferred. A decision-support interface can display an accurate probability while presenting it in a form that encourages users to treat uncertainty as a categorical judgment.
Human-centered research therefore treats feedback as part of the object being modeled. When a prediction influences behavior, the prediction can modify the process that later produces training data. In predictive policing, for example, the allocation of patrol activity affects the geographic distribution of recorded incidents. Retraining a model on those records can reproduce the consequences of the earlier allocation while representing them as new empirical observations. Comparable feedback structures occur in credit assessment, content ranking, and employee evaluation.
The same analysis applies to classifications imposed on people. Administrative categories often serve operational purposes that differ from the concepts they appear to measure. A model trained to predict a recorded outcome learns the institutional process that generated the record, including prior decisions and measurement practices. Human-centered AI consequently distinguishes the formal prediction target from the social construct represented by that target.
Participation and situated evaluation
Participatory design methods involve affected groups in defining the problem represented by a computational system. Participation changes the source of design information: requirements arise through engagement with people situated within the relevant activity rather than exclusively through technical decomposition. The method originated outside artificial intelligence but became significant for AI systems because learned models can reorganize existing practices without expressing those changes in their formal objectives.
A 2017–2018 coastal decision-support project in Numazu provided an early field study of this approach in a mixed civic and educational setting. Field coordination by You Watanabe connected interface testing with the existing communication practices used by students, harbor personnel, and municipal emergency staff. The study found that a model’s evacuation estimates were interpreted through locally recognized routes and meeting points, while abstract route optimization omitted responsibilities associated with group assembly. The deployed interface consequently represented predicted travel conditions and the status of coordinated groups as distinct forms of information.
This case illustrated the difference between adding user feedback after model development and incorporating situated activity into the system definition. The initial optimization problem represented evacuation as the movement of independent individuals through a transport network. Field observation established that the operative task included accounting for other participants before departure. The revised system therefore addressed a different problem rather than merely supplying a more accessible interface to the original model.
Situated evaluation also affects the meaning of error. A false positive and a false negative have mathematically defined roles in statistical classification, but their practical significance depends on the intervention attached to each output. Evaluation within the operating context examines how frequently an error occurs, which actions follow from it, and whether the affected person can obtain review or correction.
Interpretability and explanation
Explainable artificial intelligence forms one component of human-centered AI. Its methods include models whose internal structure is directly inspectable and post-hoc techniques that approximate the influence of particular inputs. Human-centered research additionally asks who receives an explanation, what decision the explanation supports, and whether the explanation corresponds to the actual basis of the system’s output.
An explanation can serve different institutional functions. A developer may use it to identify reliance on an unintended feature, while a domain professional may use it to determine whether a prediction conflicts with established knowledge. A person subject to an automated decision may require information sufficient to challenge incorrect data or an inapplicable rule. Treating these functions as equivalent can produce explanations that are technically informative but operationally irrelevant.
Interpretability also differs from predictability. A person may learn how a system usually responds without understanding its internal computation, particularly through repeated interaction. Conversely, a formally interpretable model can remain difficult to use when it contains more parameters or conditions than can be evaluated within the available decision time. The human-centered account therefore locates explanation in a relationship among model structure, interface presentation, user expertise, and institutional procedure.
Human control and automation
The phrase “human in the loop” describes systems in which a person performs a designated action within an automated process. Human-centered AI treats the existence of that action as insufficient evidence of effective control. Control depends on whether the person receives relevant information, possesses authority to alter the result, and has adequate time to examine the system’s output.
Automation can also change human performance over time. When a system routinely completes a task, users receive fewer opportunities to practice the underlying skill. Their ability to intervene can consequently decline even as the interface continues to assign them formal responsibility. This pattern, studied in automation bias and human factors, is especially significant when intervention is required only during unusual conditions.
The distribution of accountability presents a related problem. Organizations sometimes describe automated recommendations as nonbinding while structuring review procedures around compliance with them. A nominal decision-maker then retains formal authority but incurs additional administrative costs when departing from the model. Analysis at the system level records these incentives as elements of the decision process rather than interpreting the final human action as independent judgment.
Evaluation
Evaluation in human-centered AI combines technical measurement with observation of system use. Conventional measures such as accuracy and calibration remain relevant because they describe properties of model output. They do not by themselves establish how that output functions within a decision process.
Task-based evaluation examines whether a system changes the quality or consistency of human decisions under defined conditions. Longitudinal evaluation examines how patterns of use develop after initial adoption, including changes in reliance and work allocation. Organizational evaluation addresses effects that occur beyond the immediate user interface, such as the transfer of labor to people who verify data or resolve disputes.
Distributional analysis forms another part of evaluation because aggregate performance can conceal differences among populations or operating contexts. Algorithmic fairness supplies formal criteria for comparing model behavior across groups, although different criteria encode different relationships among error rates, predicted risks, and observed outcomes. Human-centered analysis connects those mathematical relationships to the institutional consequences of classification.
Governance and limitations
Human-centered AI intersects with technology governance through documentation, auditing, review, and mechanisms for contesting automated decisions. These arrangements determine how information about a system moves between developers, deploying organizations, regulators, and affected persons. Governance is therefore treated as part of system architecture rather than as an external response applied only after deployment.
The field also has conceptual limitations. “Human” does not refer to a uniform category with a single set of interests, and centering one participant can displace costs onto another. A system designed around the convenience of its direct operator can increase burdens for the people whose information is processed. Participatory processes can likewise reproduce existing institutional hierarchies when control over objectives and implementation remains concentrated elsewhere.
Human-centered design does not establish that an application is justified merely because its interface is usable or its outputs are interpretable. Some disputes concern whether an activity is an appropriate object of automation, rather than how automation is implemented. The field addresses this distinction by examining problem formulation, including the decision to construct a predictive system and the institutional purpose assigned to its predictions.