Behavioral human-centered AI

Behavioral human-centered AI is the design and deployment stance that treats AI adoption as a behavioral change problem, not only a technical, ethical, or interface-design problem. 2025-schweitzer-leading-ai-adoption argues that human-centered AI is incomplete if it assumes rational users, designers, and leaders will simply accept better-performing systems; bounded rationality, trust, autonomy, identity threat, loss aversion, overconfidence, escalation of commitment, and vivid failure narratives all shape whether AI becomes usable, accepted, governed, and institutionalized.

The useful move is stage-specific. In the design stage, the relevant humans are not only end users but also designers whose confirmation bias, inventor’s bias, and anchoring on technical specifications or KPIs can build systems that miss human usability, fairness, acceptance, or well-being. In the adoption stage, organizational members may overweigh vivid AI failures, job-loss stories, surveillance risks, autonomy loss, or safety concerns relative to actual evidence. In the management stage, leaders may overestimate their ability to oversee AI implementation or continue investing in systems that are ineffective or harmful because sunk costs make reversal hard.

For this wiki, behavioral human-centered AI sits between ai-adoption, ai-readiness, responsible-ai-deployment, and human-ai-collaboration. It names the behavioral layer those pages keep running into: AI systems fail not only because organizations lack use cases, data, governance, or capability, but because humans interpret, resist, trust, over-trust, defend, avoid, and institutionalize AI through predictable psychological and organizational mechanisms.

2026-farach-scaffolding-human-ai-collaboration adds an AI-specific field experiment that makes the intervention question more concrete. A required process for pairs to use Copilot together was associated with lower document quality and far fewer completed documents than unstructured use. Training individuals to treat AI as a thought partner did not improve average quality on the study’s main test, although a later, less conclusive analysis suggested that trainees may have been more likely to receive a perfect score. Behavioral human-centered design therefore cannot mean adding structure or reframing by default: the intervention has to fit the task, preserve workable autonomy and coordination, and be evaluated against completion and output quality rather than acceptance or belief movement alone.

2018-fugate-change-compliance-championing gives this behavioral layer a field-tested change-management mechanism. In a merger context, leader support did not simply reduce resistance; it shaped whether employees appraised change as threat or challenge, and challenge appraisal predicted both compliance and championing. That matters for AI because implementation success often depends on whether employees become active interpreters and advocates of new work practices, not just whether they stop resisting.

2026-liu-effects-generative-ai-knowledge-behavior supplies a direct AI-use case for the appraisal mechanism. In a three-wave survey of 428 knowledge workers, GenAI use was associated with both self-efficacy and job insecurity, which carried opposing indirect pathways to knowledge sharing minus hiding; competitive psychological climate strengthened both pathways. This is evidence that AI change can activate opportunity and replaceability appraisals together, not evidence that change management should simply suppress threat or amplify enthusiasm. The narrower behavioral outcome and its limits are developed in ai-mediated-knowledge-behavior.

2023-mkhize-examining-barriers-enablers-ai-adoption adds a small organizational case in which staff involvement was not merely rollout communication. Workers participated in early evaluation, gave domain-specific feedback, and could see that the system still required correction; the authors report cooperation and ownership alongside persistent job-security concerns. The case is weak evidence for a general intervention effect, but it sharpens a design requirement: participation can make the actual limits and repair work of AI visible, whereas an efficiency-only story may hide them.

Connections

• Relates to ai-adoption because adoption depends on behavioral mechanisms such as trust, resistance, loss aversion, availability, identity threat, and perceived autonomy, not only access, use-case discovery, or training.
• Relates to employee-change-support because behavioral human-centered AI needs to distinguish minimal compliance from championing and design for credible challenge appraisal rather than only suppressing resistance.
• Relates to ai-readiness because readiness includes whether leaders and implementation teams can diagnose human reactions, run behaviorally informed change management, and avoid managerial overconfidence or escalation of commitment.
• Relates to responsible-ai-deployment because fairness, transparency, autonomy, and control do not become responsible practice unless the deployment design accounts for how real stakeholders understand, contest, and act on those features.
• Relates to human-ai-collaboration because collaboration quality depends on how people calibrate trust, effort, authority, and vigilance when AI makes mistakes, appears confident, or changes the human role.
• Relates to ai-mediated-work-experience because behavioral human-centeredness has to preserve autonomy, competence, control, identity, and well-being during the change, not only after the system is technically live.

Contradictions & open questions

• 2026-farach-scaffolding-human-ai-collaboration makes the implementation trade-off empirical: a process intended to improve collaboration can create enough extra coordination and rule-following to worsen output. The evidence for thought-partner training is less clear because most participants already scored near the maximum, dropout differed between groups, and the earlier task may have influenced later performance. The open question is which behavioral supports improve sound AI use without turning change management into additional friction or a substitute for work redesign.
• 2025-schweitzer-leading-ai-adoption is a conceptual synthesis, so the stage model is useful but not yet a validated implementation model. 2018-fugate-change-compliance-championing supplies non-AI field evidence for appraisal-mediated change support, but the wiki still needs AI-specific field evidence on whether behaviorally informed AI change management improves adoption, trust calibration, worker outcomes, or system value.
• Behavioral design can protect workers from technosolutionism, but it can also become manipulation if organizations use behavioral insights mainly to overcome resistance rather than to redesign harmful or unwanted systems. The ethical boundary between behaviorally informed adoption and persuasion needs clearer evidence and practice rules.
• The relationship between behavioral human-centered AI and broader ai-readiness is unresolved: behavioral capability may be a readiness dimension, a change-management method, or an outcome of repeated implementation experience.