Survey Companion · 2026
Organizing Intelligence Over Time
Human–AI Collaboration as Joint Cognitive Development
Preprints.org v1 · DOI 10.20944/preprints202609.1092.v1
Abstract
As humans and AI agents work together across long-horizon tasks and accumulate interaction history, how cognitive work is organized on one task can shape later performance and later capability. A system may reason further, retrieve prior experience, verify, invoke a tool or another model, ask for human judgment, or stop. These choices also determine who gets what experience: whether the human practices a skill, whether the AI receives demonstrations or corrections, which failures are exposed for supervision, and which trajectories become reusable.
This survey connects human cognition, Human–Automation and HCI, and modern AI agents. Read together, these literatures expose a longitudinal causal sequence: organizing cognitive work determines the distribution of practice, supervision, feedback, and trajectories; that experience changes human skill and beliefs, machine memory and policy, and the coordination between them; and the updated joint system then organizes future work differently.
Joint Cognitive Development
The survey treats longitudinal Human–AI collaboration as a changing system rather than a sequence of tasks completed by a fixed team. The organization of current cognitive work shapes the experience available to both participants, and that experience changes the capabilities that return on later tasks.
Organize cognitive work
Reason, retrieve, verify, delegate, act, ask for help, or stop.
Generate selective experience
Practice, supervision, feedback, outcomes, and trajectories are distributed.
Change the participants
Human skill and beliefs, machine memory and policy, and coordination change.
Reorganize future work
The updated Human–AI system approaches later tasks with different capabilities.
Three Literatures, One Longitudinal Problem
| Tradition | What it contributes | Longitudinal implication |
|---|---|---|
| Human cognition | Bounded rationality, metareasoning, metacognitive control, expertise, and skill acquisition. | Experience changes representations, procedures, retrieval, and the cognitive operations needed later. |
| Human–Automation & HCI | Function allocation, supervisory control, mixed initiative, trust, shared autonomy, and human readiness. | How work is divided changes what the human continues to practice and how later intervention is performed. |
| Modern AI agents | Memory, workflows, skills, routing, verification, stopping, continual learning, and self-evolution. | Interaction trajectories can be retained and transformed into artifacts and policies that alter later behavior. |
Research Questions
-
How should interdependent cognitive operations be learned together?
Retrieval, reasoning, verification, routing, human involvement, and stopping interact; learning them independently can waste effort or suppress useful signals.
-
Reopening: when should experience lose authority?
Reusable workflows, skills, memories, and routing preferences save cognition only while the regularities that justified them continue to hold.
-
How should work be assigned when it also creates learning opportunities?
Delegation determines both current performance and which participant receives practice, demonstrations, corrections, and future learning value.
-
How should systems learn when their own decisions determine the feedback they receive?
Autonomy, review, routing, and takeover policies shape which errors are observed and which trajectories become supervised data.
-
How can a system distinguish user preference from adaptation to the system?
Observed user behavior may reflect stable preference, learned trust, checking cost, or behavior induced by earlier assistance.
-
How can we tell whether a Human–AI system has actually improved over time?
Longitudinal evaluation must separate history-driven capability change from extra compute, easier tasks, model upgrades, or greater human effort.
Resources
Literature map
The public companion bibliography tracks work on persistent experience, reusable skills, continual and self-evolving agents, adaptive cognitive control, repeated Human–AI adaptation, evaluation, and foundations.