Human-AI Joint Cognitive Systems

• Human-AI Joint Cognitive Systems are integrated ensembles where humans and AI collaborate as cognitive peers, sharing perception, reasoning, and decision-making.
• They rely on shared mental models, dynamic trust calibration, and joint action planning to enhance collective performance and adaptability.
• Practical applications range from autonomous vehicles to swarm robotics, emphasizing ethical design and long-term cognitive sustainability.

Human-AI Joint Cognitive Systems (HAIJCS) are integrated ensembles in which humans and artificial agents operate as cognitive peers, jointly perceiving, reasoning, deciding, and acting within dynamically coupled environments. Unlike traditional human-computer interaction, which positions the machine as an auxiliary tool or a passive interface, HAIJCS view both human and AI as autonomous, adaptive agents engaged in continuous, bidirectional collaboration to achieve shared objectives. This paradigm fundamentally reconceptualizes agency, authority, and responsibility in socio-technical systems, introducing a suite of theoretical, methodological, and empirical challenges that reshape human factors, human-centered AI, and collective intelligence research.

1. Theoretical Foundations and Formal Structure

Central to HAIJCS is the theoretical shift from unidirectional control logic to joint cognitive agency. Foundationally, this aligns with Joint Cognitive Systems (JCS) theory, which treats complex socio-technical systems as networks of intertwined human and machine cognition, rather than as simple operator-assistant hierarchies (Gao et al., 16 Jan 2026, Xu et al., 2023, Xu et al., 2022). Key cognitive constructs—situation awareness (SA), intent, trust, and decision authority—are explicitly shared and dynamically allocated between human and AI agents.

A formal instantiation is given by

$$\text{HAIJCS} = \langle H,\, A,\, \mathcal{I},\, \Omega \rangle$$

where $H$ denotes the human cognitive agent, $A$ the AI cognitive agent, $\mathcal{I}$ the multimodal cognitive interface (supporting real-time bidirectional communication and shared mental model formation), and $\Omega$ the envelope of ethical and authority constraints (asserting human ultimate authority in critical tasks) (Xu et al., 2023).

The joint perception–comprehension–projection–decision–action loops are mathematically unfolded using coupled situation awareness variables for each agent: $$SA_h(t) = \{P_h(t),\,C_h(t),\,R_h(t)\},\qquad SA_{ai}(t) = \{P_{ai}(t),\,C_{ai}(t),\,R_{ai}(t)\}$$ with shared situation awareness computed as a reliability-weighted aggregation: $SA_{shared}(t) = w_h\,SA_h(t) + w_{ai}\,SA_{ai}(t)$ where $w_h + w_{ai} = 1$ and weights adapt to agent reliability and task criticality (Gao et al., 16 Jan 2026, Xu et al., 2022).

In advanced architectural treatments, HAIJCS are cast as multi-agent POMDPs: $$S = H \times A ,\quad O = O_H \times O_A ,\quad A = A_H \times A_A$$ and policy, transition, and reward functions operate jointly over human and AI state-action pairs (Mao et al., 28 Aug 2025).

2. Core Principles and Design Paradigms

HAIJCS frameworks are anchored on several unifying principles:

• Cognitive Peerhood: Both human and AI are modeled as full cognitive agents capable of perception, memory, inference, and decision (Gao et al., 16 Jan 2026, Xu et al., 2023).
• Shared Mental Models and Dynamic Trust: The explicit engineering of shared mental models (SMMs) and real-time, calibrated trust loops are essential for sustaining resilient cooperation (Tong, 7 Nov 2025, Gao et al., 16 Jan 2026).
• Human-Centered AI (HCAI) Conformance: HAIJCS strictly maintain human final authority, integrating ethical design, accountability, and transparency constraints; the interface, decision processes, and feedback mechanisms must support interpretability and meaningful human control at all time (Xu et al., 2023, Xu et al., 2022).
• Mutual Growth and Co-Learning: Both partners iteratively adapt: the AI incrementally retrains on human corrections and the human refines mental models based on AI explanations, formalized as dual self-adapting learning loops (Huang et al., 2019).
• Bidirectional Feedback and Role Negotiation: Not limited to "AI explains, human accepts/overrides," but supporting mutual critique, real-time role reallocation, and joint sensemaking at the interface (Huang et al., 2019, Borghoff et al., 19 Feb 2025).

Pragmatically, HAIJCS span multiple architectural layers, including cognitive-agent, team/situational, social-interaction (shared social understanding), ecosystemic, and regulatory/societal levels (Gao et al., 16 Jan 2026, Xu et al., 2022).

3. Computational Models and Mechanisms

Information Flow and Cognitive Loop Architectures: System architectures instantiate parallel information-processing pipelines for both human and AI, including sensory acquisition, situation assessment, comprehension, forecasting, decision-making, execution, and adaptive control modules. These components are explicitly coupled via shared memory, trust, and intent-management structures (Xu et al., 2023, Xu et al., 2022). Joint planning and decision modules reconcile divergent strategies subject to ethical and authority constraints.

Formal Decision Models: Joint workloads and team performance are operationalized as

$W_{joint}(t) = W_H(t) + W_A(t) - \sigma_{sync}(t)$

and

$H$0

with $H$1, $H$2 the individual agent performances, $H$3 a coordination bonus (such as mutual information between agent actions), and $H$4, $H$5, $H$6 the cognitive loads (Xu et al., 2023). Synergistic performance is achieved when $H$7 exceeds both $H$8 and $H$9.

Cognitive Amplification vs. Delegation Metrics: Recent work formalizes regimes of amplification ($A$0) versus delegation ($A$1), using

$A$2

where $A$3, $A$4, $A$5 are joint, human-only, and AI-only performance, and $A$6 measures the drift in unaided human performance over time (Santi, 19 Mar 2026).

Hybrid System Models: MAS vs Centaurian: HAIJCS can be operationalized in multi-agent system (MAS) mode, maintaining agent autonomy and structured communication, or in Centaurian, deep-integration mode, in which human and AI become tightly coupled composites represented via colored Petri nets that instantiate joint-decision transitions conditioned on both human and AI tokens (Borghoff et al., 19 Feb 2025).

4. Evaluation, Performance Metrics, and Empirical Evidence

Quantitative Metrics: Joint performance, situation awareness alignment, cognitive workload, and trust calibration are foundational evaluation axes. Example metrics include

• Accuracy, response time, and error rates for collaborative versus solo agents
• $A$7 for quantifying shared situation awareness alignment
• Trust calibration error $A$8
• Human Cognitive Drift Rate (HCDR) for long-term expertise retention (Santi, 19 Mar 2026, Kural et al., 2023)

Standardized experiments in domains such as autonomous driving, creative design, and robot teleoperation report significant reductions in human workload, improved time-to-insight, and increased solution diversity—provided the interface enables bidirectional feedback and adaptive trust (Huang et al., 2019, Xu et al., 2023, Nguyen et al., 5 Aug 2025).

Meta-Analytic Dynamics: Integration studies reveal a performance paradox: in judgment/decision tasks with algorithm-in-the-loop dynamics, human-AI teams may underperform compared to AI alone due to cognitive noise, aversion/automation bias, and deskilling. Conversely, content-creation and problem-formulation tasks exhibit strong positive synergy, particularly when shared mental models are well calibrated (Tong, 7 Nov 2025).

Calibration and Alignment: Confidence-aligned arbitration (e.g., maximum-confidence selection) improves joint inference only when AI's metacognitive calibration is robust (as measured by area under the type-2 ROC curve). Misalignment sharply reduces complementarity and erodes trust (Nguyen et al., 5 Aug 2025, Kural et al., 2023).

5. Architecture, Implementation, and Application Domains

Architectural Templates: HAIJCS implementations exploit modular, layered software and hardware architectures:

• Cognitive agent modules with parallel information-processing capabilities (perception, assessment, comprehension, action)
• Multimodal human–AI interface $A$9 with real-time context fusion, transparency, and bidirectional teaching/learning loops
• Communication-space Petri nets partitioned into surface, observation, and computation layers enabling both MAS and Centaurian operation (Borghoff et al., 19 Feb 2025)

Practical Instantiations:

• Autonomous vehicles: Dynamic human–AI handover interfaces and shared dashboards elevate drivers from supervisors to full cognitive co-drivers (Gao et al., 16 Jan 2026, Xu et al., 2022).
• Intelligent cockpits: Single-pilot operations leveraging real-time shared situation assessments, collaborative planning, and adaptive function allocation (Gao et al., 16 Jan 2026).
• Swarm robotics: Human–AI “cyborg” patterns that integrate physical, sensorimotor, and cognitive control in distributed multi-robot systems (Rockbach et al., 29 Nov 2025).
• Creative and educational settings: Co-learning sandboxes for creative design, joint tutoring systems with integrated short- and long-term memory, and adaptive scaffolding frameworks (Huang et al., 2019, Salas-Guerra, 6 Feb 2025).

Collective Intelligence and Network Science: Multi-agent, multilayer network models quantify emergent collective intelligence (CI) as a function of cross-layer connectivity and agent diversity. Algebraic connectivity of the supra-Laplacian and diversity scores operationalize group cognition and creativity in human–AI collectives (Cui et al., 2024).

6. Open Challenges and Future Directions

Trust, Ethics, and Governance: Robust HAIJCS require transparent mechanisms for real-time trust calibration; explicit handling of compound human–AI biases; enforceable governance and accountability frameworks; and defense against cognitive deskilling and overreliance (Tong, 7 Nov 2025, Felten, 26 Apr 2025, Pimplikar et al., 2017).

Cognitive Sustainability Constraints: Systemic optimization must maximize joint performance while enforcing cognitive sustainability ($\mathcal{I}$0), preventing long-term erosion of human expertise even as short-term hybrid performance improves (Santi, 19 Mar 2026).

Mitigation of Compound Bias: Interactionist frameworks emphasize the management of reciprocal amplification of cognitive biases, with joint bias-mitigation strategies drawing from both psychological interventions and AI fairness-aware learning (Felten, 26 Apr 2025).

Measurement and Adaptation: Persistent challenges include measuring cognitive-layer observability in multilayer networks, dynamically adapting trust/authority allocation, and maintaining system transparency at scale in evolving, distributed environments (Cui et al., 2024, Gao et al., 16 Jan 2026, Salas-Guerra, 6 Feb 2025).

Ethically Charged Use Cases and Cultural Alignment: HAIJCS must incorporate cross-cultural cognition, value-sensitive policy frameworks, and adversarially tested watchdog mechanisms—especially in domains (e.g., law, education, critical infrastructure) where autonomy and responsibility allocation remain technologically and ethically fraught (Pimplikar et al., 2017, Mao et al., 28 Aug 2025).

Research Agenda and Methodology: Ongoing priorities include longitudinal studies of co-adaptation and extended-self absorption, development of open-source toolkits for real-time joint cognitive-systems engineering, integration of multi-agent game theory, and formalization of evaluation metrics spanning task success, well-being, and societal impact (Tong, 7 Nov 2025, Gao et al., 16 Jan 2026, Cui et al., 2024).

---

Human-AI Joint Cognitive Systems thus define a new epistemic and engineering frontier, recasting agency, responsibility, and intelligence in coupled human–machine collectives. Realizing their promise demands technical, conceptual, and ethical advances across cognitive modeling, interface architecture, trust calibration, sociotechnical system design, and long-term cognitive sustainability.